Advances in Intelligent Systems and Computing 882

N. R. Shetty · L. M. Patnaik ·
H. C. Nagaraj · Prasad Naik Hamsavath ·
N. Nalini Editors

Emerging Research
in Computing,
Information,
Communication and
Applications
ERCICA 2018, Volume 1
Advances in Intelligent Systems and Computing

Volume 882

Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland

Advisory Editors
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing, Universidad
Central de Las Villas, Santa Clara, Cuba
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
Hani Hagras, Electronic Engineering, University of Essex, Colchester, UK
László T. Kóczy, Department of Automation, Széchenyi István University,
Gyor, Hungary
Vladik Kreinovich, Department of Computer Science, University of Texas
at El Paso, El Paso, TX, USA
Chin-Teng Lin, Department of Electrical Engineering, National Chiao Tung
University, Hsinchu, Taiwan
Jie Lu, Faculty of Engineering and Information Technology, University of
Technology Sydney, Sydney, NSW, Australia
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute of
Technology, Tijuana, Mexico
Nadia Nedjah, Department of Electronics Engineering, University of Rio de
Janeiro, Rio de Janeiro, Brazil
Ngoc Thanh Nguyen, Faculty of Computer Science and Management, Wrocław
University of Technology, Wrocław, Poland
Jun Wang, Department of Mechanical and Automation Engineering, The Chinese
University of Hong Kong, Shatin, Hong Kong

[email protected]
The series “Advances in Intelligent Systems and Computing” contains publications
on theory, applications, and design methods of Intelligent Systems and Intelligent
Computing. Virtually all disciplines such as engineering, natural sciences, computer
and information science, ICT, economics, business, e-commerce, environment,
healthcare, life science are covered. The list of topics spans all the areas of modern
intelligent systems and computing such as: computational intelligence, soft comput-
ing including neural networks, fuzzy systems, evolutionary computing and the fusion
of these paradigms, social intelligence, ambient intelligence, computational neuro-
science, artificial life, virtual worlds and society, cognitive science and systems,
Perception and Vision, DNA and immune based systems, self-organizing and
adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics
including human-machine teaming, knowledge-based paradigms, learning para-
digms, machine ethics, intelligent data analysis, knowledge management, intelligent
agents, intelligent decision making and support, intelligent network security, trust
management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are
primarily proceedings of important conferences, symposia and congresses. They
cover significant recent developments in the field, both of a foundational and
applicable character. An important characteristic feature of the series is the short
publication time and world-wide distribution. This permits a rapid and broad
dissemination of research results.

** Indexing: The books of this series are submitted to ISI Proceedings,

EI-Compendex, DBLP, SCOPUS, Google Scholar and Springerlink **

More information about this series at http://www.springer.com/series/11156

[email protected]
N. R. Shetty L. M. Patnaik
• •

H. C. Nagaraj Prasad Naik Hamsavath

• •

N. Nalini
Editors

Emerging Research
in Computing, Information,
Communication
and Applications
ERCICA 2018, Volume 1

123
[email protected]
Editors
N. R. Shetty L. M. Patnaik
Central University of Karnataka National Institute of Advanced Studies
Kalaburagi, Karnataka, India Bangalore, Karnataka, India

H. C. Nagaraj Prasad Naik Hamsavath

Nitte Meenakshi Institute of Technology Nitte Meenakshi Institute of Technology
Bangalore, Karnataka, India Bangalore, Karnataka, India

N. Nalini
Nitte Meenakshi Institute of Technology
Bangalore, Karnataka, India

ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-981-13-5952-1 ISBN 978-981-13-5953-8 (eBook)
https://doi.org/10.1007/978-981-13-5953-8

Library of Congress Control Number: 2018966829

© Springer Nature Singapore Pte Ltd. 2020

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore

[email protected]
Organizing Committee

ERCICA 2018

The Fifth International Conference on “Emerging Research in Computing,

Information, Communication and Applications”, ERCICA 2018, was held during
July 27–28, 2018, at the Nitte Meenakshi Institute of Technology (NMIT),
Bangalore, and organized by the Departments of CSE and MCA, NMIT

Chief Patrons

Dr. N. V. Hegde, President, Nitte Education Trust, Mangalore, India

Dr. N. R. Shetty, Chancellor, Central University of Karnataka, Kalburgi, and
Advisor, Nitte Education Trust, Mangalore, India

Conference Chair

Dr. H. C. Nagaraj, Principal, NMIT, Bangalore, India

Program Chairs

Dr. Prasad Naik Hamsavath, HOD, MCA, NMIT, Bangalore, India

Dr. N. Nalini, Professor, CSE, NMIT, Bangalore, India

[email protected]
vi Organizing Committee

Publisher

Springer

Advisory Chairs

Dr. K. Sudha Rao, Advisor, Admin and Management, NMIT, Bangalore, India
Mr. Rohit Punja, Administrator, NET, Mangalore, India
Dr. Jharna Majumdar, Dean (R&D), NMIT, Bangalore, India
Mr. K. A. Ranganatha Setty, Dean (Academic), NMIT, Bangalore, India

Advisory Committee

Dr. L. M. Patnaik, INSA Senior Scientist, NIAS, Bangalore, India

Dr. B. S. Sonde, Former Vice Chancellor, Goa University, Goa, India
Dr. D. K. Subramanian, Former Dean and Professor, IISc, Bangalore, India
Dr. K. D. Nayak, Former OS & CC, R&D (MED & MIST), DRDO, India
Dr. Kalidas Shetty, Founding Director of Global Institute of Food Security and
International Agriculture (GIFSIA), North Dakota State University, Fargo, USA
Dr. Kendall E. Nygard, Professor of Computer Science and Operations Research,
North Dakota State University, Fargo, USA
Dr. Sathish Udpa, Dean and Professor, Michigan State University, Michigan, USA
Dr. K. N. Bhat, Visiting Professor, Center for Nano Science and Engineering-
CeNSE, IISc, Bangalore, India
Dr. K. R. Venugopal, Principal, UVCE, Bangalore, India
Dr. C. P. Ravikumar, Director, Technical Talent Development at Texas Instruments,
Bangalore, India
Dr. Navakanta Bhat, Chairperson, Center for Nano Science and Engineering-
CeNSE, IISc, Bangalore, India
Dr. Anand Nayyar, Professor, Researcher and Scientist in Graduate School, Duy
Tan University, Da Nang, Vietnam

Program Committee

Dr. Savitri Bevinakoppa, Professional Development and Scholarship Coordinator,

School of IT and Engineering, Melbourne Institute of Technology (MIT), Australia
Dr. P. Ramprasad, Professor, Department of CSE and IT, Manipal University,
Dubai

[email protected]
Organizing Committee vii

Dr. Ohta Tsuyoshi, Department of Computer Sciences, Shizuoka University, Japan

Dr. Sonajharia Minz, Professor, School of Computer and Systems Sciences,
Jawaharlal Nehru University, New Delhi, India
Dr. Sanjay Kumar Dhurandher, Professor and Head, Department of Information
Technology, Netaji Subhas Institute of Technology, New Delhi, India
Dr. Ramesh R. Galigekere, Professor and HOD, Department of Biomedical
Networks, University Putra Malaysia, Malaysia
Dr. K. G. Srinivasa, CBP Government Engineering College, New Delhi, India
Dr. S. Ramanarayana Reddy, HOD, Department of CSE, IGDTU for Women,
Kashmere Gate, Delhi, India
Dr. K. Raghavendra, Professor, School of Computing, National University of
Singapore, Singapore
Dr. Abhijit Lele, Principal Consultant, Robert Bosch, Bangalore, India
Dr. Bappaditya Mandal, Faculty of Science, Engineering and Computing, Kingston
University, London

Organizing Co-chairs

Dr. M. N. Thippeswamy, Professor and Head, Department of CSE, NMIT,

Bangalore, India
Dr. H. A. Sanjay, Professor and Head, Department of ISE, NMIT, Bangalore, India
Dr. S. Sandya, Professor and Head, Department of ECE, NMIT, Bangalore, India
Dr. H. M. Ravikumar, Professor and Head, Department of EEE, NMIT, Bangalore,
India

[email protected]
Theme Editors

Computing

Dr. N. Nalini
Dr. Thippeswamy M. N.

Information

Dr. Sanjay H. A.
Dr. Shakti Mishra

Communication

Dr. H. C. Nagaraj
Dr. Raghunandan S.
Prof. Sankar Dasiga

Application

Dr. Prasad Naik Hamsavath

Prof. Sitaram Yaji

ix

[email protected]
Preface

The Fifth International Conference on “Emerging Research in Computing,

Information, Communication and Applications,” ERCICA 2018, is an annual event
organized at the Nitte Meenakshi Institute of Technology (NMIT), Yelahanka,
Bangalore, India.
ERCICA aims to provide an interdisciplinary forum for discussion among
researchers, engineers and scientists to promote research and exchange of knowl-
edge in computing, information, communication and related applications. This
conference will provide a platform for networking of academicians, engineers and
scientists and also will enthuse the participants to undertake high-end research in
the above thrust areas.
ERCICA 18 received more than 400 papers from all over the world, viz. from
China, UK, Africa, Saudi Arabia and India. The ERCICA Technical Review
Committee has followed all necessary steps to screen more than 400 papers by
going through six rounds of quality checks on each paper before selection for
presentation/publication.
The acceptance ratio is only 1:3.

Kalaburagi, India N. R. Shetty

Bangalore, India L. M. Patnaik
July 2018 H. C. Nagaraj
Prasad Naik Hamsavath
N. Nalini

xi

[email protected]
Acknowledgements

First of all, we would like to thank Prof. N. R. Shetty who has always been the
guiding force behind this event’s success. It was his dream that we have striven to
make a reality. Our thanks to Prof. L. M. Patnaik, who has monitored the whole
activity of the conference from the beginning till its successful end.
Our special thanks to Springer and especially the editorial staff who were patient,
meticulous and friendly with their constructive criticism on the quality of papers
and outright rejection at times without compromising the quality of the papers as
they are always known for publishing the best international papers.
We would like to express our gratitude to all the review committee members of
all the themes of computing, information, communication and applications and the
best-paper-award review committee members.
Finally, we would like to express our heartfelt gratitude and warmest thanks to
the ERCICA 2018 organizing committee members for their hard work and out-
standing efforts. We know how much time and energy this assignment demanded,
and we deeply appreciate all the efforts to make it a grand success.
Our special thanks to all the authors who have contributed to publishing their
research work in this conference and participated to make this conference a grand
success. Thanks to everyone who have directly or indirectly contributed to the
success of this conference ERCICA 2018.

Regards
Program Chairs
ERCICA 2018

xiii

[email protected]
About the Conference

ERCICA 2018
The Fifth International Conference on “Emerging Research in Computing,
Information, Communication and Applications,” ERCICA 2018, is an annual event
jointly organized by the Departments of CSE and MCA during July 27–28, 2018, at
the Nitte Meenakshi Institute of Technology (NMIT), Yelahanka, Bangalore, India.
ERCICA 2018 is organized under the patronage of Prof. N. R. Shetty, Advisor,
Nitte Education Trust. Dr. L. M. Patnaik, Technical Advisor, NMIT, and Dr. H. C.
Nagaraj, Principal, served as the Conference Chairs, and the Program Chairs of the
conference were Dr. Prasad Naik Hamsavath, Professor and Head, MCA, and Dr.
N. Nalini, Professor, CSE, NMIT, Bangalore, Karnataka.
ERCICA aims to provide an interdisciplinary forum for discussion among
researchers, engineers and scientists to promote research and exchange of knowl-
edge in computing, information, communication and related applications. This
conference will provide a platform for networking of academicians, engineers and
scientists and also will enthuse the participants to undertake high-end research in
the above thrust areas.

xv

[email protected]
Contents

A Computational Segmentation Tool for Processing Patient

Brain MRI Image Data to Automatically Extract Gray
and White Matter Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Ayush Goyal, Sunayana Tirumalasetty, Disha Bathla, Manish K. Arya,
Rajeev Agrawal, Priya Ranjan, Gahangir Hossain and Rajab Challoo
Developing Ontology for Smart Irrigation of Vineyards . . . . . . . . . . . . 17
Archana Chougule and Debajyoti Mukhopadhyay
An Intensive Review of Data Replication Algorithms
for Cloud Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Chetna Dabas and Juhi Aggarwal
Integrated Cryptography for Internet of Things Using TBF
Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Santosh Kumar Sharma, Muppidi Somasundara Rao,
Lavudi Poorna Chandar Rao and Pendyala Chaitanya
Data Governance on Local Storage in Offsite . . . . . . . . . . . . . . . . . . . . . 53
G. Priyadharshini and K. Shyamala
Extraction of Character Personas from Novels Using Dependency
Trees and POS Tags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Nikhil Prabhu and S. Natarajan
A Multifactor Authentication Model to Mitigate the Phishing
Attack of E-Service Systems from Bangladesh Perspective . . . . . . . . . . . 75
Md. Zahid Hasan, Abdus Sattar, Arif Mahmud and Khalid Hasan Talukder
Study on Energy-Efficient and Lifetime-Enhanced Clustering
Algorithm-Based Routing Protocols in Wireless Sensor
Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Ishita Banerjee and P. Madhumathy

xvii

[email protected]
xviii Contents

Role of Fog Computing in IoT-Based Applications . . . . . . . . . . . . . . . . 99

Charul Thareja and N. P. Singh
Single Horizontal Camera-Based Object Tracking Quadcopter
Using StaGaus Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Lakshmi Shrinivasan and N. R. Prasad
IoT-Enabled Medicine Bottle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
A. R. Shreyas, Soumya Sharma, H. Shivani and C. N. Sowmyarani
Revamp Perception of Bitcoin Using Cognizant Merkle . . . . . . . . . . . . . 141
J. Vijayalakshmi and A. Murugan
A Novel Algorithm for DNA Sequence Compression . . . . . . . . . . . . . . . 151
K. Punitha and A. Murugan
DigiPen: An Intelligent Pen Using Accelerometer for Character
Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Ankur Agarwal and Swaroop Sudhanva Belur
Design of FPGA-Based Radar and Beam Controller . . . . . . . . . . . . . . . 171
Adesh Panwar and Neha Goyal
Effect of Lattice Topologies and Distance Measurements
in Self-Organizing Map for Better Classification . . . . . . . . . . . . . . . . . . 183
Sathiapriya Ramiah
Evaluation and Classification of Road Accidents Using Machine
Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Jaspreet Singh, Gurvinder Singh, Prithvipal Singh and Mandeep Kaur
Multi-language Handwritten Recognition in DWT Accuracy
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
T. P. Umadevi and A. Murugan
A Novel H-∞ Filter Based Indicator for Health Monitoring
of Components in a Smart Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
E. Ranjini Warrier, P. V. Sunil Nag and C. Santhosh Kumar
A Survey on Intelligent Transportation System Using Internet
of Things . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Palak Patel, Zunnun Narmawala and Ankit Thakkar
CRUST: A C/C++ to Rust Transpiler Using a “Nano-parser
Methodology” to Avoid C/C++ Safety Issues in Legacy Code . . . . . . . . 241
Nishanth Shetty, Nikhil Saldanha and M. N. Thippeswamy
Species Environmental Niche Distribution Modeling for Panthera
Tigris Tigris ‘Royal Bengal Tiger’ Using Machine Learning . . . . . . . . . 251
Shaurya Bajaj and D. Geraldine Bessie Amali

[email protected]
Contents xix

Organizational Digital Footprint for Traceability, Provenance

Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Sheetal Arya, Kumar Abhishek and Akshay Deepak
Bidirectional Long Short-Term Memory for Automatic English
to Kannada Back-Transliteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
B. S. Sowmya Lakshmi and B. R. Shambhavi
A Dominant Point-Based Algorithm for Finding Multiple Longest
Common Subsequences in Comparative Genomics . . . . . . . . . . . . . . . . . 289
Manish M. Motghare and Preeti S. Voditel
Fast and Accurate Fingerprint Recognition in Principal
Component Subspace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
S. P. Ragendhu and Tony Thomas
Smart Meter Analysis Using Big Data Techniques . . . . . . . . . . . . . . . . . 317
Neha Pandey, Nivedita Das, Sandeep Agarwal, Kashyap Barua,
Manjusha Pandey and Siddharth Swarup Rautray
Movie Recommendation System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
S. Rajarajeswari, Sharat Naik, Shagun Srikant, M. K. Sai Prakash
and Prarthana Uday
A Semiautomated Question Paper Builder Using Long Short-Term
Memory Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Rajarajeswari Subramanian, Akhilesh P. Patil,
Karthik Ganesan and T. S. Akarsh
Time-Critical Transmission Protocols in Wireless Sensor Networks:
A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
Archana R. Raut, S. P. Khandait and Urmila Shrawankar
Impact of Shuffler Design Pattern on Software Quality . . . . . . . . . . . . . 365
G. Priyalakshmi, R. Nadarajan, Joseph W. Yoder, S. Arthi
and G. Jayashree
Efficient Algorithms for Text Lines and Words Segmentation
for Recognition of Arabic Handwritten Script . . . . . . . . . . . . . . . . . . . . 387
Amani Ali Ahmed Ali and M. Suresha
Privacy-Preserving Lightweight Image Encryption
in Mobile Cloud . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
M. Sankari and P. Ranjana
Performance Evaluation of Ensemble-Based Machine Learning
Techniques for Prediction of Chronic Kidney Disease . . . . . . . . . . . . . . 415
K. M. Zubair Hasan and Md. Zahid Hasan

[email protected]
xx Contents

Proficient Cooperative Caching in SWNET Using Twin Segments

Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
B. N. Lakshmi Narayan, Prasad N. Hamsavath, Meher Taj, E. G. Satish,
Vivek Bharadwaj and S. Rabendranath
A Particle Swarm Optimization-Backpropagation (PSO-BP) Model
for the Prediction of Earthquake in Japan . . . . . . . . . . . . . . . . . . . . . . . 435
Abey Abraham and V. Rohini
Analysis and Detection of Diabetes Using Data Mining
Techniques—A Big Data Application in Health Care . . . . . . . . . . . . . . . 443
B. G. Mamatha Bai, B. M. Nalini and Jharna Majumdar
Cyclic Scheduling Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Ravin Kumar
A Framework for Monitoring Clustering Stability Over Time . . . . . . . . 467
K. Namitha and G. Santhosh Kumar
Efficient Dynamic Double Threshold Energy Detection of Cooperative
Spectrum Sensing in Cognitive Radio . . . . . . . . . . . . . . . . . . . . . . . . . . 479
Shahbaz Soofi, Anjali Potnis and Prashant Diwivedy
ConvFood: A CNN-Based Food Recognition Mobile Application
for Obese and Diabetic Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
Kaiz Merchant and Yash Pande
Segmentation and Recognition of E. coli Bacteria Cell in Digital
Microscopic Images Based on Enhanced Particle Filtering
Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
Manjunatha Hiremath
Automatic Bengali Document Categorization Based on Deep
Convolution Nets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
Md. Rajib Hossain and Mohammed Moshiul Hoque
Artist Recommendation System Using Hybrid Method:
A Novel Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
Ajay Dhruv, Aastha Kamath, Anuja Powar and Karan Gaikwad
Anomaly Detection of DDOS Attacks Using Hadoop . . . . . . . . . . . . . . . 543
Y. S. Kalai vani and P. Ranjana
Pedestrian Detection and Tracking: A Driver Assistance
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
Shruti Maralappanavar, Nalini C. Iyer and Meena Maralappanavar
Exploiting Parallelism Available in Loops Using Abstract
Syntax Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563
Anil Kumar and Hardeep Singh

[email protected]
Contents xxi

A Study on Cooperation and Navigation Planning for Multi-robot

Using Intelligent Water Drops Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 577
D. Chandrasekhar Rao and Manas Ranjan Kabat
Customer’s Activity Recognition in Smart Retail Environment
Using AltBeacon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
M. Lakshmi, Alolika Panja, Naini and Shakti Mishra
An Active Mixer Design For Down Conversion in 180 nm CMOS
Technology for RFIC Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 605
B. H. Shraddha and Nalini C. Iyer
Analysis of PAPR for Performance QPSK and BPSK Modulation
Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621
K. Bhagyashree, S. Ramakrishna and Priyatam Kumar
Implementation of Modified Array Multiplier for WiMAX
Deinterleaver Address Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629
Patil Nikita, Arun Kakhandki, S. Ramakrishna and Priyatam Kumar
Link Quality-Based Mobile-Controlled Handoff Analysis
Using Stochastic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
S. Akshitha and N. G. Goudru
A Comparative Analysis of Lightweight Cryptographic
Protocols for Smart Home . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663
Rupali Syal
Algorithm Study and Simulation Analysis by Different
Techniques on MANET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671
Nithya Rekha Sivakumar and Abeer Al Garni
Cloud-Based Agricultural Framework for Soil Classification
and Crop Yield Prediction as a Service . . . . . . . . . . . . . . . . . . . . . . . . . 685
K. Aditya Shastry and H. A. Sanjay

[email protected]
About the Editors

Prof. N. R. Shetty is the Chancellor of Central University of Karnataka,

Kalaburagi, and Chairman of the Review Commission for the State Private
University Karnataka. He is currently serving as an advisor to the Nitte Meenakshi
Institute of Technology (NMIT), Bangalore. He is also founder Vice-President
of the International Federation of Engineering Education Societies (IFEES),
Washington DC, USA. He served as Vice Chancellor of Bangalore University for
two terms and President of the ISTE, New Delhi for three terms. He was also a
member of the AICTE’s Executive Committee and Chairman of its South West
Region Committee.

Prof. L. M. Patnaik obtained his PhD in Real-Time Systems in 1978, and his DSc
in Computer Systems and Architectures in 1989, both from the Indian Institute of
Science, Bangalore. From 2008 to 2011, he was Vice Chancellor of the Defense
Institute of Advanced Technology, Deemed University, Pune. Currently he is an
Honorary Professor with the Department of Electronic Systems Engineering, Indian
Institute of Science, Bangalore, and INSA Senior Scientist and Adjunct Professor
with the National Institute of Advanced Studies, Bangalore.

Dr. H. C. Nagaraj completed his B.E in Electronics & Communication from the
University of Mysore in 1981, his M.E in Communication Systems from P.S.G
College of Technology, Coimbatore in 1984. He was awarded Ph.D (Biomedical
Signal Processing and Instrumentation) from Indian Institute of Technology
Madras, Chennai in 2000. Dr. Nagaraj has teaching experience spanning more than
35 years. He was the Chairman, BOS of IT/BM/ML of Visvesvaraya Technological
University, Belagavi for 2010-13 and Member, Academic Senate of VTU for 06
years w.e.f. April 2010. Further extended for a period of three years w.e.f
02-06-2016. He has the credit of publishing more than 40 technical papers and has
also published a book- “VLSI Circuits”, Star- Tech Education, Bangalore in 2006.
He has won the Best Student Paper Award at the 5th National Conference of
Biomechanics held at I.I.T. Madras, Chennai in 1996 and Best Paper Award at the
Karnataka State Level Seminar on “Introduction of Flexible System in Technical

xxiii

[email protected]
xxiv About the Editors

Education” under Visveswaraiah Technological University in 1999 at P.E.S.

Institute of Technology, Bangalore. Presently, he is the Dean, Faculty of
Engineering, Visvesvaraya Technological University Belgaum, for three years from
2016 to 2019 and Member of the Court, Pondicherry University. He is the Member
of Karnataka State Innovation Council, Government of Karnataka and Member of
NAAC (UGC) Peer Team to assess the institutions for Accreditation. He has also
visited as an Expert Member of the UGC, New Delhi for inspecting the colleges
seeking Autonomous Status.

Dr. Prasad Naik Hamsavath is a Professor and Head of the Department of Master
of Computer Applications at Nitte Meenakshi Institute of Technology, Bangalore.
He completed his PhD at Jawaharlal Nehru University, New Delhi, India.
Dr. Prasad N H has more than 12 years of experience in different roles in both
public and private sector enterprises, including the Ministry of Human Resource
and Development, New Delhi, Government of India. He received the prestigious
“Dr. Abdul Kalam Life Time Achievement Award” and also received a
“Young Faculty” award at the 2nd Academic Brilliance Awards.

Dr. N. Nalini is a Professor at the Department of Computer Science and

Engineering at Nitte Meenakshi Institute of Technology, Bangalore. She received
her MS from BITS, Pilani in 1999 and her PhD from Visvesvaraya Technological
University in 2007. She has more than 21 years of teaching and 14 years of research
experience. She has written numerous international publications, and Received
“Bharath Jyoti Award” by India International Friendship Society, New Delhi
on 2012, from Dr. Bhishma Narain Singh, former Governor of Tamilnadu and
Assam. She received the “Dr. Abdul Kalam Life time achievement National
Award” for excellence in Teaching, Research Publications, and Administration
by International Institute for Social and Economic Reforms, IISER, Bangalore
on 29th Dec 2014. She is also the recipient of “Distinguished Professor” award by
TechNext India 2017 in association with Computer Society of India-CSI, Mumbai
Chapter and “Best Professor in Computer Science & Engineering “ award by
26th Business School Affaire & Dewang Mehta National Education Awards
(Regional Round) on 5th September 2018, at Bangalore. She is a lifetime
member of the ISTE, CSI, ACEEE and IIFS.

[email protected]
A Computational Segmentation Tool
for Processing Patient Brain MRI Image
Data to Automatically Extract Gray
and White Matter Regions

Ayush Goyal, Sunayana Tirumalasetty, Disha Bathla, Manish K. Arya,

Rajeev Agrawal, Priya Ranjan, Gahangir Hossain and Rajab Challoo

Abstract Brain MRI imaging is necessary to screen and detect diseases in the brain,
and this requires processing, extracting, and analyzing a patient’s MRI medical image
data. Neurologists and neurological clinicians, technicians, and researchers would be
greatly facilitated and benefited by a graphical user interface-based computational
tool that could perform all the required medical MRI image processing functions
automatically, thus minimizing the cost, effort, and time required in screening dis-
ease from the patient’s MRI medical image data. Thus, there is a need for automatic
medical image processing software platforms and for developing tools with applica-
tions in the medical field to assist neurologists, scientists, doctors, and academicians
to analyze medical image data automatically to obtain patient-specific clinical param-
eters and information. This research develops an automatic brain MRI segmentation
computational tool with a wide range of neurological applications to detect brain
patients’ disease by analyzing the special clinical parameters extracted from the
images and to provide patient-specific medical care, which can be especially helpful
at early stages of the disease. The automatic brain MRI segmentation is performed
based on modified pixel classification technique called fuzzy c-means followed by
connected component labeling.

Keywords Segmentation · Medical imaging · Fuzzy c-means · Neurological

application

A. Goyal (B) · S. Tirumalasetty · G. Hossain · R. Challoo

Texas A&M University, Kingsville, TX, USA
e-mail: [email protected]
D. Bathla · P. Ranjan
Amity University Uttar Pradesh, Noida, UP, India
M. K. Arya · R. Agrawal
G.L. Bajaj Institute of Management and Research, Greater Noida, UP, India

© Springer Nature Singapore Pte Ltd. 2019 1

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_1

[email protected]
2 A. Goyal et al.

1 Introduction

In the field of medical image processing, the most challenging task to any neurologist
or a doctor or a scientist is to detect the patient’s disease by analyzing the patient’s
clinical information. Patient’s data is extracted and analyzed to detect the abnormal-
ities and to measure the illness of the disease which helps a medical practitioner to
cure the disease at its early stages. Extraction of brain abnormalities in brain MRI
images is performed by segmentation of gray and white matter regions in patient’s
brain MRI images. After segmentation is performed, patient’s clinical data such as
the area of the cortex, size of tumor, type of tumor (malignant or benign), and position
of tumor are determined which help a [1] doctor to take early decisions for surgery
or treatment to cure any brain disease.
During initial days, these segmentation techniques were performed manually by
subject matter experts or neurological experts, which consumes time and effort of
neurological specialists in the field. The segmentation results obtained [2, 3] from the
manual segmentation techniques may not be accurate due to vulnerable and unsat-
isfactory human errors which may lead to inappropriate surgical planning. There-
fore, it has become very much necessary for a neurologist or an academician or a
researcher to introduce automatic segmentation techniques which give accurate seg-
mentation results. These segmentation techniques that are performed automatically
are of two types typically known as semiautomatic and fully automatic segmentation
techniques. In a semiautomatic segmentation process, partial segmentation is per-
formed automatically, and then, the results thus obtained are checked by neurological
experts to modify for obtaining final segmentation results. In a fully automatic seg-
mentation technique, there is no need for manual checking by neurological experts
which minimizes his time and effort. These fully automatic segmentation techniques
are classified as threshold-based, region-based, pixel classification based, and model-
based techniques which are determined by the computer without any human partic-
ipation.
In this paper, regions in the brain are segmented automatically using a technique
called Fuzzy C-Means (FCM) algorithm, which is a pixel classification technique
followed by component labeling technique which is used widely in biomedical image
processing to perform fully automatic segmentation in brain MRI images. This clus-
tering mechanism is the most widely used technique for segmentation and detection
of tumor, lesions, and other abnormalities in brain MRI scans. The above pixel classi-
fication technique gives accurate results especially while analyzing non-homogenous
and dissimilar patterns of [4–7] brain MRI images. FCM is a unique method that
can be implemented in most of the MRI images to perform segmentation and obtain
efficient results even for the noisy MRI images. The main concept of clustering algo-
rithm is grouping of similar components (in this research, it is pixels of an image)
within the same cluster. This simple idea is implemented in this work to develop
a disease prediction framework that can automatically segment various regions of
multidimensional brain MRI scans.

[email protected]
A Computational Segmentation Tool for Processing … 3

2 Automatic Segmentation

Recent studies have shown that the atrophy rate in the brain is the valid parameter
to measure the severity of diseases such as dementia, Alzheimer’s and other brain
disorders from brain MRI images. Therefore, the necessity to calculate the atrophy
rate has been increased which is the measurement of abnormalities in gray and white
matter regions in brain MRI images of the patient [8]. The method herein presents a
disease prediction framework that can automatically segment gray and white matter
regions of patient’s brain using modified adapted pixel clustering method. In the
proposed method, the gray and white matter regions of cerebral structures are auto-
matically segmented using a form of adaptive modified pixel clustering technique
called Fuzzy C-Means (FCM) in which the pixels having similar intensity values are
grouped [9] into similar clusters and followed by connected component labeling in
which each pixel of gray and white matter regions are labeled.
A. Image Acquisition
The patient’s brain MRI image and neurological data used in this research work
was obtained from the Image and Data Archive (IDA) powered by Laboratory
of Neuro Imaging (LONI) provided by the University of Southern California
(USC) and from the Department of Neurosurgery at the All India Institute of
Medical Sciences (AIIMS), New Delhi, India. The data was anonymized as well
as followed all the ethical guidelines of the participating research institutions.
B. Segmentation Methodology
The segmentation methodology used in this research for automatically perform-
ing segmentation of gray and white matter regions in brain MRI images using
fuzzy c-means clustering algorithm is shown as a block diagram in Fig. 1. The
preliminary step in the process of segmentation is to remove the external sections
of the image which is not required for brain MRI image analysis. Therefore,
it is necessary to detect and remove the skull outline from the patient’s brain
MRI image. This mechanism is performed using elliptical Hough transform
which is used in digital image processing applications that identify the arbitrary
shapes such as circles, ellipses, and lines in an image data. After the skull outline
removal, the inner brain slice is subjected to adapted fuzzy c-means clustering
algorithm which is one of the pixel classification techniques mentioned above.
In this process, the brain internal slice is separated into different regions using
clustering mechanism which is based on the intensity values of the pixels in this
research.
Among the above-described pixel classification segmentation techniques,
clustering-based fuzzy c-means algorithm is used for segmentation of gray and
white matter regions in this research. Also, this technique generated accurate,
reliable, and robust results even with the noisy MR images of patients’ brain.
After clustering, the next step in the segmentation process is to perform con-
nected component labeling based on the connectivity of the neighboring pixels.
Even after performing clustering, some of the pixels positioning adjacent to
each other different similar intensity values of pixels may be in the same clus-

[email protected]
4 A. Goyal et al.

Flowchart of Automatic Segmentation Methodology

Fig. 1 Block diagram of automatic gray and white matter segmentation

ter. Therefore, it is necessary to perform connected component labeling in which

each component of the image is labeled and given a membership to each of the
pixels in the clusters to determine and differentiate the pixels accurately. Also,
among many other techniques, this modified adaptive pixel classification tech-
nique called fuzzy c-means is used for both multi-featured and single featured
extraction and analysis using spatial data. The segmentation technique [8, 9]
used in this research is fully automatized unsupervised segmentation that can
perform feature analysis, clustering in many medical imaging applications. A
medical image data is formed with the combination of set of components or
data points that have similar or dissimilar parametric values. These similar and
dissimilar data points of the image are classified into various similar clusters
which can be performed based on similarity criteria. Image pixels of medical
image data can be correlated to each other which have similar characteristics
or feature information to the data points that are sitting next to the data point in
an image. In this segmentation mechanism, spatial data of the adjacent pixels
is taken to perform clustering. This research work presents an algorithm for
clustering of various regions of brain MRI images into various classes followed
by connected component labeling using a knowledge-based algorithm. Below
are the steps for fully automatic segmentation algorithm:
(a) Skull outline detection:

[email protected]
A Computational Segmentation Tool for Processing … 5

Skull Detection Methodology

Fig. 2 Skull outline detection in brain MRI images

Skull outline present in the brain MRI scan is not required for analyzing brain
abnormalities in this research. Therefore, detection of skull outline and removing
it play a vital role during the segmentation process so that feature extraction and
analysis becomes easier and results thus obtained without the skull outline part
will be accurate. This skull outline sections in brain MRI regions are not our
region of interest as this section are filled with fat, skin, and other unwanted
materials. This step allows a researcher to focus more on the actual brain sections
which are responsible for brain disorders and to obtain reliable outputs [10]. In
this skull removal process, a widely used image feature extraction tool in digital
image processing is used to detect the superfluous components of the brain
MRI image data with in different shapes such as circle, ellipse, and lines. In this
research, we have used the elliptical Hough transforms to extract the unwanted
material (data objects or pixel components of an MRI image) from the actual
brain MRI image data. This elliptical Hough transform is applied to the original
brain MRI slice using a voting process in a parametric space [11]. Figure 2
shows the results obtained in the first step of the segmentation process:
(b) Adaptive fuzzy c-means clustering:

[email protected]
6 A. Goyal et al.

Once the skull outline sections are detected and removed from the original
brain MRI image, the next and very important step in the segmentation process
is to perform clustering to the image that is obtained from the first step of the
segmentation process. In this clustering, the medical image is classified into
various regions of brain such as gray matter, white matter, and cerebrospinal
fluid. The concept of clustering helps a researcher especially in digital image
processing technique to classify different patterns of the image and for the
segmentation of any medical image data. It is a widely used technique for
various purposes for medical data analysis in the field of medical sciences. The
process of classifying different clusters by grouping the similar components into
same cluster based on some criteria is defined as clustering. In this research,
clustering of the medical MRI image data having different regions such as gray
matter and white matter is performed based on similar intensity values of the
pixels. Due to several internal and external parameters, patient’s MRI scans in
the field of biomedical sciences may have more noise which when analyzed
further may produce inappropriate results [12]. This is highly unacceptable as
these inappropriate results may lead to improper diagnosis and surgical planning
of the patient. And hence, an effective algorithm is required to avoid inaccurate
results during the segmentation process. There are several types of clustering
techniques available in the field to perform segmentation of brain MRI images
in the medical field. In this research, we have used a modified pixel classification
technique called fuzzy c-means, which is based on the clustering mechanism.
This technique that is used for segmentation generates accurate results equally
for noisy MRI patient data [13–18]. Among many other clustering algorithms,
fuzzy c-means algorithm is the most popular technique which has a wide number
of benefits comparatively as it performs well even with the uncertain medical
image data. This technique used in our research enhances the features of fuzzy
c-means algorithm minimizing computational errors during the segmentation
process and this modified algorithm is called adaptive fuzzy c-means clustering
algorithm [19].
(c) Connected component labeling:
The next step in the proposed automatic segmentation is to perform connected
component labeling to the clustered image based on pixel connectivity mecha-
nism. In this stage, positions of several pixels which are located on the clustered
image are extracted and classified. In this process, several disjoint and also con-
nected components are labeled based on the connectivity procedures, which
is a very essential step in the segmentation process in order to reduce inaccu-
racy [20]. Every medical MRI image consists of pixels that are located side
by side sitting together forming connected components will have similar inten-
sity values. Therefore, in this method, the image is scanned such that every
pixel is detected and examined to extract the connected pixel regions of the
MRI image that are positioned adjacent to each other having similar intensity
values [21–25]. Each and every pixel component of the image irrespective of
which group it belongs to are labeled based on the connectivity of pixels. In this
research, connected component labeling is performed using two-dimensional

[email protected]
A Computational Segmentation Tool for Processing … 7

eight-connectivity measures to determine the way in which each pixel is related

to its neighboring pixel in the medical MRI image.
(d) Final segmentation mask after removing noise:
Last but not the least, the step after skull outline detection, clustering, and
connected component labeling is to generate the required segmented gray and
white matter regions by superimposing gray and white matter masks on the
actual brain MRI image that has to be analyzed. Our major goal in this step is to
remove all background pixels and to only keep foreground pixels of region of
interest of the original MRI image [25–28]. This process of overlaying masks
on original brain MRI image and removing the background pixels from the
image improves the segmentation process by further increasing the quality of
separability of gray and white matter regions, and thus, accurate segmentation
results are obtained. The results obtained by the process of clustering using
fuzzy c-means followed by connected component labeling to extract gray and
white matter regions as masks and when these masks are further processed for
final segmentation of gray and white matter regions are shown in Fig. 3.

3 Graphical Computational Tool

A software tool is developed that can automatically perform the entire process of
feature extraction, classification, preprocessing, and segmentation as an effective
graphical computational tool with a user interface (GUI). This application is inde-
pendent and a standalone GUI application that can be installed on the users’ machine
acting as a desktop application. Neurologists or any user can load the brain MRI
image from his local machine and perform automatic segmentation to obtain various
results instantaneously. This automated segmentation tool can perform segmentation
and display the results as mask, color images, or boundaries of gray and white matter
regions of brain MRI image with just clicks of buttons that takes very less amount
of time and efforts of neurologists. The developed GUI system assists neurologists
or any user making it easy to upload patient’s brain image from his local computer,
viewing and obtaining the results in very less time reducing efforts due to manual
tracings [29–33] by the experts. The GUI has the following features:
(1) Segmentation of brain MRI images is provided as a software.
(2) It is freely accessible to all researchers in the medical field and neurologists,
radiologists, and doctors in any part of the world.
(3) It is user-friendly and easy to use.
(4) It automatically segments the brain images and so no manual tracing is required
by the user.
(5) It supports all medical image data types (nifty, dicom, png, etc.).
(6) Neurologists disease prediction framework is provided in this software tool.
(7) Automatically calculates the areas of gray and white matter regions or lesions
or tumors based on which it predicts disease in brain.

[email protected]
 8

Automatic Gray and White Matter Segmentation Methodology

[email protected]
Fig. 3 Gray and white matter segmentation in brain MRI images
A. Goyal et al.
A Computational Segmentation Tool for Processing … 9

GUI of Computational Software for Automatic Brain MRI Segmentation

(a) (b)

(c) (d)

Fig. 4 Graphical User Interface (GUI) of computational brain segmentation software developed
to a load MRI image, b gray and white matter regions, c gray and white matter masks, and d gray
and white matter boundaries

Below are the three screenshots which show running the GUI for loading the brain
MRI image (Fig. 4a), viewing the gray and white matter segmented regions (Fig. 4b),
viewing the gray and white matter extracted masks (Fig. 4c) and viewing the gray
and white matter region boundaries (Fig. 4d).

4 Manual Segmentation

In this research work, the manual segmentation is performed for several patients’
MRI image data to validate and verify the automatic segmentation techniques using
fuzzy c-means followed by connected component labeling performed in this research
with the manual segmentations done by neurological tracings of gray matter and
white matter regions by experts [34–38]. Figure 5 presents the manual segmentations

[email protected]
10 A. Goyal et al.

Manual Segmentation Illustration

Fig. 5 Manual segmentation (labeling) by neurologist expert of the gray and white matter regions
in brain MRI images. Gray matter region (left) and white matter region (right)

performed by neurological experts to segment gray and white matter regions of brain
MRI images.

5 Validation

This research work presents the comparison of manual and automatic segmentation
results of five different patients’ sample MRI images. To compare these results, a
statistical measure called Dice coefficients is used to calculate the similarity mea-
sures of these two techniques. Figure 6a–c shows the sample manual and automatic
segmentation of three of the patients. The automatic segmentations are obtained from
the proposed algorithm, and the manual segmentation is obtained here from the neu-
rological tracings by experts in this research. For the validation purpose, five different
sample image data are considered, and the manual and automatic segmentations are
performed for the [39–42] same to compare both the segmentation results that are
obtained. For each of the patient MRI images, a dice coefficient value is calculated
between manual and automatic segmentation of patient brain MRI images. Manual
segmentation is performed three times by neurological experts for each of the sample
patient images among the five different MRI images. Finally, the Dice coefficient
values are plotted using box plots for each of the patient brain images that com-
pare manual and automatic segmentations for all sample patient images considered.
Figure 7 shows the box plots of the Dice coefficients calculated as the similarity

[email protected]
A Computational Segmentation Tool for Processing … 11

Fig. 6 Sample manual and automatic segmentation of three specimens

[email protected]
12 A. Goyal et al.

Fig. 6 (continued)

measure to compare manual and automatic segmentation of the brain MRI images
for the five sample patients.

6 Discussion and Conclusion

The research presented in this work facilitates efficient and effective automatic seg-
mentation of gray and white matter regions from brain MRI images, which has
several clinical neurological applications. A fully automatic segmentation method-
ology using elliptical Hough transform along with pixel intensity and membership-
based adapted fuzzy c-means clustering followed by connected component labeling
and region analysis has been implemented in this research to perform segmentation
of gray and white matter regions in brain MRI images. The algorithm was tested
and verified for several sample brain MRI images. Manual segmentations were per-
formed by neurological experts for several patient brain MRI images. These manual
segmentations were used to compare and validate with the results obtained from the
automatic segmentations in this research work. Validations were performed by calcu-
lating several Dice coefficient values between the automatic segmentation results and
the manual segmentation results. The Dice coefficient values are similarity measures

[email protected]
A Computational Segmentation Tool for Processing … 13

Validation: Dice Coefficients as Similarity Measures

Fig. 7 Box plots for Dice coefficients to compare manual and automatic segmentation of brain
MRI images of five patients

that are represented statistically using box plots in this research work. The automa-
tized computational segmentation tool developed in this research can be employed in
hospitals and neurology divisions as a computational software platform for assisting
neurologist in detection of disease from brain MRI images post MRI segmentation.
This tool obviates manual tracing and saves the precious time of neurologists or
radiologists. This research presented herein is foundational to a neurological dis-
ease prediction and disease detection framework, which in the future, with further
research work, can be developed and implemented with a machine learning model-
based prediction algorithm to detect and calculate the severity level of the disease,
based on the gray and white matter region segmentations and estimated gray and
white matter ratios to the total cortical matter, as outlined in this research.

References

1. Shasidhar, M., Raja, V. S. and Kumar, B. V., (June 2011). MRI brain image segmentation using
modified fuzzy c-means clustering algorithm. In 2011 International Conference on Communi-

[email protected]
14 A. Goyal et al.

cation Systems and Network Technologies (CSNT) (pp. 473–478).

2. Despotović, I., Goossens, B., Philips, W. (2015) MRI segmentation of the human brain: chal-
lenges, methods, and applications. Computational and mathematical methods in medicine,
2015.
3. Gordillo, N., Montseny, E., & Sobrevilla, P. (2013). State of the art survey on MRI brain tumor
segmentation. Magnetic Resonance Imaging, 31(8), 1426–1438.
4. Suhag, S., & Saini, L. M. (May 2015). Automatic detection of brain tumor by image processing
in matlab. In SARC-IRF International Conference.
5. Hassan, E., & Aboshgifa, A. (2015). Detecting brain tumour from MRI image using matlab
gui programme. International Journal of Computer Science & Engineering Survey (IJCSES)
6(6).
6. Sharma, N., & Aggarwal, L. M. (2010). Automated medical image segmentation techniques.
Journal of medical physics/Association of Medical Physicists of India, 35(1), 3.
7. Tsai, C., Manjunath, B. S., & Jagadeesan, B. (1995). Automated segmentation of brain MR
images. Pattern Recognition, 28(12), 1825–1837.
8. Despotović, I., Goossens, B., & Philips, W. (2015). MRI segmentation of the human
brain: challenges, methods, and applications. Computational and mathematical methods in
medicine, 2015.
9. Chuang, K. S., Tzeng, H. L., Chen, S., Wu, J., & Chen, T. J. (2006). Fuzzy c-means clus-
tering with spatial information for image segmentation. Computerized medical imaging and
graphics, 30(1), pp. 9–15.
10. Mokbel, H. A., Morsy, M. E. S., & Abou-Chadi, F. E. Z. (2000). Automatic segmentation and
labeling of human brain tissue from MR images. In 17th NRSC’2000. Seventeenth National
Radio Science Conference, 2000 (pp. K2–1). IEEE.
11. Antolovic, D., (2008). Review of the Hough transform method, with an implementation of the
fast Hough variant for line detection. Department of Computer Science, Indiana University.
12. Kumar, N., & Nachamai, M. Noise Removal and filtering techniques used in medical images.
Oriental Journal of Computer Science and Tachnology 10(1).
13. Wang, H. R., Yang, J. L., Sun, H. J., Chen, D., & Liu, X. L. (August 2011). An improved region
growing method for medical image selection and evaluation based on Canny edge detection. In
2011 International Conference on Management and Service Science (MASS) (pp. 1–4). IEEE.
14. Mubarak, D. M. N., Sathik, M. M., Beevi, S. Z., & Revathy, K. (2012). A hybrid region
growing algorithm for medical image segmentation. International Journal of Computer Science
& Information Technology, 4(3), 61.
15. Wong, K. K., Tu, J., Kelso, R. M., Worthley, S. G., Sanders, P., Mazumdar, J., et al. (2010).
Cardiac flow component analysis. Medical Engineering & Physics, 32(2), 174–188.
16. Zanaty, E. A. (2013). An Approach based on fusion concepts for improving brain magnetic
resonance images (MRIs) segmentation. Journal of Medical Imaging and Health Informatics,
3(1), 30–37.
17. Zanaty, E. A., & Ghiduk, A. S. (2013). A novel approach for medical image segmentation based
on genetic and seed region growing algorithms. Journal of Computer Science and Information
Systems ComSIS, 10(3), 1319–1342.
18. Zanaty, E. A., & Afifi, A. (2013). A watershed approach for improving medical image segmen-
tation. Computer methods in biomechanics and biomedical engineering, 16(12), 1262–1272.
19. Zanaty, E. A. (2013). An adaptive fuzzy C-means algorithm for improving MRI segmentation.
Open Journal of Medical Imaging, 3(04), 125.
20. [Online] Available: https://en.wikipedia.org/wiki/Connected-component_labeling [Accessed
November 9, 2017].
21. Wu, K., Otoo, E., & Shoshani, A. (2005). Optimizing connected component labeling algorithms.
Lawrence Berkeley National Laboratory.
22. Suzuki, K., Horiba, I., & Sugie, N. (2003). Linear-time connected-component labeling based
on sequential local operations. Computer Vision and Image Understanding, 89(1), 1–23.
23. Goyal, A., Lee, J., Lamata, P., van den Wijngaard, J., van Horssen, P., Spaan, J., et al. (2013).
Model-based vasculature extraction from optical fluorescence cryomicrotome images. IEEE
Transactions on Medical Imaging, 32(1), 56–72.

[email protected]
A Computational Segmentation Tool for Processing … 15

24. Sikarwar, B. S., Roy, M. K., Ranjan, P., & Goyal, A. (2016). Automatic Disease Screening
Method Using Image Processing for Dried Blood Microfluidic Drop Stain Pattern Recognition.
Journal of Medical Engineering & Technology, 40(5), 245–254.
25. Sikarwar, B. S., Roy, M. K., Ranjan, P., & Goyal, A. (2016). Imaging-based method for precur-
sors of impending disease from blood traces. In Advances in Intelligent Systems and Computing
(Vol. 468, pp. 411–424). Springer.
26. Sikarwar, B. S., Roy, M. K., Ranjan, P., & Goyal, A. (2015). Automatic pattern recognition
for detection of disease from blood drop stain obtained with microfluidic device. In Advances
in Intelligent Systems and Computing (Vol. 425, pp. 655–667). Springer.
27. Bhan, A., Bathla, D., & Goyal, A. (2016). Patient-specific cardiac computational modeling
based on left ventricle segmentation from magnetic resonance images. Advances in Intelligent
Systems and Computing (Vol. 469, pp. 179–187). Springer.
28. Ray, V., & Goyal, A. (2015) Automatic left ventricle segmentation in cardiac MRI images using
a membership clustering and heuristic region-based pixel classification approach. In Advances
in Intelligent Systems and Computing (Vol. 425, pp. 615–623). Springer.
29. Chhabra, M., & Goyal, A. (2017) Accurate and robust iris recognition using modified classical
hough transform. In Lecture Notes in Networks and Systems (Vol. 10, pp. 493–507). Springer.
30. Goyal, A., & Ray, V. (2015). Belongingness clustering and region labeling based pixel clas-
sification for automatic left ventricle segmentation in cardiac MRI images. Translational
Biomedicine, 6(3).
31. Goyal, A., Roy, M., Gupta, P., Dutta, M. K., Singh, S., & Garg, V. (2015) Automatic detection
of mycobacterium tuberculosis in stained sputum and urine smear images. Archives of Clinical
Microbiology, 6(3).
32. Bhan, A., Goyal, A., Chauhan, N., & Wang, C.W. (2016) Feature line profile based automatic
detection of dental caries in bitewing radiography. In: International Conference on Micro-
Electronics and Telecommunication Engineering (ICMETE), pp. 635–640, IEEE.
33. Bhan, A., Goyal, A., Dutta, M. K., Riha, K., Omran, Y. Image-Based Pixel Clustering and
Connected Component Labeling in Left Ventricle Segmentation of Cardiac MR Images. In
7th International Congress on Ultra-Modern Telecommunications and Control Systems and
Workshops (ICUMT), pp. 339–342, IEEE, 2015.
34. Ray, V., & Goyal, A. (2015). Image-Based fuzzy c-means clustering and connected component
labeling subsecond fast fully automatic complete cardiac cycle left ventricle segmentation in
multi frame cardiac MRI images. In International Conference on Systems in Medicine and
Biology (ICSMB), IEEE.
35. Goyal, A., van den Wijngaard, J., van Horssen, P., Grau, V., Spaan, J., & Smith, N. (2009).
Intramural spatial variation of optical tissue properties measured with fluorescence microsphere
images of porcine cardiac tissue. In Annual International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC), pp. 1408–1411.
36. Sharma, P., Sharma, S., Goyal, A. (2016) An MSE (mean square error) based analysis of
deconvolution techniques used for deblurring/restoration of MRI and CT Images. In 2nd
International Conference on Information and Communication Technology for Competitive
Strategies (ICTCS-2016), March 04–05, 2016, Udaipur, India, Conference Proceedings by
ACM—ICPS Proceedings Vol. ISBN 978-1-4503-3962-9/16/03, http://dx.doi.org/10.1145/
2905055.2905257.
37. Goyal, A., Bathla, D., Sharma, P., Sahay, M., & Sood, S. (2016). MRI image based patient
specific computational model reconstruction of the left ventricle cavity and myocardium.
In 2016 International Conference on Computing, Communication and Automation (ICCCA),
pp. 1065–1068, IEEE.
38. Duta, M., Thiyagalingam, J., Trefethen, A., Goyal, A., Grau, V., & Smith, N. (2010) Parallel
simulation for parameter estimation of optical tissue properties. In Euro-Par 2010-Parallel
Processing (pp. 51–62).
39. Atkins, M. S., & Mackiewich, B. T. (1998). Fully automatic segmentation of the brain in MRI.
IEEE Transactions on Medical Imaging, 17(1), 98–107.

[email protected]
16 A. Goyal et al.

40. Wagner, M., Yang, P., Schafer, S., Strother, C., & Mistretta, C. (2015). Noise reduction for
curve-linear structures in real time fluoroscopy applications using directional binary masks.
Medical Physics, 42(8), 4645–4653.
41. Meijs, M., Patel, A., Leemput, S. C., Prokop, M., Dijk, E. J., Leeuw, F. E., et al. (2017). Robust
segmentation of the full cerebral vasculature in 4D CT of suspected stroke patients. Scientific
reports, 7(1), 15622.
42. Bhan, A., Goyal, A., & Ray, V. (2015) Fast fully automatic multiframe segmentation of left
ventricle in cardiac mri images using local adaptive k-means clustering and connected compo-
nent labeling. In 2nd International Conference on Signal Processing and Integrated Networks
(SPIN), pp. 114–119, IEEE.

[email protected]
Developing Ontology for Smart
Irrigation of Vineyards

Archana Chougule and Debajyoti Mukhopadhyay

Abstract As the availability of groundwater is getting reduced these years, proper

scheduling of irrigation is very important for survival and improvement of vineyards
in the southern part of India. Well calculated irrigation scheduling also improves
quality of grapes. The knowledge about irrigation to vineyards is available in various
documents, and it is scattered. This knowledge can be made available to grape grow-
ers through computer-based applications. As Semantic Web is playing an important
role, information sharing among varying automated systems, extraction of informa-
tion from such documents and representing it as ontology is a good idea. This paper
presents techniques for automated extraction of knowledge from text resources using
natural language processing technique for building vineyard ontology. Smart irriga-
tion system can be developed using IoT sensors and other ICT devices. The paper
explains ontology built for resources used under smart irrigation system. It suggests
how smart irrigation systems can be built by grape growers by utilizing vineyard
ontology and smart irrigation ontology.

Keywords Ontology building · Irrigation scheduling · Grapes · Knowledge base ·

Natural language processing

1 Introduction

Water is the most important resource in agriculture. Irrigation scheduling is basically

deciding on the frequency and duration of water supply to any crop. Proper scheduling
of irrigation can minimize wasting water and can also help to improve the quality of
grapes. There are different methods of irrigation. Drip irrigation is used by almost all
vie yards in India. Manerajuri is a village in India, which is famous for grapes, but

A. Chougule (B)
Walchand College of Engineering, Sangli, India
e-mail: [email protected]
D. Mukhopadhyay
NHITM, Mumbai University, Mumbai, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 17

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_2

[email protected]
18 A. Chougule and D. Mukhopadhyay

more than 10,000 bore wells have run dry in year 2016. This fact motivated authors
to develop the knowledge base and decision support system which can be used for
micromanagement of available water.
Knowledge base generation for scheduling water supply to grapes in hot tropical
region in India will help automation of water scheduling. Good irrigation practices
in ontology will be good education material for farmers. Structured representation of
knowledge is more useful than unstructured one. Ontologies can be used for struc-
tured representation of concepts in any domain. Along with concepts, relationships
between concepts can also be mentioned using ontology. Maintenance and sharing
of information can be facilitated using ontologies. Manual construction of domain
ontology is time-consuming and error-prone. Semi-automated approach will help
in enriching ontology building, reducing required time and enhancing the quality of
built ontology. For IoT-based automation of irrigation systems, formal representation
and management of IoT techniques and devices used important.
Automated irrigation system needs information like ambient temperature and
humidity, atmospheric pressure, soil temperature, soil moisture and leaf wetness.
System can be built with such information using sensors, actuators, monitors, col-
lectors and transmitters. Data generated and required from such varying sources can
be easily integrated using ontology [1].
This paper describes building ontology for developing IoT-based automated irri-
gation systems. It details mechanism used for constructing ontology. Two types of
ontologies, namely, vineyard irrigation ontology and smart irrigation ontology, are
described. These ontologies can be used for generating knowledge base and estab-
lishment of automated irrigation system based on available resources. It can also be
used to educate grape growers about principles of vineyard irrigation.

2 Research Method

Building ontology consists of five basic steps as defining domain, scope and objec-
tive of ontology, listing important keywords from the domain, finding hierarchies
between keywords, defining relations between them and the last step is to add those
keywords as concepts in ontology and define relationships between concepts. Various
formats for building ontology are available as RDF, RDF-S, OWL and OWL-DL.
Here ontology is developed in Web Ontology Language (OWL) format. Classes,
individuals, data properties, object properties and axioms are the main parts of OWL
ontology. Classes represent core concepts in the domain. Individuals are instances
of any specific class. Data properties are attributes of a class with specific values
and object properties hold values as objects of some other class. Two ontologies are
maintained separately as vineyard ontology and smart irrigation ontology.

[email protected]
Developing Ontology for Smart Irrigation of Vineyards 19

2.1 Building Vineyard Irrigation Ontology

Water management is also important to maintain yield and quality of grapes. This
section details how water management knowledge is represented in ontology form.
Knowledge about irrigation scheduling of grapes is available in various documents
published by researchers from organizations like National Research Center for
Grapes. To make use of this knowledge in decision support system, it must be trans-
formed to accessible form. As mentioned earlier, knowledge stored in ontology can
be accessed and shared by automated systems. This section discusses converting the
irrigation scheduling details available in text documents to ontology. As a reference
ontology, AGROVOC [2] is used. AGROVOC is an agricultural vocabulary available
as ontology. It is available in resource description framework (RDF) format.
The documents for irrigation scheduling are taken as input and the most relevant
documents are only considered. Relevancy of documents is ranked using TF-IDF
[3] algorithm. TF-IDF stands for term frequency, inverse document frequency. List
of keyword related to agriculture taken from agriculture experts is considered as
input for this purpose. Term frequency is considered as count of word occurring in
document and inverse document frequency is the count of documents containing the
word versus the total number of documents. Natural language processing techniques
are used then to extract important keywords from the text. The text is converted to
a number of tokens using StringTokenizer. The text is broken into words, phrases
and symbols under tokenization. Tokenization helps in finding meaningful keywords
from the text.
After tokenization, stop words are removed from the list of tokens. Stop words
are words which are used for joining words together in a sentence. They occur very
frequently in documents, but are of no importance for ontology building. Stop words
like ‘and’, ‘or’, ‘this’ and ‘the’ are not used for the classification of documents,
so they are removed. The process of conflating the variant forms of a word into
a common representation is called stemming. There can be the same words with
different forms. Such words are converted to common form using Porter’s stemmer
[4]. For example, the words: ‘irrigation’, ‘irrigated’ and ‘irrigating’ are reduced to
a common representation ‘irrigate’. OpenNLP [5] POS tagger is used for extracting
important keywords from tokens. POS tagger assigns labels to keywords as NN for
noun and NNP as proper nouns. Assigned labels are then considered for probable
classes, data properties, individuals or object properties.
Along with concepts relationships between concepts are also found by process-
ing sentences. Formal concept analysis [6] is used for deciding hierarchy among
concepts. Formal concept analysis is a technique used for defining conceptual struc-
tures among data sets [7]. Extracted individuals, data properties and object properties
are used for this purpose. The top-down approach is used here. Extracted concepts
are stored in the concept table. Extracted relations are used for defining hierarchy
among classes and among properties. For example, Table 1 has five concepts and
four attributes, showing which concepts carry which corresponding attributes, and
Fig. 1 shows corresponding concept lattice.

[email protected]
20 A. Chougule and D. Mukhopadhyay

Table 1 Example of formal Amount Unit Timestamp Level

concept analysis
Sensor X X
Irrigation X X X
Groundwater X X X X
Surface water X X
Soil moisture X X X

Fig. 1 Concept lattice corresponding to context in Table 1

The ontology is built using ProtegeOWL APIs [8] available in java language.
The fundamental concepts related to irrigation are added as classes in ontology. The
ontology is built using the top-down approach. More general term is added at higher
level in the hierarchy, followed by more specific concepts at lower levels.
Vineyard ontology contains 16 classes, 56 individuals, 12 object properties and
19 data properties. The built ontology is further edited using Protégé 5.1 ontology
editor. Protégé editor has facilities to add, edit and delete concepts, change hierarchy,
view ontology and use reasoners on ontology. The irrigation knowledge is extracted
from research documents published by national research for grapes, Pune, India [9].
Irrigation schedule can be decided based on knowledge stored in vineyard irriga-
tion ontology in terms of individuals, object properties and data property values.
Requirements for irrigation for grapes are mentioned as individuals of Irrigation-
Schedule class. Under each object, water requirement is specified as data property
value. Growth stage-wise water requirements of grapes are considered and mentioned
for each irrigation schedule.
Figure 2 shows classes under vineyard ontology. Water-Shortage-Symptoms class
stores all symptoms shown on grape leaves, stems, berries and in soil due to the

[email protected]
Developing Ontology for Smart Irrigation of Vineyards 21

Fig. 2 Classes in vineyard ontology

Fig. 3 Example of irrigation schedule defined in vineyard ontology as data and object property
assertions

shortage of water. Specific information about yard like root depth, soil type and
available water resources are covered by Yard-Details class. All stages of vine growth
are stored as individuals of Grape-Growth-Stage class and are referenced by irrigation
requirement schedule.
Irrigation-Requirement-Schedule is the most important class for modelling and
predicting irrigation requirements. It contains water requirement schedules as indi-
viduals detailing days after pruning of schedule, water requirement quantity in litres
per day per hectare, growth stage of vines, irrigation period in terms of month
and evapotranspiration measure. Figure 3 shows the example of irrigation sched-
ule defined in vineyard ontology.

[email protected]
22 A. Chougule and D. Mukhopadhyay

Fig. 4 Sensor class in smart irrigation ontology

2.2 Smart Irrigation Ontology

In order to provide a knowledge base of tools and techniques that can be used for
automation of irrigation scheduling, Smart Irrigation ontology is built. It contains
all concepts including IoT, for creating and managing smart irrigation system for
vineyard. The time and amount of water to apply to vineyard are generally based on
four methods as monitoring of soil moisture levels, measuring water status in plants,
based on evapotranspiration measurement and based on the estimated vineyard water
use rate and soil water storage. As irrigation decision depends on various factors in
vineyard, sensors are used for such measurements. Figure 4 shows types of sensors
that can be used for smart irrigation. Unit-of-measure, type-of-sensing, locating-of-
sensing and sensor-description are data properties of the Sensor class. The ontology
provides all support for storing information about IoT-based tools. It also provides
data and object properties for storing measurements read by such tools. It comes
under devices and communication techniques classes and their sub-classes. Knowl-
edge about irrigation techniques that can be used under vineyards and possible water
resources comes under Irrigation-Techniques and Water-Resources classes. Obser-
vations class contains sub-classes, data and object properties for storing measure-
ments taken by sensors. System-Design and User-Interface are the classes suggesting
options for building irrigation automation system using available resources. As irri-
gation management depends on climate and weather details in specific region two
classes, namely, Climate and Weather-Details, are added in the ontology. Annual pre-
cipitation and annual rainfall are data properties of climate class. Rainfall, humidity
and temperature are parts of Weather-Details class. Figure 5 shows all classes under
smart irrigation ontology.

[email protected]
Developing Ontology for Smart Irrigation of Vineyards 23

Fig. 5 Classes in smart irrigation ontology

[email protected]
24 A. Chougule and D. Mukhopadhyay

3 Conclusion

Ontology plays an important role in IoT-based automation of agricultural systems.

The paper demonstrated how ontology can be built by using natural language pro-
cessing techniques and formal concept analysis. Paper demonstrated how irrigation
details about vineyards can be very well represented using ontology. The informa-
tion about sensors can also be stored in ontology and used for automated systems
building, which is shown in smart irrigation ontology. Based on given water require-
ments, the automated irrigation system can be built using knowledge from vineyard
ontology and smart irrigation ontology as proposed in the paper.

References

1. Cornejo, C., Beck, H. W., Haman, D. Z., & Zazueta, F. S. (2005). Development and application
of an irrigation ontology. 2005 EFITA/WCCA Joint Congress on IT and Agriculture, Vila Real,
Portugal, 25–28 July 2005.
2. AGROVOC Thesaurus. http://aims.fao.org/agrovoc#.VF29AvmUc2U.
3. TFIDF Algorithm. http://en.wikipedia.org/wiki/Tf-idf.
4. Porter, M. F. (1980). An algorithm for suffix stripping. Program, 14(3), 130–137.
5. OpenNLP. https://opennlp.apache.org/.
6. Obitko, M., Snasel, V., Smid, J. (2004). Ontology design with formal concept analysis. In V.
Snasel & R. Belohlavek (Eds.), CLA 2004, pp. 111–119, ISBN 80-248-0597-9. VSB—Technical
University of Ostrava, Deportment of Computer Science.
7. Ganter, B., & Wille, R. (1999). Formal concept analysis, mathematical foundation. Berlin:
Springer Verlag.
8. Protege-OWL API Programmer’s Guide. https://protegewiki.stanford.edu/wiki/ProtegeOWL_
API_Programmers_Guide [Accessed on 17 May 2017].
9. National Research Centre for Grapes. http://nrcgrapes.nic.in.

[email protected]
An Intensive Review of Data Replication
Algorithms for Cloud Systems

Chetna Dabas and Juhi Aggarwal

Abstract Cloud computing is a technological paradigm that facilitates a universal

platform for hosting and accessing services and data by means of a large number of
shared resources by an equally large number of users across the globe. Since a large
amount of data is involved, it becomes crucial to make the data available to the users
on the go without compromising with the integrity of data. Data replication is used
to make copies of the data and files such that they can be accessed without delay
and also act as a backup in case of any failure. Here we have discussed crucial data
replication algorithms, namely, adaptive data replication strategy (ADRS), high QoS-
first replication (HQFR), build time algorithm, cost-effective incremental replication
strategy (CIR) and dynamic replica control strategy (DRCS).

Keywords Quality of service · Hadoop file system (HDFS) · Replication factor

1 Introduction

Cloud has become increasingly popular as a data storage unit as many business
organizations and companies have decided to store their data in cloud data centres.
Some of the well-known cloud storage providers are the Amazon S3 [1], Google
cloud storage [2] and Microsoft Azure [3]. These storage providers meet the needs
of the companies by securely storing and analysing data while providing 99.9999%
durability in case of Amazon S3. The benefit of using such a system for storing data is
that data can be retrieved from anywhere at any time with help of a web application
or mobile phone. The services provided by these cloud service providers include
Software as a Service (SaaS), Platform as Service (PaaS) and Infrastructure as a
Service (IaaS). These services are provided as pay per use services, i.e. the customer
pays according to the amount of resources she/he uses of the cloud. Some of the
resources which are used by the client are bandwidth, storage space, networks and
servers, to name a few.

C. Dabas (B) · J. Aggarwal

Department of CSE, Jaypee Institute of Information Technology, Noida, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 25

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_3

[email protected]
26 C. Dabas and J. Aggarwal

The cloud infrastructure faces many challenges owing to a lot of factors such as
a large number of users, highly dispersed data centres, large numbers of data access
requests, heterogeneity of data and load balancing, among others [4]. Secure storage
of data is one of the major challenges faced by the cloud environment. Since the
data is stored in the public cloud, issues like data integrity, data confidentiality and
data privacy are of great concern. To deal with the aforementioned issues, cloud
platforms use data encryption algorithms to encrypt the data before storing it in
cloud as is done in [5] using Shamir’s secret sharing algorithm. Data availability and
reliability challenges are also related to cloud storage. Data availability means that the
data should be available with its agreed quality of service requirements irrespective
of any failure. Data reliability considers that the data is complete and error-free
whenever the client retrieves it. These issues of data availability and reliability are
well handled by two schemes: (1) erasure coding and (2) data replication. Erasure
coding [6, 7] partitions a data entity into k uniformly sized chunks in addition to
m extra chunks to hold the parity bit. The total (k + m  n) chunks are placed on
different nodes such that in case of failure data can be recovered using these parity
bits.
Data replication improves data availability and performance of the cloud storage
system by serving data request by various replicas placed at different locations either
at the same site or different sites. Replicating data onto different sites also provides as
a means for fault tolerance. Failures can be of the type disk failure, site failure or node
failure. Data replication provides the benefits that there is more access to the shared
data. Since availability is improved therefore the waiting times are reduced, and
hence, the latency is also reduced. For data replication, there are several factors which
need to be considered like how many copies of a data item should be maintained,
how often should the replication process be called, in case of failure from where will
the data be retrieved, for how long should the copy of a data unit be maintained.
The main objective of this entire data replication scheme should be to enhance the
performance while keeping the number of replicas as minimum, as more number of
replicas will increase the storage overhead.
In this work, we have analysed five different data replication algorithms for replica
creation, replica placement, replica replacement and replica management. Following
are algorithms that we have reviewed (1) adaptive data replication strategy (ADRS),
(2) high quality of service-first replication (HQFR), (3) build time algorithm, (4)
cost-effective incremental replication strategy (CIR) and (5) dynamic replica control
strategy (DRCS). These algorithms take different approaches for generating best
replicas while optimizing the time constraints.

2 Related Work

Data reliability has been explored in various literatures [8, 9] which use concepts
like Markov models, Bayesian approaches, Monte Carlo method and graph theory
for modelling service reliability. However, in [8], the optimal resource scheduling

[email protected]
An Intensive Review of Data Replication Algorithms … 27

strategies have been missed out for multiple computing intensive services. In [9],
load influence of the nodes has not been performed. In [10], however, data relia-
bility is examined in terms of variable disk failure rate. Herein, a proactive replica
checking for reliability (PRCR) strategy is applied such that reliability is achieved
by minimizing the number of replicas. This algorithm provides a reasonable solution
for creating minimum replicas as opposed to the traditional three-replica strategy
used by Hadoop file system [11]. In [10] the concept of location replicas needs to be
further addressed. The reliability of cloud storage over a period of time is expressed
as an exponential distribution with the failure rate λ which has its own disadvan-
tages. It is assumed that the reliability of data adopts memory-less property, i.e. the
reliability of data is same at time t + t 1 as it was at time t. If the storage duration
of a data item is small then no replication is needed but if it is long enough then
PRCR is applied. Here the storage nodes are periodically scanned and checked by a
unit called the replication manager. After the scanning operation, it is decided which
replica has to be sent for checking. This bears the issue of overhead.
Some other studies have discussed data replication in terms of bandwidth con-
sumption and energy consumption. Energy consumption in cloud storage data centres
is immensely big since a huge amount of data are stored and the systems run con-
tinuously to provide availability. In [12], database replication is done according to
the popularity of data. The architecture adopted in this study is a level architecture
having a central database at the top most level; data centres at the middle level and
racks at the bottom most level. The system is modelled for performing data replica-
tion for minimizing the bandwidth utilization, energy utilization and network delay.
Energy consumed by the servers while performing its operation an even when sitting
idle is considered along with the power consumption of different types of switches.
Bandwidth is utilized during data transfer operations, i.e. when the request for a
data replica is invoked by the client to the server and when the server produces the
response to the request in the form of a replica. Overhead is a consistent issue with
the approach used by the authors. Moreover, in [12], the test bed implementation is
not addressed in the proposed work of the authors.
In the work of [13], a graph-based approach is used for data replication. This work
employs a locality replica manager (LRM) which coordinates the tasks of the system
and takes care of the quality of service (QoS) requirements. Herein, data replication
considers the physical neighbourhood of blocks for the query. With each inquiry
request, a full graph is constructed for the blocks for which the inquiry is raised and
these graphs are hosted by some data nodes. The decision as to who will host the
inquiry graph is made based on parameters like the QoS requirement of the inquiry,
delay of the graph, probability of a graph being available, storage space and load of
the node on which graph is to be hosted. In this paper, a genetic algorithm based
approach is used for maintaining the availability and delay of the system at desired
level but this approach lacks efficiency. In [13], no real-time implementation of the
proposed work has been performed.

[email protected]
28 C. Dabas and J. Aggarwal

3 Data Replication Algorithms

This section presents an insight into five data replication algorithm providing data
replication and management functionalities. The main focus of these algorithms is
to ensure data availability and reliability by providing a minimum number of quality
of service satisfied replicas.

3.1 Adaptive Data Replication Strategy (ADRS)

Author, namely, Najme Mansouri in [13] has proposed a replica placement and
replica replacement strategy for improving the availability of data with low cost.
For the purpose of replica placement, five parameters are considered, namely, mean
service time, failure probability, load variance, latency and storage usage. The afore-
mentioned parameters are calculated as below:

m  
Acc(i, j)
Mean Service Time (MST)  ST(i, j) ∗ (1)
j1
AR(i)

In Eq. (1), ST(i, j) is the expected time a file serves on a data node, Acc(i, j) is
the access rate of read requests coming from a data node asking for a file on it and
AR(i) is the mean access rate.

Load variance (LV)  Acc(i, j) ∗ ST(i, j) (2)


m
Size(i)
Latency  1/ri ∗ d(i, j) ∗ ∗ Acc(i, j) (3)
j1
B( j)

Using the above equations, a cost function is calculated for each site. Smaller the
value of this cost function, the better is its fitness value. The new replica is placed at
a site with minimum cost function value.

Cost function (CF)  w1 × MST + w2 × LV + w3 × SU + w4 × FP + w5 × L

(4)

where SU is the storage unit and FP is the failure probability and w1, w2, w3, w4,
w5 are weight which are randomly assigned.
For replica replacement, the parameters considered are availability of file, last
time the replica was requested, number of accesses of the replica and file size of
replica. Based on these parameters, a V value is calculated and the replica whose V
value is the least is replaced. The V value is calculated as below:

[email protected]
An Intensive Review of Data Replication Algorithms … 29

V  (w1 ∗ NA + w2 ∗ P)/(w3 ∗ (CT − LA) + w4 ∗ S) (5)

where NA is the number of access, CT is the current time, LA is the last request time
of replica and P is availability of file [14]. The working of this algorithm is further
explained conceptually and in detail in Fig. 1.
In this work, the number of times the replication process need to be called is
reduced because only when there is not enough space on the best site, then a replica
is created. Response time to access the data is reduced because it is available in the
best site and is easily available. Also only read-only data facility is provided which
limits the capability of the user to write any changes to the data. No file partitioning
is done thereby causing entire data to be stored at one storage location which can
lead to high storage space cost.

3.2 Build Time Algorithm

This algorithm considers that data has already been placed at the right data centres by
using some existing replication algorithm. Data in this algorithm is considered to be
datasets distributed over data centres. The data sets are classified into two types: free
flexible dataset (FFD) and constraint flexible dataset (CFD). The size of the storage
space decides which category the data set belongs to. If the size of the data set is less
than storage space it belongs to free flexible dataset category; otherwise, it belongs
to constrained flexible dataset. The storage unit itself is divided into two, i.e. original
data space and replicated data space. The dependency between two datasets and
access frequency are the parameters considered for this algorithm. The dependency
between two datasets is given as
    
Dep Di , D j  Count T (di ) ∩ T d j (6)

In the above equation, T (d i ) represents the number of tasks which belong to that
dataset. Access frequency is frequency with which a dataset is accessed, which is
defined as follows:

Acc f di  AccNumdi /times (7)

Here, AccNumdi represents the number of times a dataset is accessed and times
is the time interval of data transfer [15]. Figure 2 gives the working flow of this
algorithm in a conceptual manner.
The strategy is compared with a normal data replication scheme and observed a
substantial decrease in total cost. Transfer cost is reduced because a dataset is placed
at a place where the task it is used for is located. The algorithm does not discuss
how the tasks are distributed among different data centres but it only discusses the
datasets.

[email protected]
30 C. Dabas and J. Aggarwal

Fig. 1 Working of ADRS algorithm

[email protected]
An Intensive Review of Data Replication Algorithms … 31

Fig. 2 Working of build time algorithm

[email protected]
32 C. Dabas and J. Aggarwal

3.3 High QoS-First Replication Algorithm

In the QoS aware data replication (QADR) problem [16] the HQFR algorithm is used
for producing QoS aware data replicas. QoS can be defined differently according to
the user’s requirements; here it is taken to be the access time. This work also aims at
minimizing the count of those replicas that violate QoS requirement. This algorithm
is based on the HDFS [11] architecture and each data block has two copies of the
original data. A replication request is generated when an application running on the
disk attempt to write a data block. For the nodes to fulfil the replication request, they
should be QoS qualified. To qualify as QoS nodes following two conditions should
be met:
1. The requested node and the node on which data is to be placed should not be in
the same rack so as to avoid any possible rack failure.
2. The time to access a node qj, for a requested node r i, should be less than the time
required to access a QoS requested node.
The working of this algorithm is explained in Fig. 3 with the help of a conceptual
approach.
Minimum cost and minimum number of replicas can be achieved in polynomial
time.
Also, the system is highly scalable and can ensure that QoS requirements are
fulfilled for a large number of nodes. Every time the higher quality of service node
will be preferred so the load distribution is less.

3.4 Dynamic Replica Control Strategy

The algorithm proposed by Wei in [17] finds a cost-effective replica placement strat-
egy which is based on capacity and blocking probability of nodes containing data.
Capacity of the node is defined as the number of sessions it can allow during a time
and period. Blocking probability is the probability of a request for a session being
blocked by the data node as its capacity exceeded. The algorithm imposes a con-
straint on the bandwidth that can be used for session requests that it should be less
the total bandwidth available for the network.


c
BWnet > S(i)/T (i) (8)
i1

where BW net is the network bandwidth, c is the upper bound on the number of sessions
of a node S and T is the permissible delay. The blocking probability is calculated as
 

c
B Pi (λi τi )c /c! ∗ 1/ (λi ∗ τi )k /k! (9)
k0

[email protected]
An Intensive Review of Data Replication Algorithms … 33

Fig. 3 Working of HQFR algorithm

[email protected]
34 C. Dabas and J. Aggarwal

where λ is the arrival rate of requests and τ is the service time. The replica will
be stored at the node with the least value of blocking probability. The conceptual
workflow of the algorithm is explained in Fig. 4.
The availability of the data is increased when the number of replicas is kept small.
A few sessions are supported simultaneously which may lead to a lot of requests
getting blocked and thus consume more bandwidth and time.

3.5 Cost-Effective Incremental Replication (CIR)

The CIR algorithm is proposed by Li in [18] works by creating replicas in incremental

steps for the purpose of increasing data reliability. Data reliability is based on many
factors such as the time data duration for which the data is stored, number of replicas
to be stored and rate of failure of a node. Data reliability is calculated as below:

k
 
X 1− 1 − e−λi ∗ Tk (10)
i1

Here, k represents the number of replicas, λi is the failure rate and T k is the storage
duration. This algorithm finds the storage reliability for up to three replicas. Working
flow of this algorithm is explained in a conceptual manner in Fig. 5. This algorithm
is able to cater to very large data requirements and works well for data-intensive
applications. As the replica is created only when the replica creation point is reached
therefore number of replicas is reduced which further improves the storage capacity.
While the reliability of data is discussed, the reliability of the metadata table which
contains the information about actual replicas is not considered.

4 Analysis

A comparative analysis of the five replication algorithms discussed above is pre-

sented in Table 1. We observe that the major objective of all the algorithms discussed
thus far is to maintain the availability and reliability requirements of the data while
distributing the load evenly in a cost-efficient way. The ADRS algorithm optimizes
various parameters like mean service time, failure probability, load variance and stor-
age usage. It produces a better response time when the number of nodes is increased
as compared to CIR, build time and DRCS algorithms.
The HQFR algorithm gives less computational complexity when calculating the
number of replicas. The CIR algorithm performs better than build time algorithm in
terms of network utilization. This algorithm also provides a very high reliability.

[email protected]
An Intensive Review of Data Replication Algorithms … 35

Fig. 4 Working of DRCS algorithm

[email protected]
36 C. Dabas and J. Aggarwal

Fig. 5 Working of CIR algorithm

[email protected]
An Intensive Review of Data Replication Algorithms … 37

Table 1 Comparison of data replication algorithms

Algorithm Author Objective Approach Parameters Advantage
considered
ADRS [14] To minimize Multi- Mean service Load
response time objective time, failure variance,
and make optimization probability, response
load balanced load variance, time, network
cloud storage storage usage usage is
improved as
the number of
files increases
HQFR [16] To Maintaining Access time Less
continuously quality of of data block, computation
support service disk access complexity
quality of requirement latency,
service network
requirement latency
of a data after
corruption
DRCS [17] To maintain Placing the Blocking Improves
load replicas with probability, system’s
balancing nodes having replication availability
while placing lowest factor
replicas in a blocking
cost-efficient probability
way
CIR [18] To reduce Incremental Storage Provides
storage cost replication duration, 99.9999%
and improve failure rates reliability
reliability with three
replicas
Build time [15] To reduce the Divide data Access Greatly
cost of data storage space frequency, reduces the
storage and and datasets dataset cost for data
transfer for into parts dependency, management
work storage
applications capacity, size
of dataset

5 Conclusions and Future Scope

Data replication for cloud computing systems has become popular because of its
availability to provide high data availability and ensure reliability. Replication is
done at file level as well as block level. The algorithms discussed in this paper
have particularly focused on maintaining the quality of service of the replicas stored
in the data nodes. Quality of service is maintained by ensuring that the time and
bandwidth required to retrieve a replica should not be more than the time to retrieve

[email protected]
38 C. Dabas and J. Aggarwal

the original copy. Apart from this, the creation of minimum number of replicas for an
original data is taken into account. By taking all these parameters into consideration,
a cost-effective reliable data replication can be performed. In future, the authors
wish to overcome the problem of creating too many replicas by restricting the replica
creation procedure. The work studied so far focuses mainly on read-only data thereby
giving no importance to coordination among replicas. The future work will take into
consideration consistency among replicas along with the best practices.

References

1. Amazon S3. (2017 March). [Online]. Available: http://aws.amazon.com/s3/.

2. Google Cloud storage. (2017, March). [Online]. Available: https://cloud.google.com/storage/.
3. Microsoft Azure. (2017, March). [Online]. Available: https://azure.microsoft.com/en-in/
overview/what-is-azure/.
4. Kapoor, S., & Dabas, C. (2015, August). Cluster based load balancing in cloud computing. In
2015 Eighth International Conference on Contemporary Computing (IC3) (pp. 76–81). IEEE.
5. Feng, K., & Zhang, J. (2017, April). Improving availability and confidentiality of shared data
under the multi-cloud environment. In 2017 IEEE 2nd International Conference on Cloud
Computing and Big Data Analysis (ICCCBDA), (pp. 6–10). IEEE.
6. Mansouri, Y. (2017). Brokering algorithms for data replication and migration across cloud-
based data stores (Doctoral dissertation).
7. Li, J., & Li, B. (2013). Erasure coding for cloud storage systems: a survey. Tsinghua Science
and Technology, 18(3), 259–272.
8. Luo, L., Li, H., Qiu, X., & Tang, Y. (2016, August). A resource optimization algorithm of cloud
data center based on correlated model of reliability, performance and energy. In 2016 IEEE
International Conference on Software Quality, Reliability and Security Companion (QRS-C),
(pp. 416–417). IEEE.
9. Liu, Y., Li, X., Kang, R., & Xiao, L. (2016, October). Service reliability modeling of the IT
infrastructure of active-active cloud data center. In Prognostics and System Health Management
Conference (PHM-Chengdu), 2016 (pp. 1–7). IEEE.
10. Li, W., Yang, Y., & Yuan, D. (2016). Ensuring cloud data reliability with minimum replication
by proactive replica checking. IEEE Transactions on Computers, 65(5), 1494–1506.
11. Shvachko, K., Kuang, H., Radia, S., & Chansler, R. (2010, May). The hadoop distributed
file system. In 2010 IEEE 26th symposium on Mass storage systems and technologies
(MSST), (pp. 1–10). IEEE.
12. Boru, D., Kliazovich, D., Granelli, F., Bouvry, P., & Zomaya, A. Y. (2015). Energy-efficient
data replication in cloud computing datacenters. Cluster Computing, 18(1), 385–402.
13. Sookhtsaraei, R., Artin, J., Ghorbani, A., Faraahi, A., & Adineh, H. (2016). A locality-based
replication manager for data cloud. Frontiers of Information Technology & Electronic Engi-
neering, 17(12), 1275–1286.
14. Mansouri, N. (2016). Adaptive data replication strategy in cloud computing for performance
improvement. Frontiers of Computer Science, 10(5), 925–935.
15. Xie, F., Yan, J., & Shen, J. (2017, August). Towards cost reduction in cloud-based workflow
management through data replication. In 2017 Fifth International Conference on Advanced
Cloud and Big Data (CBD), (pp. 94–99). IEEE.
16. Lin, J. W., Chen, C. H., & Chang, J. M. (2013). QoS-aware data replication for data-intensive
applications in cloud computing systems. IEEE Transactions on Cloud Computing, 1(1),
101–115.
17. Wei, Q., Veeravalli, B., Gong, B., Zeng, L., & Feng, D. (2010, September). CDRM: A cost-
effective dynamic replication management scheme for cloud storage cluster. In: 2010 IEEE
International Conference on Cluster Computing (CLUSTER), (pp. 188–196). IEEE.

[email protected]
An Intensive Review of Data Replication Algorithms … 39

18. Li, W., Yang, Y., & Yuan, D. (2011, December). A novel cost-effective dynamic data replication
strategy for reliability in cloud data centres. In Dependable, Autonomic and Secure Computing
(DASC), 2011 IEEE Ninth International Conference on (pp. 496–502). IEEE.

[email protected]
Integrated Cryptography for Internet
of Things Using TBF Approach

Santosh Kumar Sharma, Muppidi Somasundara Rao,

Lavudi Poorna Chandar Rao and Pendyala Chaitanya

Abstract In current trends, Internet-based portable system is on the huge demand,

and in coming future, it will be the most demanded technology among the smart
society. It is very difficult to say that entire world will use the smart technology but
majority of the developed and developing countries will base on smart work, using
smart technology. The people can control their smart devices from remote location
and can be causation free with security issues, which is associated with their valu-
ables and other belongings. But all these are not so easy to implement practically
to solve security pitfall and other issues to maintain reliable usage of smart devices.
Here, in our proposed system, we are providing data security on data layer by using
the concept of excessive cryptography and implementing the integrated approach
of two cryptography techniques, i.e. TBF-transposition and blowfish cryptography
techniques which can make powerful security barrier against the vulnerability and
can stop the illegal usage of data. Here first of all, the plain text will go through trans-
position after that blowfish block cipher technique will be applied on consequential
value from transposition outcome. This application will use in highly sensitive secu-
rity purpose over the smart secure communication as we know that security issues
are major challenges among smart devices where the data is more sensitive, and if it
will go to wrong hands, there will be loss of human life in some cases. Finally, our
work will contribute to integrate the crypto techniques for maintaining more security
without breach of services.

Keywords Cryptography · IOT security · Transposition · Blowfish · Security

services · Block cipher

S. K. Sharma (B) · M. S. Rao · L. P. C. Rao · P. Chaitanya

Department of MCA, Vignan Institute of Information Technology,
Visakhapatnam 530049, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 41

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_4

[email protected]
42 S. K. Sharma et al.

1 Introduction

In present scenario, Internet-based technology and resources are reachable to almost

every corner of the world. Due to the heterogeneous structure of IOT connectivity,
it creates service reliability along with several security challenges such as authen-
tication for every connected objects and privacy of data. Even though it is easy to
access Internet-based devices, it invites lots of security challenges when it comes to
daily life of human being; later on if the device control will go in wrong hand, then it
may create serious damage to the connected objects. IOT is an integrated approach
of hardware and software such as sensors, actuators, transistors and processors. IOT
needs a high degree of computational aptitude to perform the smart task. IoT is a
integrated product of electronic and electromagnetic devices such as sensors, actu-
ators and microcontrollers with internet connectivity which are embedded on high
degree circuit to conduct specific task. IoT is having major applications to make the
system smart that can be explore in face of smart city, smart health management
system, smart farming, smart home appliances, smart displays and smart security.
For making the system elegant, there is a need to collaborative the smart technology
but it should be mange through permitted users only to avoid the malicious usage
and access of resources.

2 Related Work

By studying several scientific journals and articles related to database security and
wireless sensor network security is exposing the several security threats to authentic
data accessing and transmission. Here Huang [1, 2] has discussed the application of
signature-based authenticated technique and the successful test on IOT devices with
the help of Burrows–Abadi–Needham logic for security and using of automated tools
and applications (AVISPA) tool. Instead of analysing a single-level security, multiple
nodes can also be secured, which provides security in communication, application
interface and cryptography for IOT security. Situation is not so far and by the year
of 2020, IOT devices will go beyond the billions counts, which create the collective
issues related to collaboration techniques. Security is a major challenge in presence
of many anomalies & malicious user. Kang and Lee [3, 4] has discussed the security
at different stages such as communication security, application level security. Where
as Kim. S has discussed deep impact of block chain technique on IoT devices. Zahra,
Majeed and Mohsin [5–7] has done the risk analysis and projected the framework
for security of IoT as future proof and Quality of services is provided by the Maliki,
Khanna with trust management approach [8–10]. Nawir, Baldini and Zada Khan
[11–13] they have presented several taxonomy for preventing the attacks with secu-
rity certificate. Metongnon, Nakagawa and Ben [14–16] has kept the heterogeneous
behaviour of IoT and how to secure this with the help of agent platform mechanism
for social IoT. Zonari, Midi and Urien [17–19] has discussed the different aspects

[email protected]
Integrated Cryptography for Internet of Things … 43

of homomorphic cryptographic technique and intrusion detection with threat assess-

ment.
With the continuous growing of IOT technology, ‘Things’ alarming several critical
issues subsequently regarding how to access and monitor the authentication manage-
ment to transmit and access the data in wireless environment security management
is difficult in wireless environment which leads the service segment issues towards
the devices and human life. It turn out to be so risky once we fail to provide service
communication among them, Sklavos [20] has discussed several security issues to
solve this. Sharma and Baik [21, 22] have discussed Massey–Omura cryptography
approach as a private key encryption technique with its application and strength. M-
Omura works on the precept of the prime modulus with exponential system, where
the message is generated with secret key and transmits to the receiver side by using
prime modulo. Our study found different types of attacks and threats to our valuables
along with its possible security measures such as bootstrap modelling, cryptography
(symmetric/asymmetric) authentication protocol, security proxy servers and scan-
ners; in spite of all these available security tools, we cannot confident about data
security and we need more complex security techniques to safe our things and data.
Gong, Hu, Wen and Bose [23–25] then presented security mechanism at physical
layer and how the data can be transmit intelligently over the network by focusing the
different security techniques in different layers of IoT architecture. Andreas and Atat
[26, 27] proposed how to analyse the risk assessment and security certification with
Armour project in IoT. Xu and Moon [28, 29] introduced the concept of two-factor
authentication protocol whose working principle is to work in multi-server system to
provide more security for known attacks. Further, Jongho discussed the concept of
online enemy attack with process saver authentication protocol using AVISPA and
random Oracle. Jesus, Caminha and Du [30–32] revealed the strategy of the block
chain for securing the Internet of things, and Jean discussed the trust management
to show the issue of on-off attack for the IoT devices. Zhou, Wang and Li [33–35]
worked with the quantum cryptography for analysing the upcoming future security
challenges. Chen Wang provided the concept of novel security scheme for IoT by
implementing on instant encrypted transmission.

3 Proposed System

3.1 Transposition/Permutation

Transposition work is to produce the transpose of given matrix unless by supporting

column values with row values. First, we take a message and change the order of
message as a transposition and split this transposition message into four equal halves
and named it as ‘P(x)’.

[email protected]
44 S. K. Sharma et al.

EXAMPLE Take a message which is string

Message  “IOT FOR MILITARY SERVICES”
Perform or split this message into transposition and the result will be as

I O L R R E
O R I Y V Z
T M T S I Z
F I A E C Z

where ‘Z’ are taken as dummy values, and the splitting result is

IOTF ORMI LITA RYSE RVIC E Z ZZ

3.2 Blowfish Algorithm

The working principle of blowfish is symmetrically block cipher approach, and it is

originated to replace the IDEA algorithm in cryptography. This blowfish algorithm
is used to convert plain text into cipher text which is entrenched with 32–448 bit
encryption and work with the block size of 64 bit, operational activity for making
final cipher is 16 rounds. Operational behaviour of blowfish as follows:

3.3 Algorithm

Step-1: Initiate the cryptography process.

Step-2: Let message (M)ˆT  P(x) //Apply transposition.
Step-3: Arrange the message in permutation order (transposition). If the permu-
tation order is not filled, place a dummy value.
Step-4: New confuse matrix [m]ˆCM  cipher value.
Step-5: Split the message as blocks in horizontal order so that each block contains
4-values that is taken as P(X).
Step-6: P(X) ¥ 64 bits.
Step-7: Divide 64 bit into 2 equal halves (32 bit) and named as left (XL ) and right
(XR ).
Step-8: For Pn  1 to 16 //if n  true then next step otherwise go to Step-11.
Step-9: Perform operation XL  XL ⊕ Pi .

[email protected]
Integrated Cryptography for Internet of Things … 45

Table 1 Comparison of available algorithms with blowfish

Cryptography Key size (BITS) Block size Security Speed
Algorithms (BITS)
DES 56 64 Less secure Low
RC6 128,192 128 Less secure Low
AES 128,192,256 128 More secure Fast
RSA Maximum Minimum Less secure Fast
1024 Bits 512 bits
Idea 128 64 More secure Fast
Blowfish 32-448 64 More secure Fast

Step-10: Perform operation XR  F (XL ) XOR XR

{ Where F(XL )  ((s1 [a] + s2 [b] mod 2∧ 32) ⊕ s3 [c]) + s4 [d] mod 2∧ 32
Logically it can be expressed as F(XL )
  
 s1 [a] + s2 [b]%232 ⊕ s3 [c] + s4 [d]%232 }

Step-11: Run Swap() XL XR .

Step-12: Perform XL and XR operations up to i value 16 and swap operation (undo
the last swap).
Step-13: Perform operation XR  XR ⊕ P17.
Step-14: Perform operation XL  XL ⊕ P18.
Step-15: Recombine XL and XR .
Step-16: Stop.

3.4 Application and How It Works

This article says about IOT security that helps to make secure data services. They are
equipped with latest crypto techniques to fight against the vulnerability of services.
Instead, constant improvement leveraging advanced technology is needed for the
tools to enable security forces edge out the amateurs. Security is a major challenge
during the development of IOT application where huge number of unprivileged user
try to use the resources illegaly. So our work is to create the complexity for malicious
users in case of intercepting the confidential data services. TBF creating the two-level
encryption technology where first level work is to create confusion for the amateur
using transposition, and the second level provides the encryption algorithm to secure
the data, after a long course of action making the concrete secure system for the IOT
(Fig. 1; Table 1).
DES: Data encryption is a old encryption algorithm which work on the symmetric
key where only one key is use for encryption and decryption process which is shared

[email protected]
46 S. K. Sharma et al.

Fig. 1 Secure communication using TBF

by both sender and receiver as a private key. Here we are processing the block size
of 64 bit on which DES is applying 56 bit key length.
RC6: RC is stand for Rivest cipher which is available in market from RC 1 to RC6
versions (since 1987 to 1997) which is symmetric key block size 128 bits cipher and
support the size from 128, 192, 256 and 2040 bits up to maximum length.
AES: Advanced encryption standard algorithm is the powerful and well known block
cipher techniques which implement on block cipher with symmetric key concept.
AES is considerably faster than DES with small key size which provides fast and
secure communication with the help of strong cryptography process with 16-rounds
of encryption process.
RSA: RSA is named on three scientist Rivest, Shamir and Adleman who has given the
widely accepted first public key cryptography for secure data transaction . Among all
the available public key cryptography it is most popular algorithm which work prime
number products and other operation. RSA algorithm is used as a basic algorithm to
generate other advance algorithm such as Massy Omura and Diffie hellman algorithm.
IDEA: Apart from all the cryptography algorithm the IDEA ( International Data
Encryption Algorithm ) is having its own unique importance which implement on
64 bit block size with 128 bit key along with 8 number of rounds to transform the
plaintext into subsequence strong cipher text.
Blowfish: Blowfish agorithm is also a symmetric block cipher technique which is
invented in 1993 to produce big number of cipher sets. Its block size is 64 bit and
takes key length from 32 bit to 448 bit and complete the encryption process in number
of rounds with fixed size of boxes such as Feistal algorithm.

[email protected]
Integrated Cryptography for Internet of Things … 47

Example

(a) Encryption:
The process of encryption is converting plain text into cipher text because it should
not be visible to the third parties.

I N I Y V S
O M T S I Z
T I A E C Z
I L R R E Z

Here ‘Z’ is taken as a dummy value. The decimal values of above message is
9 + 15 + 20 + 9 + 14 + 13 + 9 + 12 + 9 + 20 + 1 + 18 + 25 + 19 + 5 + 18 + 22 + 9
+ 3 + 5 + 19  274
The value ‘274’ is converting into binary value as 64 bit the value is

X  00000000000000000000000000000
00000000000000000000000000100010010

Now spilt the X into two halves as 32 bit named as XL and XR

where

XL  00000000000000000000000000000000 (32 bit)

XR  00000000000000000000000100101010 (32 bit)

P1  0 × 243f6a88 the binary value of p1 is

00100100001111110110101010001000 (Fig. 2)
Round 1:
Compile XL ⊕ p1

XL  00000000 00000000 00000000 00000000

⊕ P1  00100100 00111111 01101010 10001000
00100100 00111111 01101010 10001000

Compile F(XL ) ⊕ XR

F(XL )  00000000 00000000 00000000 00000000

⊕ XR  00000000 00000000 00000001 00010010
00000000 00000000 00000001 00010010
Swap XL → XR

[email protected]
48 S. K. Sharma et al.

Fig. 2 Encryption process

XL ← XR

Round 2:
(a) Now perform ⊕ (XOR) operation for XL and p2

XL  00000000 00000000 00000001 00010010

⊕ P2  10000101 10100011 00001000 11010011
10000101 10100011 00001001 11000001

[email protected]
Integrated Cryptography for Internet of Things … 49

F(XL )  00000000 00000000 00000001 00010010

↓ ↓ ↓ ↓
A B C D

Add( ) values of A and B

00000000 00000000 00000000 00000000


00000000 00000000 00000000 00000000
00000000 00000000 00000000 00000000

Perform ⊕ (XOR) operation of A + B value with C

00000000 00000000 00000000 00000000

⊕ 00000000 00000000 00000000 00000001
00000000 00000000 00000000 00000001

Fig. 3 Decryption process

[email protected]
50 S. K. Sharma et al.


Add ( ) C value with D

00000000 00000000 00000000 00000001


00000000 00000000 00000000 00000010
00000000 00000000 00000000 00000011

(b) Now perform ⊕ (XOR) operation for F(XL ) and XR

F(XL )  00000000 00000000 00000000 00010011

⊕ XR  10000101 10100011 00001000 11010011
10000101 10100011 00001000 11000000

Swap XL and XR values so the values of XL and XR are interchanged.

Perform above operations up to the P18.
(b) Decryption:
The decryption means converting cipher text to plain text. In the above encryption
process, we performed operations as a result we got cipher text. Now convert this
cipher text into plain text.
From the above context Fig. 1 is giving the demonstration for secure communi-
cation using TBF, Fig. 2 showing the Encryption process and Fig. 3 is presenting the
decryption process.

4 Conclusion

By traversing the numerous factors related to security breach, we proposed and

concluded that single security cannot be work, and demanded to create the multiple
approaches to provide the security at every level to provide the maximum security of
data resources. In future work, our proposal is to design the secure communication
channel for the military vehicles to protect themselves from any casualties using
end-end security mechanisms.

References

1. Huang, H. (2017). Private and secured medical data transmission and analysis for wireless
sensing healthcare system. Citation information: Transactions on Industrial Informatics 1–10.
2. Challa, S., Wazid, M., Das, A. K., & Kumar, N. (2017). (Member, IEEE), Secure signature-
based authenticated key establishment scheme for future iot applications. IEEE Access, 1–16.
3. Sain, M., Kang, Y. J., Lee, & H. J. (2017). Survey on security in internet of things: State of the
art and challenges, pp. 699–704.

[email protected]
Integrated Cryptography for Internet of Things … 51

4. Huh, S., Cho, S., & Kim, S. (2017). Managing IoT devices using block chain platform,
pp. 464–467.
5. Zahra, B. F., Fatima Zahra, B. (2017). Risk analysis in internet of things using EBIOS.
6. Majeed, A. (2017). Internet of Things (IoT): A verification framework.
7. Mohsin, M., Anwar, Z., Husari, G., Al-Shaer, E., & AshiqurRahman, M. (2016). IoTSAT: A
formal framework for security analysis of the internet of things (IoT).
8. Abels, T., Khanna, R., & Midkiff, K. (2017). Future proof IoT: Composable semantics, security,
QoS and reliability, pp. 1–4.
9. El-Maliki, T. (2016). Efficient security adaptation framework for internet of things,
pp. 206–211.
10. Ben Abderrahim, O., HoucineElhdhili M., & Saidane, L. (2017) TMCoI-SIOT: A trust man-
agement system based on communities of internet for the social internet of things, pp. 747–752.
11. Nawir, M., Amir, A., Yaakob, N., & Bi Lynn, O. (2016). Internet of things (IoT): Taxonomy
of security attacks, pp. 321–326.
12. Baldini, G., Member, IEEE, SKarmeta, A., Fourneret, E., Neisse, E., Legeard, B., & Gall, F.
L. (2017). Security certification and labelling in internet of things, pp. 627–632.
13. Zada Khan, W., Mohammed Zangoti, H., Aalsalern, M. Y., Zahid, M., Arshad, Q. (2016).
Mobile RFID in internet of things: security attacks, privacy risks, and countermeasures,
pp. 36–41.
14. Metongnon, L., Eziny, E. C., & Sadre, R. (2017). Efficient probing of heterogeneous IoT
networks, pp. 1052–1058.
15. Nakagawa, I., & Shimojo, S. (2017). IoT agent platform mechanism with transparent cloud
computing framework for improving IoT security, pp. 684–689.
16. Ben Abderrahim, O., Elhedhili, M. H., & Saidane, L. (2017). CTMS-SIOT: A context-based
trust management system for the social internet of things, pp. 1903–1908.
17. Zouari, J., Hamdi, M., & Kim, T.-H. (2017). A privacy-preserving homomorphic encryption
scheme for the internet of things, pp. 1939–1944.
18. Midi, D., Rullo, A., & Mudgerikar, A. (2017). Kalis a system for knowledge-driven adaptable
intrusion detection for the internet of things, pp. 656–666.
19. Dorsemaine, B., Gaulier, J. P., & Urien, P. (2017). A new threat assessment method for inte-
gration an IoT infrastructure in an information system, pp. 105–112.
20. Sklavos, N., & Zaharakis, I. D. (2016). Cryptography and, security in internet of things (IoTs):
Models, schemes, and implementations.
21. Sharma, S. K. (2017). A survey on layered approach for internet of things security. SERSC,
ASTL, SMART DSC-2017 (Vol. 147, pp. 26–33).
22. Sharma, S. K., Baik, N., Khuntia, B. (2018) Encrusted security for internet of things using
MAC-OMURA. IJCA, SERSC-Australia, pp. 45–54.
23. Gong, B., Wang, Y., Liu, X., Qi, F., & Sun, Z. (2018, February). A trusted attestation mechanism
for the sensing nodes of internet of things based on dynamic trusted measurement, pp. 100–121.
24. Hu, L., & Wen, H. (2018, February). Cooperative jamming for physical layer security enhance-
ment in internet of things (pp. 219–228).
25. Bose, R. (2017). Channel-based mapping diversity for enhancing the physical layer security in
the internet of things (2017 IEEE).
26. Andreas. (2018, January). Wireless communication and security issues for cyber–physical
systems and the internet-of-things, pp. 38–60.
27. Atat, R., Liu, L., Ashdown, J., Medley, M., Matyjas, J., Yi, Y. (2017). A physical layer security
scheme for mobile health cyber-physical systems, pp. 1–15 (2017 IEEE).
28. Xu, G., & Li, W. (2018, 15 April). A secure and anonymous two-factor authentication protocol
in multi server environment, pp. 1–15.
29. Moon, J., & Youngsook, L. (2017, 27 September). Improving an anonymous and provably
secure authentication protocol for a mobile user, pp. 1–13.
30. Jesus, E. F., Chicarino, V. R., & de Albuquerque, C. V. (2018, 8 April). A survey of how to use
blockchain to secure internet of things and the stalker attack, pp. 1–27.

[email protected]
52 S. K. Sharma et al.

31. Caminha, J., & Perkusich, M. (2018, 15 April) A smart trust management method to detect
on-off attacks in the internet of things, pp. 1–10.
32. Du, Q., & Song, H. (2018, 8 February). Security enhancement for multicast over internet of
things by dynamically constructed fountain codes, pp. 1–11.
33. Zhou, T., Shen, J., Li, X., & Wang, C. (2018, 21 February). Quantum cryptography for the
future internet and the security analysis, pp. 1–7.
34. Wang, C., & Ren, T. L. (2018, 17 May) A novel security scheme based on instant encrypted
transmission for internet of things, pp. 1–7.
35. Li, Q., Zhu, H., & Zhang, T. (2018, 6 June) Traceable cipher text-policy attribute-based encryp-
tion with verifiable outsourced decryption in e-health cloud, pp. 1–12.

[email protected]
Data Governance on Local Storage
in Offsite

G. Priyadharshini and K. Shyamala

Abstract The major challenges faced by software industry are how to restrict
confidential data and protect copyright or intellectual property information between
customer location and delivery center. Typically customer will have multiple ven-
dors spread across various geographical locations. In global delivery model, customer
sensitive information will be exchanged between teams and there is possibility of
data breach from customer network. Though these delivery centers are firewall seg-
regated or air gap network, it is difficult to restrict end user to store information in
local device. Thin client installation at delivery center or Virtual Desktop Infrastruc-
ture (VDI) setup at customer location are trivial solution to ensure information is not
moved out of network. All traditional offshore development centers may not have
thin client set up and they use workstation with local storage. Converting these work-
stations into thin client is expensive and customer may not be ready to provide VDI
setup for offsite location. This paper provides simple solution with zero investment
for local workstations to act like Thin Client, and also to ensure that there is no data
or information leakage through local storage.

Keywords Data security · Local storage restriction · Group policy object · Thin
client · VDI · Data protection

1 Introduction

For any organization, one of the key strategic assets is data. Considering the way
data are growing and shared through multiple channels, it is equally important to do
data classification. We also need to take utmost care for protecting sensitive, confi-
dential, and restricted data. In information security, we need to secure both tangible
and intangible assets to ensure it is trustworthy to use for business. Tangible assets
include workstation, access points, servers, router, firewall, switches, etc. Intangible

G. Priyadharshini (B) · K. Shyamala

Research Department of Computer Science, Dr. Ambedkar Government
Arts College, Chennai 600039, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 53

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_5

[email protected]
54 G. Priyadharshini and K. Shyamala

Fig. 1 Sensitive information flow across all locations

asset includes software code, development life cycle documents, customer details,
contracts, invoices, statement of works, personally identifiable information, and so
on. To secure assets, we also need to consider cost incurred for securing information.
Securing intangible asset plays vital role and challenge to keep them with restricted
access.
In this digital world, personal and industrial sensitive information spread across
the world and are stored in the cloud. Considering vulnerabilities and public threat,
there is no guarantee that these information are secured. There are several regula-
tory bodies and acts established across the globe to guide and product data. These
regulatory bodies want to give control over how individual data are used by orga-
nization. The organization will comply with these regulatory bodies based on the
geography or industry it operates. For example, healthcare industries follow HIPAA
(Health Insurance Portability and Accountability Act) which is established in 1996.
The organization which stores credit card information is following payment card
industry data security standard (PCI DSS).
United Kingdom (UK) came up with The Data Protection Act in 1998 [1] to
protect personal data stored on computers or in an organized paper filing system.
Europe Union (EU) establishes GDPR (Global Data Protection Regulation) which
will apply for all EU states from May 2018 (Fig. 1).
A customer may have multiple vendors for their operations like information tech-
nology, marketing, administration, human resources, etc. Each vendor will operate
from different locations spread across customer site, nearshore, and offshore. Cus-
tomer data will flow across all these locations in the form of document, data, and
source code which include sensitive detail.
As per “Infowatch Analytic center research report on confidential data leakage”
[2], 66% of leaks are from employees who include intentional as well as accentual.
The only way to prevent such data leakage is not to store any data locally and monitor
local storage on periodic intervals (Fig. 2).

[email protected]
Data Governance on Local Storage in Offsite 55

Fig. 2 Sources for data leaks

As per data protection act, data [3]:

(a) is being processed by means of equipment operating automatically in response
to instructions given for that purpose,
(b) is recorded with the intention that it should be processed by means of such
equipment,
(c) is recorded as part of a relevant filing system or with the intention that it should
form part of a relevant filing system,
(d) does not fall within paragraph (a), (b) or (c) but forms part of an accessible
record as defined by Section 68, or
(e) is recorded information held by a public authority and does not fall within any
of paragraph (a) to (d).
Personal data [3] means data which relate to a living individual who can be
identified:
(a) from those data, or
(b) from those data and other information which is in the possession of, or is likely
to come into the possession of, the data controller,
(c) any expression of opinion about the individual and any indication of the inten-
tions of the data controller or any other person in respect of the individual.

2 Related Work

There are several tools available in the market to prevent data leakage from local
network. Most of leading products are licensed version and expensive to implement.
These software demonstrate the need for a high-profile acceptable use policy to
prevent data leakage and give practical guidance on how to allot funds for IT security.

[email protected]
56 G. Priyadharshini and K. Shyamala

The challenges of protecting sensitive data from internal users, preventing from
insiders and monitoring data leakage are the most difficult area in information secu-
rity. Security personnel finding it difficult to identify potential insider threats by
investigating suspicious activities in a person’s physical behavior. Malicious insiders
continue to steal data from network leaking personally identifiable information (PII).
Whatever the approach, researchers are struggling to define exact problem and find
the solution for the problem for all scenarios. The research of Huth et al. [4] gives a
different solution to address components of the problem, with the goal of minimizing
the potential malicious insiders can inflict on an organization.
It is not easy to identify insider who is trying for security breach data and rea-
son for data leakage through the privilege and access given for their regular task.
There are few solutions derived through ontological approach, and utilize the notion
of semantic associations and their discovery among a collection of heterogeneous
documents. New approaches have been documented seeking to address components
of the problem, with the goal of reducing the harm malicious insiders can inflict
on an organization. An Ontological Approach to the Document Access Problem of
Insider Threat [5] describes the research and prototyping of a system that takes an
ontological approach and is primarily targeted for use by the intelligence community.
David Clark et al. did research on Quantitative Analysis of the Leakage of Con-
fidential Data [6] in which the team uses information theory to analyze the sensitive
and confidential information which may be leaked by software applications written in
very simple conditional statements and improper iterative statements. This analysis
helps to determine the bounds on the quantity of information leaked.
Data leakage detection model has been developed [7] for data protection. This
research did a detailed study on watermarking technology and proved that this tech-
nique will not help for data protection. This study also gives the difference between
watermarking technology and data leakage detection model.
None of these studies provide the solution on preventing data leakage from local
storage through a potential insider. This research focuses on a simple solution for
such a scenario without any additional investment. This solution is fully automated
and there is no need for any manual intervention.

3 Problem Statement

When the vendor is working on customer’s data which is sensitive or contains per-
sonal detail like personal Identifiable Information (PII), vendor must comply with
data protection act based on Industry or geography. When an organization subcontract
or outsource their work, they need to ensure that vendor also support data protection
act and data is not misused or leaked through subcontractor or vendor organization.
As per the report published by breachlive [8], 59 records are stolen every second.
As per global leakage report 2017, published by infowatch [2], data leakage
through instant messaging, network is increased in the last 1 year. The share of data
leaks resulting from theft/loss of equipment, as well as through removable media

[email protected]
Data Governance on Local Storage in Offsite 57

Fig. 3 Distribution of Leaks by channel

and paper documents must be reduced [2]. It is necessary to find a solution to avoid
data leakage through network (Fig. 3).
Breachliveindex [9] research report presents top ten data breaches in the last
5 years which are listed in Table 1. Sources for these breaches are malicious outsider
and malicious insider.
The printing area is 122 mm × 193 mm. The text should be justified to occupy
the full line width, so that the right margin is not ragged, with words hyphenated as
appropriate. Please fill pages so that the length of the text is no less than 180 mm, if
possible.
To secure customer data, few organizations prefer to establish the connection
with the vendor through VDI (Virtual Desktop Infrastructure), citrix receiver or
forcing vendor to have air gap offshore development center. In air gap ODC or
shared network, several security controls like providing shared folder, spot check are
established to avoid local storage.
VDI [10] effectively addresses and resolves key problems in terminal server-
based approaches to server-based computing. VMware ESX Server allows multiple
user desktops to run as separate virtual machines while sharing underlying physical
hardware resources such as CPU, memory, networking, and storage. This isolates
users from each other, giving each user their own operating system, allowing gran-
ular resource allocation and protecting users from application crashes and operation
system faults caused by the activities of other users.
VDI interacts with remote virtual machine and local desktop acts as dump termi-
nal. The operating system, application, and data are stored on virtual machine hosted
on central server. This saves desktop management cost, software licensing cost and
most importantly it improves security and data leakage from customer location. It
requires additional investment on customer side. When customer does not want to
invest in setting up VDI and keen only on data leakage, we need solution at local
desktop used by vendor at various geographical locations.

[email protected]
58 G. Priyadharshini and K. Shyamala

Table 1 Top 10 data breaches

Organization Records Date of breach Source of breach Industry
breached breached
Equifax 143,000,000 7/15/2017 Malicious Financial
Outsider
Friend Finder 412,214,295 10/16/2016 Malicious Entertainment
Networks Outsider
Anthem 78,800,000 1/27/2015 State Sponsored Health care
Insurance
Companies
(anthem blue
cross)
Home Depot 109,000,000 9/2/2014 Malicious Retail
Outsider
JPMorgan 83,000,000 8/27/2014 Malicious Financial
Chase Outsider
eBay 145,000,000 5/21/2014 Malicious Retail
Outsider
Korea Credit 104,000,000 1/20/2014 Malicious Financial
Bureau, NH Insider
Nonghyup Card,
Lotte card, KB
Kookmin card
Target 110,000,000 11/4/2013 Malicious Retail
Outsider
Adobe Systems, 152,000,000 9/18/2013 Malicious Technology
Inc. Outsider
MySpace 360,000,000 6/11/2013 Malicious Other
Outsider

Users tend to download various data from customer site and refer them for particu-
lar release. The data can be in the form of document, manuals, procedures, processes
or sensitive information. Most of the users will not delete these files and there is a
chance that some sensitive information stored in these folders without user knowl-
edge. In addition to default files, the users will download customer information for
project reference. After downloading the file, user will come to know that few files
may not be needed for project purpose, but felt good to have in the system. These
will lead to data leakage as per customer contract and lead to regulatory compliance
issue. To avoid such scenario, it is important to wipe out unwanted files from location
storage on regular basis. Table 2 gives some scenarios which lead to keep sensitive
data locally without the user’s knowledge.

[email protected]
Data Governance on Local Storage in Offsite 59

Table 2 Common files Scenario Description

stored locally by the user
Temporary files Temporary files created by application
Downloaded Any files downloaded from Internet will
be stored in default location
Instant messages All individual or group chat transcription
and file shared between users will be
stored here

3.1 Research Objectives

The objective of this research is to find a low-cost solution to prevent data storage
from local drive where data flows between various vendor group and customer. In
a scenario where customer does not have VDI, citrix environment, vendor needs to
implement multiple controls including monitoring and log review to restrict sensitive
data within customer network.
This solution provides following benefits to vendor and customer:
• Restrict local storage with zero cost.
• Does not require any additional infrastructure set up.
• There is no manual process.
• Zero maintenance.
• No need for any additional review or log monitor.
• Easily portable and replicate to multiple locations.
• User group and file type are configurable.
• Policies and groups can be modified in no time.

4 Proposed Solution

Hackers and malware attackers are a major threat to data loss. These attackers target
right from global customers (particularly banks, health care, retail, and insurance)
to novice users. However, accidental breaches caused by employee error or data
breached while controlled by third-party suppliers continue to be a major problem,
accounting for 30% of breaches overall.
To avoid data theft/leakage from the network, no information should be stored
locally for a longer duration. The data includes user created, system generated, down-
loaded, temporary files, and cookies. This can be achieved by implementing different
Group Policy Objects to various user groups.
Group Policy Object (GPO) is a collection of settings that define what a system
will look like and how it will behave for a defined group of users. Microsoft provides a
program snap-in that allows you to use the Group Policy Microsoft Management Con-
sole (MMC). The selections result in a Group Policy Object. The GPO is associated

[email protected]
60 G. Priyadharshini and K. Shyamala

with selected Active Directory containers, such as sites, domains, or organizational

units (OUs). The MMC allows you to create a GPO that defines registry-based poli-
cies, security options, software installation and maintenance options, scripts options,
and folder redirection options [11].
To form a user group, we need to do detailed analysis based on nature of work
by the end user. The users may belong to any of below categories but not limited to
these groups alone:
• Users who are working on heavyweight software which has dependency on local
storage for installation and keeping libraries and APIs.
• Users who require more storage for coding, compilation, and testing.
• Users do not require any storage and work on maintenance and production support.

4.1 Group Classification

• Group 1: Those who do not construct any code and does not require any local
storage.
• Group 2: Those who develop code, store libraries, and make changes in config-
uration and settings. These groups can store files for their development purpose,
but should not store any documents or references on permanent basis.
• Group 3: Third group includes those who are considered as exceptions and require
storage for a longer duration. This group may work on heavy weight applications
or working on document management related activities.

4.2 Solution for Each Group

• Group 1: Local storage for this group should have two logical drives. Primary
drive should be hidden for this group and used for operating system and software
installation. Only secondary drive should be visible to this group, and they are
allowed to store their files in secondary drive or desktop. The files stored in desktop
are redirected internally to secondary drive. All files stored in secondary drive
should be deleted through scheduler on periodic basis.
• Group 2: This group will have access to all drives and they can store files which
are needed for their development work. They are not allowed to store any office
documents or pdf files. Scheduler will run every day to wipe out all documents
and pdf files irrespective of location including temp folder, downloaded or recycle
bin.
• Group 3: These are exception groups. These users should be educated to know data
classification and security risk on storing sensitive information. These machines
will be monitored on a periodic basis and appropriate action should be taken to
ensure compliance.

[email protected]
Data Governance on Local Storage in Offsite 61

5 Design Approach

5.1 Workstation Analysis

Installed software and required storage size for applications need to be analyzed
for all workstations. Table 3 gives minimum information to be collected from each
workstation. This will help to map them into different groups.

5.2 Group Classification

As mentioned in the proposed solution, all workstation should be mapped to particular

group. The mapping details can have either list of workstations to be included or
exempted from each group.

5.3 Group Policy Creation

Workstations belong to Group 1 and will not have any files which are created by the
user. All user files will be deleted on scheduled time from the local drive. This can
be achieved through the following settings:

Table 3 Settings for group policy I

Field name Description Usage
Hostname Hostname of workstation This hostname will be
included in GPO
Software All software installed in the Helps to analyze whether it
workstation has dependency on local
storage
Additional storage required for This will have “Yes” or To know whether software
executing software “No” is needed for additional
storage for storing
Libraries/configuration
settings or API
Additional storage required for Is there any additional space This will help us to classify
the user needed for the user to store correct group
reference files
Projected space required for Local storage space This is needed for future
storing user file required in MB work to calculate required
bandwidth to move files to
centralized location

[email protected]
62 G. Priyadharshini and K. Shyamala

• USB, CD, and primary drive will be hidden normal user and accessible only for
admin.
• Only secondary drive will be visible and accessible.
• Files can be saved either in Desktop or secondary drive.
• Data stored in the desktop are automatically redirected to secondary drive.
• Scheduler will run from the workstation and all files will be deleted automatically
to ensure local storage restriction as part of security requirement.
Workstations belong to Group 2; specific file groups will be wipeout from local
storage. This can be achieved through the following settings:
• USB, CD, and primary drive will be hidden for normal user and accessible only
for admin.
• By using administrative template, below options should be disabled.
• Auto recover delay.
• Auto recover save location.
• Keep the last autoSaved version of files for the next session.
• AutoRecover time.
• Save AutoRecover info.
• Default file location.

6 Execution Approach

6.1 Group Policy Object-1

The basic requirement to implement this policy is workstation must have two parti-
tions (Primary and Secondary drive). In addition to basic function listed in Sect. 5,
the following sub-functions should also be applied. If the below sub-functions are
not applied, the user can do workaround to save file in primary location.
• Hide libraries from Windows explorer (documents, downloads, music, videos, and
picture).
• Restrict adding files to the root folder of the user.
• Prohibit users from manually redirecting profile folders.
• Remove “My Documents”, “Recycle bin” from desktop.
• Remove “Documents”, “Music”, “Picture” from startup.
• Remove “Pinned program”, “Recent items”, “Network” from startup.
• Remove properties icon from context menu.
• Restrict access to control panel.

[email protected]
Data Governance on Local Storage in Offsite 63

6.2 Group Policy Object-2

This group policy is used for Group 2. To implement GPO to restrict on office files,
Microsoft Office administrative template files (ADMX/ADML) is required which
are available in Microsoft download center [12]. The policy can be implemented
using the steps given below.
• Download ADMX from Microsoft download center.
• Import ADMX to Domain Controller.
• Write PowerShell script in Netlogon to remove all file extensions used for MS
Office, Open Office, and pdf.
• This file extension can be customized as per user requirement. The sample code
given below is used for removing documents and pdf files alone.
$objShell  New-Object -ComObject Shell.Application
$objFolder  $objShell.Namespace(0xA)
$WinTemp  get-childitem C:\, D:\ - include *.doc,
*.docx, *.docm, *.dotm, *.odt, *.dotx, *.pdf -recurse
Remove-Item -Recurse $winTemp -Force
• This can be extended to remove other types of files as well.
• Configure GPO as mentioned in Sect. 5.3.
• Write batch file to clean files and place them in Windows startup script.
• Include workstation details to GPO.
• Apply GPO and enable scheduler to run on a regular interval.

7 Conclusion and Future Work

To avoid data leakage from customer network where we do not have VDI or citrix
infrastructure, it can be achieved through GPO implementation. This is easy to man-
age and scalable to future expansion without any additional investment. By using
existing windows group policy settings, administrative template files, and simple
PowerShell scripts and batch files, data leakage from the network can be restricted
and compliance to security requirement with zero cost. This approach does not require
any change on network model and there is no impact on business continuity. This
Group policy can be configured to add additional file types.
As the next step, these policies can be enhanced to include exception on retaining
list of files so that all types of files can be wipeout. Also if the user renames the file
extension to accepted file types, there is no provision to delete those files through
this policy. We need to write a script which reads metadata of file to identify the
file type and implement same GPO. This is applicable only for Windows operating
system environment and does not apply for other operating systems. We need to find
a similar solution for other operating systems.

[email protected]
64 G. Priyadharshini and K. Shyamala

References

1. http://www.legislation.gov.uk/ukpga/1998/29/contents.
2. https://infowatch.com/report2016_half#-DataLeaksource.
3. https://ico.org.uk/media/for-organisations/guide-to-data-protection-2-9.pdf.
4. Huth, C. L., Chadwick, D. W., Claycomb, W. R., et al. (2013). Information system Frontiers-
springer US-arch, 15(1), 1–4.
5. leman-Meza, B., Burns, P., Eavenson, M., Palaniswami, D., & Sheth, A. (2005). An Ontological
approach to the document access problem of insider threat. In: Kantor P., et al. (Eds.), Intel-
ligence and Security Informatics. ISI 2005. Lecture Notes in Computer Science, (Vol 3495).
Berlin, Heidelberg: Springer.
6. Clark, D., Hunt, S., & Malacaria, P. (2002) Quantitative analysis of the leakage of confidential
data. Electronic Notes In Theoretical Computer Science, 59(3), Elsevier.
7. Kale, S. A., & Kulkarni S. V. (2012, November) Data leakage detection. International Journal
of Advanced Research in Computer and Communication Engineering 1(9).
8. http://breachlevelindex.com.
9. http://breachlevelindex.com/top-data-breaches.
10. https://www.vmware.com/pdf/vdi_strategy.pdf.
11. http://searchwindowsserver.techtarget.com/definition/Group-Policy-Object.
12. https://www.microsoft.com/en-in/download/details.aspx?id=35554.

[email protected]
Extraction of Character Personas
from Novels Using Dependency Trees
and POS Tags

Nikhil Prabhu and S. Natarajan

Abstract Novels are a rich source of data for extracting interesting information.
Besides the plot, the characters of a novel are its most important elements that shape
the story and its message. An interesting task to consider is extracting these charac-
ters from novels in the form of the personas they embody. In this paper, we define
and introduce a method to extract such personas of characters in fiction novels, in the
form of descriptive phrases. These personas are divided into three types of descrip-
tion—facts, states and feelings. We show that such a model performs satisfactorily
returning an extraction precision of 91% and average classification accuracy of 80%.
The algorithm uses universal dependency trees, POS tags and WordNet to capture
semantically meaningful descriptions of characters portrayed. The results have the
potential to serve as input for future NLP tasks on literature fiction like character
clustering and classification using techniques such as sentence embeddings.

Keywords Universal dependency trees · Part-of-speech tags · WordNet · Natural

language processing

1 Introduction

A persona is defined as the aspect of a person’s character that is presented to or

perceived by others. It can include a person’s behaviour in different situations, opin-
ions on different matters, body language and background, among a score of other
parameters. In literary fiction, the persona a character displays is much easier to
define. Here, the motives, beliefs, behaviour and quirks of a character are explicitly
presented, unless intentionally hidden by the author. This paper presents an algo-
rithm that exploits the descriptions of characters given in a novel in order to extract
their personas. A persona is defined here as an encapsulation of three facets—facts,
states and feelings. Facts include character traits and attributes, (physical, cognitive,
or otherwise) as well as indisputable certainties describing the person. Facts could

N. Prabhu (B) · S. Natarajan

Department of CSE, PES University, Bangalore 560085, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 65

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_6

[email protected]
66 N. Prabhu and S.Natarajan

be ‘hates exercise’, or ‘She was a squat, smiling woman’. States include physical
and emotional states, including ‘howled with laughter’, and ‘ran home as fast as
he could’. Feelings refer to directly affective states, both physical and emotional.
Examples of valid feelings are ‘awkwardness’, and ‘emptiness’ with phrases such as
‘jealous and angry’, and ‘was hungry’ as descriptions.

2 Background and Related Work

Bamman et al. [1, 2] divided a persona into four categories based purely on typed
dependency relations. These are:
(i) ‘agent’ for verbs that have an (nsubj)/agent relation with a character,
(ii) ‘patient’ for verbs that have a (dobj)/nsubjpass relation with a character,
(iii) ‘predicative’ for a noun/adjective that has an (nsubj) relation with the character,
with an inflection of the word be as a child, and
(iv) ‘possessive’ for words that have a poss relation with a character.
The personas extracted through this mechanism, however, are constituted by only
single words, like ‘strangled’, or ‘angry’, including many trivial ones like ‘told’,
and ‘came’, which, when they do provide a meaningful pointer to the character’s
personality, do so in a very limited sense, lacking context.

3 Proposed System

The philosophy behind the algorithm is that every non-trivial feature worth describ-
ing a character is done so with an adjective. The algorithm implements this by finding
valid mappings between characters and adjectives in each sentence based on several
factors as shown in Step 1. These come together to achieve the goal of the task—ex-
tracting concise and comprehensive phrases that serve as descriptions. POS tags and
universal dependency trees used are extracted using the Stanford CoreNLP toolkit
(Chen and Manning [3], Manning et al. [4], Schuster and Manning [5], Toutanova
et al. [6]). The four steps of the algorithm are outlined below.
Step 1: Extracting Valid Mappings
In each sentence containing at least one person and adjective, each possible pair
is checked for validity by examining their dependency tree relationships, among
other conditions. The four relationships exploited for extraction and classification of
phrases are shown in Table 1.
Here, a person is any word with the dependency nsubj that refers to a character
in the novel. An adjective is any word with the POS tag JJ. In Fig. 1, Harry and
Malfoy are persons, while ‘bad’ and ‘quick’ are adjectives. Each mapping that meets
the conditions for the relationship it has, as shown in Table 2, is considered valid

[email protected]
Extraction of Character Personas from Novels Using … 67

Table 1 Valid relationships for a person-adjective mapping

descendant ancestor

(g)_niece sibling

Fig. 1 Bare dependency tree

for the phrase

and proceeds to the next step. nsubj children refer to those nodes that the adjective
in a person-adjective mapping actually describes, rather than the person itself. This
occurs when an adjective is related to a node as its parent, or as its sibling or niece
along with the additional conditions described in Table 3 (Tables 2, 3 and 4 are
described in Prolog, some of whose syntax is described in Table 5. These tables are
described at the end). Checking for these weed out invalid mappings like that shown
in Fig. 1, where the adjective ‘quick’ describes Malfoy instead of Harry, making the
extracted phrase ‘quick to notice’ inappropriate for Harry, but apt for Malfoy.
Step 2: Adding Dependents
There are several dependent words which are added to the describing phrase that are
vital to the coherence and conciseness of the description returned. These dependents
and their conditions of addition are described in Table 6. For example, in the sentence
‘He drifted into an uneasy sleep, thinking about what happened’, adding the parents
of the person and the adjective extracts ‘drifted into an uneasy sleep’ instead of just
‘uneasy’, which gives a very different meaning. One particular condition checked

[email protected]
68 N. Prabhu and S.Natarajan

Table 2 Procedure for ensuring validity and classifying type of person-adjective mapping
relationship(adjective, person, X):-
(sibling(adjective, person),
non_trivial_dep(adjective) ->
X = sibling;
(ancestor(adjective, person),
not(nsubj_child(parent, adjective, NODE)),
(no_valid_splitters(adjective, person, punctuations);
not(VB_in_lineage(adjective, person)))) -> X = ancestor;
(g)_nib_desc_condition(parent, adjective) ->
dep(adjective, dobj)) ->
((g)_niece(adjective, parent) -> X = (g)_niece);
ancestor(parent, adjective) -> X = descendant).

(g)_nib_desc_condition(parent, adjective):-
(non_trivial_dep(adjective), not(nsubj_child(parent, adjective, NODE))).

VB_in_lineage(A, B):-
(ancestor(A, B), ancestor(V, B), ancestor(A, V), pos(V, ‘VB’)).

non_trivial_dep(A):-
not(dep(A, det); dep(A, case)).

sibling(A, B):- parent(adjective, PA), parent(person, PA).

(g)_niece(A, B):- ancestor(B, X), sibling(X, A).

ancestor(A, B):- parent(A, B); parent(A, X), ancestor(X, B).

Table 3 Definitions of sub-procedures and terminology used

nsubj_child(person, adjective, NODE):-
(not(same_name(person, NODE)), dep(NODE, nsubj), dep(person, conj)),
not(parent(person, NODE), (parent(NODE, adjective);
((relationship(NODE, P, sibling); parent(NODE, P)),
(pos(P, ‘VB’), no_valid_splitters(person, NODE))),
present_after(person, NODE), parent(adjective, P))).

no_valid_splitters(X, Y, PUNCT LIST):-

(word_in_between(X, Y, Z),
not(conjunction splitter(Z); punctuation_splitter(Z, PUNCT LIST))).

conjunction_splitter(NODE):-
is(NODE, conjunction), adjacent words condition(NODE).

punctuation_splitter(NODE, PUNCT LIST):-

(is(NODE, punctuation), not(descriptive comma(NODE)),
adjacent_words_condition(NODE), not(descriptive pair(PUNCT LIST))).

descriptive_comma(NODE):-
(punct_type(NODE, comma), dep(NODE, D), member(D, [amod, conj, advcl]),
pos(NODE, ‘RB’); pos(NODE, ‘JJ’)).

descriptive_pair(PUNCT_LIST):-
(length(PUNCT LIST, 2), type(PUNCT_LIST[0], P1), type(PUNCT_LIST[1], P1)).

adjacent_words_condition(NODE):-
(previous_word(NODE, P), next_word(NODE, N), not(pos(P, ‘RB’), pos(P, ‘JJ’), pos(N, ‘RB’),
pos(N, ‘JJ’)), dep(P, DP), not(dep(N, DP))).

same_name(person, NODE):-
name(NODE, NM), name(person, NM)).

[email protected]
Extraction of Character Personas from Novels Using … 69

Table 4 Procedure for classifying extracted phrases into fact, state or feeling
facet(phrase, X):-
(person_condition(person) -> X = Fact;
relationship(adjective, person, sibling) ->
(feeling condition -> X = Feeling;
parent(adjective, P), pos(P, ‘VB’) ->
(word_in_between(person, adjective, ‘in’),
no_valid_splitters(person, adjective, punctuations);
root(sentence, adjective) -> X = State; X = Feeling);
sibling_state_condition -> X = State);
relationship(adjective, person, ancestor) ->
(previous_word(adjective, PR), dep(PR, det) -> X = Fact;
feeling_condition -> X = Feeling;
desc_state_condition -> X = State; X = Fact);
(relationship(adjective, person, (g)_niece);
relationship(adjective, person, descendant)) ->
feeling_condition -> X = Feeling;
parent(person, P), pos(P, ‘VB’) -> X = States; X = Fact).

person_condition(PER):-
(POS condition(PER); succeeding_word_condition(PER); have condition(PER)).

POS_condition(PER):-
((parent(PER, P), adjacent_word(PER, A), (pos(P, ‘VBZ’); pos(A, ‘VBZ’))).

succeeding_word_condition(PER):-
(succeeding word(PER, S), (inflection(S, ‘be’); inflection(S, ‘have’)),
(root(sentence, S); parent(S, P), not(pos(P, ‘VB’)))).

have_condition(PER):-
((relationship(‘have’, PER, descendant); relationship(‘have’, P, (g) niece),
parent(PER, P)), no_valid_splitters(‘have’, PER, punctuations)).

feeling_condition:-
(present_in(adjective, WordNet feelings), (not(dep(P, dobj); pos(P, ‘NN’)),
parent(adjective, P)); (inflection(N, feel), nearby word(adjective, N))).

sibling_state_condition :-
(not(dep(adjective, dep)), child(adjective, C), not(dep(C, nsubj)),
(dep(adjective; nmod) -> parent(adjective, P), not(pos(P, ‘JJ’)))).

desc_state_condition:-
(child(adjective, C), not((inflection(C, ‘be’)); inflection(C, ‘have’));
root(sentence, adjective); no_valid_splitters(person, adjective, punctuations),
word_in_between(person, adjective, ‘in’); pos(adjacent words(adjective), ‘VBG’);
adjacent_words(adjective, A), pos(A, ‘VBD’)).

[email protected]
70 N. Prabhu and S.Natarajan

Table 5 Prolog syntax

X ; Y => Try X first; if it fails, try Y.

X -> Y => If X, then try Y, otherwise fail.

X -> Y ; X => If X, then try Y, else try Z

X, Y => Try X, then try Y

for is the presence of ‘splitters’ in the form of punctuations or conjunctions that split
the phrase, eliminating redundant details. A conjunction that splits a phrase into two
parts—of which one part describes something other than the ‘person’, is considered
as a valid splitter. A descriptive conjunction that has its adjacent words as describers,
like ‘and’ in ‘kind and sweet’, is invalid. A punctuation may similarly split a phrase
but will be considered invalid if it is either a descriptive comma, an aside-describing
identical pair, or fails the conditions for a conjunction splitter described in Table 3.
Examples of invalid punctuation splitters are present in ‘good, honest, thoughtful
man’, and ‘Bob, sitting on the chair, laughed’, with the first containing descriptive
commas and the second being a descriptive pair.
Step 3: Refinement
The words extracted till now may not form a contiguous phrase, so any gaps are
filled. The resulting phrase is divided into sub-phrases by valid punctuation splitters,
if present. The sub-phrases that satisfy any of the following conditions are then added
to the description:
(i) It contains the adjective of the mapping.
(ii) It immediately follows the person of the mapping.
(iii) The first word in the sub-phrase is the person.
Again, any gaps present are filled and any stop words present at the ends are
removed, giving the final description.
Step 4: Classification
The procedure for classification shown in Table 4 follows from validity checking,
where three of the initial conditions for classification follow from the type of relation-
ship between the person and adjective as found in Table 1, along with a new condition
person_condition(PER) on words directly related to the person in the mapping. Word-
Net_feelings is a set used in the decision to classify a phrase as a feeling. It is the
set of all the hyponyms of the synset feeling.n.01 present in the WordNet database.
Checking if a word is present here is complicated by the fact that adjectives are stored
as clusters rather than trees, making a direct search ineffective. Table 7 describes a
procedure that uses the presence of lemmas present for each word to check if a word
is present in this set.

[email protected]
Extraction of Character Personas from Novels Using … 71

Table 6 List of dependents to be added to the descriptive phrase

1 All the words between the person and the adjective if the dependency of the adjective is
xcomp
2 All negative words that are either
i. child or sibling of the adjective, or
ii. are present between the person and the adjective in the sentence
3 i. All words with dependency nsubj that don’t represent a character and do not have
POS tag based on any of PRP, NNP, WP, DT
ii. All words with POS tags VBZ, VBD or VBG that are adjacent to the person
iii. All nodes with the POS tag JJ and the dependency advcl or acl:relcl, if they are
either siblings or children of the adjective or the person, respectively
4 All valid prepositions—children of the adjective with POS tag IN that don’t have the
dependency mark
5 i. All adverb children of the adjective
ii. All siblings or children of the adjective that have either advmod or det dependency
which are preceded by a word having an an inflection of ‘be’ or ‘have’
6 i. All nmod nodes that are either children, siblings, or parents of the adjective
ii. All conj children related in the same way as long as they don’t have POS tags VB
or VBD
iii. In addition, all prepositions that mark these nmod or conj words, for each of these
found
7 i. The parent of the person if it has POS tag VB and there are no valid punctuation
splitters between it and its parent. (Table 3, punctuation_splitter(X, Y, PUNCT
LIST))
ii. The parent of the adjective if the adjective’s dependency is any of nmod, conj,
xcomp, dobj, amod, advmod and there are no valid splitters between it and its
parent
8. i. All valid prepositions that are either siblings or children of the adjective, and
ii. all words with dependency cc that are either
(a) related to the adjective in the same way or
(b) are the parent of any node added to the list till now, as long as the words are present
after the person and the adjective in the sentence
9 All punctuations as long as the only punctuations extracted are not an aside-describing
identical pair, (Table 3, descriptive pair(PUNCT LIST)) as well as all ‘as…as’
comparative pairs
10 i. All words with dependency neg that are parents of any node in the list, and
ii. All neg words present before the adjective as long as the adjective has no valid
splitters between it and its parent, and the person in the mapping is not the first
person described in the sentence
11 All children nodes of nodes in the list that have either
i. dependencies xcomp, ccomp, nmod, conj, dobj, cc or aux, or
ii. have POS tags VB, RB, WRB, extracted recursively until there are no more to be
added

[email protected]
72 N. Prabhu and S.Natarajan

Table 7 Procedure for checking if a word is present in WordNet_feelings

1 Get the synsets of the word, and convert each to its corresponding lemma and store them
in a list
2 For each lemma, get its derivationally related forms and store them along with the
existing lemmas in the list
3 Convert each lemma in the list to its corresponding synset, if it exists
4 If any of the synsets are present in WordNet_feelings, the word is considered present. If
not, extract synsets of words

Table 8 Performance of the algorithm

Task Precision Recall Negative Specificity F1 score Accuracy
predicate value
Validity of 0.91 0.62 0.39 0.77 0.74 0.67
phrase
Classification 0.66 0.53 0.85 0.91 0.55 0.80
as fact
Classification 0.75 0.86 0.73 0.58 0.79 0.73
as state
Classification 0.60 0.48 0.91 0.95 0.53 0.88
as feeling
Average 0.67 0.62 0.83 0.81 0.63 0.80
classification

4 Results and Inferences

The algorithm was evaluated using an unbiased human evaluator who marked a
randomly generated subset containing both valid phrases (extracted by the algorithm)
and invalid ones (phrases filtered out by the conditions of the algorithms as in Table 4)
for the books ‘A Picture of Dorian Grey’ and ‘The Adventures of Huckleberry Finn’.
Each phrase was marked as valid or invalid, and if valid, it was marked as one class
out of Fact, State or Feeling. The total number of phrases evaluated amounted to
985. To perform the evaluation, one must be able to distinguish whether a phrase is
describing a given person or not. In addition, one must be able to confidently state
whether such a description falls under the category of a fact, state or feeling as clearly
distinguished in the Introduction. As this task would be simple for a person with
sufficient command of the English language and would bring very few ambiguities
that need to be resolved, one evaluator was considered sufficient for the purpose of
evaluation. Table 8 shows the results of the evaluation of validity of phrase extraction
and phrase classification. The precision of phrase validity was 91% while the average
accuracy of classification was 80%. Table 9 shows a handpicked subset of phrases
that represent the persona of the character Dorian Grey from the novel ‘The Picture
of Dorian Grey’, as extracted by the algorithm.

[email protected]
Extraction of Character Personas from Novels Using … 73

Table 9 Sample phrases from the persona of the character ‘Dorian Grey’
Facts Is never more present in my work than when no image of him is there
As a rule, he is charming to me
Now and then, however, he is horribly thoughtless
Seems to take a real delight in giving me pain
Is my dearest friend
Has a simple and a beautiful nature
Does not think his natural thoughts, or burn with his natural passions
Was brilliant, fantastic, irresponsible
But he has been most pleasant; is far too wise not to do foolish things now and then
States Burying his face in the great cool lilac-blossoms
Saw it creeping into the stained trumpet of a Tyrian convolvulus; murmured,
flushing at his own
Boldness
Had recognized himself for the first time
Stood there motionless and in wonder, dimly conscious that Hallward was
speaking to him; was Reclining in a luxurious arm-chair
Will tell you more
Answered in his slow melodious voice
Replied, touching the thin stem of his glass with his pale, fine-pointed fingers
Had grown years older
Grew pale as he watched her
Feelings Far too charming
But he felt afraid of him, and
Ashamed of being afraid
Seemed quite angry
Was puzzled and anxious
Looked pale, and proud, and indifferent
Felt that he was on the brink of a horrible danger

5 Conclusion and Future Work

Extracting semantic summaries from the text is a useful task, and the results of the
algorithm described can be used as a basis for further literary and computational anal-
ysis. Skip-thought vectors (Kiros et al. 2015) [7] are representations of sentences in
the form of embeddings similar to word embeddings. These represent sentences in
the form of vectors where a sentence is represented by the sentences surrounding
it. They have been used to successfully perform tasks of semantic relatedness, para-
phrase detection and classification. This model can be used on the personas extracted
for the purpose of a variety of tasks. These can include clustering similar personas
of characters across novels, classifying character personas into a set of archetypes.

Acknowledgements We would like to thank Ms. V. Sruthi for her help in evaluation of the 985
phrases present in the test set for the purpose of determining the performance of the algorithm.

[email protected]
74 N. Prabhu and S.Natarajan

References

1. Bamman, D., O’Connor, B., & Smith, N. A. (2014a). Learning latent personas of film characters.
In: Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL)
(p. 352).
2. Bamman, D., Underwood, T., & Smith, N. A. (2014). A Bayesian mixed effects model of literary
character. ACL, 1, 370–379.
3. Chen, D., & Manning, C. D. (2014). A fast and accurate dependency parser using neural net-
works. In: EMNLP (pp. 740–750).
4. Manning, C. D., Surdeanu, M., Bauer, J., Finkel, J. R., Bethard, S., & Mc-Closky, D. (2014).
The Stanford CoreNLP natural language processing toolkit. In ACL (System Demonstrations)
(pp. 55–60).
5. Schuster, S., & Manning, C. D. (2016). Enhanced English universal dependencies: An improved
representation for natural language understanding tasks. In Proceedings of the 10th International
Conference on Language Resources and Evaluation.
6. Toutanova, K., Klein, D., Manning, C. D., & Singer, Y. (2003). Feature-rich part-of-speech
tagging with a cyclic dependency network. In Proceedings of the 2003 Conference of the
North American Chapter of the Association for Computational Linguistics on Human Language
Technology-Volume 1. Association for Computational Linguistics (pp. 173–180).
7. Kiros, R., Zhu, Y., Salakhutdinov, R. R., Zemel, R., Urtasun, R., Torralba, A., & Fidler, S. (2015).
Skip-thought vectors. In Advances in neural information processing systems (pp. 3294–3302).

[email protected]
A Multifactor Authentication Model
to Mitigate the Phishing Attack
of E-Service Systems from Bangladesh
Perspective

Md. Zahid Hasan, Abdus Sattar, Arif Mahmud and Khalid Hasan Talukder

Abstract A new multifactor authentication model has been proposed for

Bangladesh taking cost-effectiveness in primary concern. We considered two-factor
authentications in our previous e-service models which were proven to be insufficient
in terms of phishing attack. Users often fail to identify phishing site and provide con-
fidential information unintentionally, resulting in a successful phishing attempt. As a
result, phishing can be considered as one of the most serious issues and required to be
addressed and mitigated. Three factors were included to form multifactor authentica-
tion, namely, user ID, secured image with caption, and one-time password. Through
the survey, the proposed multifactor model is proven to be better by 59% points for
total users which comprises 55% points for technical users and 64% points for non-
technical users in comparison to traditional two-factor authentication model. Since
the results and recommendations from the user were reflected in the model, user
satisfaction was achieved.

Keywords Phishing attack · E-banking · E-service · Online banking

1 Introduction

At present, e-services can be considered as the most substantial concerns in our day-
to-day life. The efficient utilizations of computerized knowledge and tools expand
along with our requests. However, the ideas and tools will not be adequate to upkeep
novel technologies with the hi-tech devices in modern e-service system.

Md. Zahid Hasan (B) · A. Sattar · A. Mahmud · K. H. Talukder

Daffodil International University, 102 Sukrabadh, Mirpur Road, Dhaka 1207, Bangladesh
e-mail: [email protected]
A. Sattar
e-mail: [email protected]
A. Mahmud
e-mail: [email protected]
K. H. Talukder
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 75

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_7

[email protected]
76 Md. Zahid Hasan et al.

We proposed models for two e-services, namely, e-healthcare [1] and mobile pay-
ment [2] system that guarantees accomplishing the objectives and requirements, as
supported by our previously suggested framework models [3]. Besides, the models
can have an important role in e-commerce sector which can lead to a giant accom-
plishment in terms of upgrading of e-care services. However, the suggested models
should be secure and reliable in terms of prevention of phishing attack which has
become a concern now a day.
Phishing is such kind of attack which exploits the weakness of end user of a
system [4]. A huge number of internet users cannot able to differentiate phishing and
original site, and they ignore passive warnings like toolbar indicator [5]. The aim of
phishing is to steal sensitive information like password, credit card information from
Internet user, etc. Therefore, authentication needs to be ensured in order to reduce
phishing attack.
Most of the e-services in Bangladesh use two-factor authentications (2FA) which
are insufficient to defend the phishing attack. As a result, multifactor authentication
needs to be implemented. Numerous security factors are being used for authentication
in economically advanced countries but it is not viable in a developing country like
Bangladesh where the cost-effectiveness in authentication is a challenging task.
The main purpose of a multifactor authentication solution is to verify the iden-
tity of the user by other means apart from the traditional username and password
authentication. These factors are usually combined with a PIN code that generates
one-time passwords (OTP). Based on the earlier papers, we have investigated the
login behavior of users and how the user compromises their login credentials with
phishing site through the survey. Along with this, we proposed a new authentication
mechanism in the context of developing country like Bangladesh to support our pre-
viously developed models. The specific objectives of this research can be listed as
follows:
• To propose a new multifactor authentication mechanism being supported by pre-
viously proposed e-service model.
• To exploit the login behavior and to figure out the user’s experience after applying
the proposed approach.
This model can be applied to stop deceitful emails, texts, and copycat websites
which share valuable information such as social security numbers, account numbers,
login IDs, and passwords. Besides, it will be beneficial in prevention of stealing
money and identity or both.
This paper is designed in the following way: Sect. 2 illustrates the background;
Sect. 3 clarifies the previously proposed models; Sect. 4 describes the proposed MFA
model; Sect. 5 illustrates the results; the discussions and conclusions are added in
Sect. 6.

[email protected]
A Multifactor Authentication Model to Mitigate … 77

2 Background

Phishing can be considered as the social engineering attack which targets at manip-
ulating the limitation of the system procedure created by system users [4]. As for
illustration, a system might be technically safe against the attack like password theft
but passwords might get leaked from the end user when an attacker insists to update
their respective passwords through a fake HTTP link. So, this phenomenon threatens
the complete system security. In addition, we can take DNS cache poisoning, for
example, which can also be utilized by the attackers to create more influence on
social engineering messages [4].
According to Weider [6], the major motives behind phishing attack are as follows:
i. Financial benefit: Phishers can utilize the users’ banking credentials to gain
financial profits.
ii. Identity hiding: Phishers can advertise the users’ identities in the market and
can be used later by criminals who hide their authentication and activities from
the authority.
iii. Reputation and disrepute: Phishers can attack the victims for the peer recogni-
tion.
Due to abovementioned reasons, phishing attack can be considered as serious one
and an effectual lessening will be required.
Multifactor authentication involves the use of two or more independent security
factors to authenticate an information system and to lessen phishing attack. Unfor-
tunately, several technological, economical, and usability limitations exist that resist
sector-wide acceptance of MFA.

3 Previous Model

We have anticipated frameworks for e-care systems in our earlier works those mini-
mize the misinterpretation about design, development, and communication systems
and develop the harmonization among data, service, and management. E-services
can be considered as one of the most important factors, and the projected framework
ensures the ease of access of services anytime, anywhere without being delimited by
any periphery. These two models are described below in brief.

3.1 Contextual Mobile Payment System Implementation

Five players or entities are included in the projected system shown in Fig. 1 named
user, network operator, merchants, bank, and cell phone manufacturer that are placed
in the management level of ICTization model. Integration of mobile payment system

[email protected]
78 Md. Zahid Hasan et al.

Fig. 1 Contextual mobile payment system [3]

entities is the significant factor to develop a global open solution model in lieu of a
closed system given a small scope. The functionalities of these entities are as briefly
followed:
• Customers or mobile users are capable to purchase the products from merchants
and can make transaction even with others through mobile phones.
• The merchant play the role of a middleman who provides the requested services
for the users.
• The network operators are responsible for the network connectivity in mobile
devices and manage the user’s authenticity.
• Bank plays the role of a financial institute that is responsible for payment procedure
and monitors the flow of transaction.
• Device manufacturers are responsible to develop a generic ground for payment
functions as the demand of users.
The projected model allows subscribers to buy products, balance inquiry, and
money transfer to bank accounts through mobile SMS. It will enable users to check
the transaction status and history such as the amount is left to use, the amount has
been spent in the last day, week, and month, transaction acknowledgement, balance
notification, etc.

[email protected]
A Multifactor Authentication Model to Mitigate … 79

For instance, when a subscriber desires to buy a transportation card, he/she can use
mobile phone to pay directly from his/her own bank account. According to the steps
in Fig. 8, user plans to obtain any services and/or goods utilizing mobile devices with
projected application from different kinds of services integrated with the application.
The apps/icons which are included in the mobile phones will be used to provide the
desired facilities to the consumer. Users pay for the applications or services and
proceed to advance activities. The procedure of these sending and receiving services
is happened between the user and the provider where merchant acts as the third party
who becomes connected with both ends and ensures authentication.
If a consumer desires to pay for the products and/or services, merchant will inform
the bank about the particular client and provide support to accomplish the financial
transaction. Afterward, the bank will conclude the financial transaction with con-
sumer and provider and inform the merchant to give the specific products or services
to the client. At last, the merchant will offer the services or products to the consumer
and the consumer can consume the products and/or services without any disturbance.

3.2 Contextual E-Healthcare Model

The anticipated model supports ambient healthcare services as supported by our for-
merly planned model frameworks. Five entities are included in our system as shown
in Figs. 3 and 5 named as patient, merchants, home network, network operators, and
healthcare centers. The functions of these entities are given below in brief:
Sensors are integrated into patients and patients can utilize their home networks
for requesting and receiving the desired services. The home network can be consid-
ered as the combination of mobile devices like cell phone, laptop, and PDA with
Internet functionalities. The merchants plays the role of a middleman to define and
provide the services as requested by the patients. They will also communicate with
banks regarding financial transactions. The network operators like mobile operators
and ISPs will provide the network connectivity to mobile devices and manages the
subscriber’s authenticity (Fig. 2).
The healthcare service centers can be considered as the collection of hospitals and
diagnostics center together with expert physicians and researchers. These centers are
scattered in several cities and countries as well. To be noted, the patient information
(located in databases) is shared among these centers. As for illustration, patients can
appeal the services through their home network both manually and automatically.
If we consider the automatic service, the sensors are used to capture the raw data
like heart bit, pulse rate, motion, etc. and transmit the data to the home network. On
the contrary, patients can utilize their home networks to demand explicit services
manually where home network forwards the data to the merchant in order to verify
the validity and authenticity of the client. The data can be sent in two ways such as
IP traffic and mobile traffic as chosen by the patient. As a result, user can demand
the services both using email and cell phones.

[email protected]
80 Md. Zahid Hasan et al.

Fig. 2 Contextual e-healthcare model [2]

The procedure of sending and receiving services and application happened

between consumer and provider, and the merchant acts as the third party that commu-
nicates with both sides and ensure authentication. The merchant can communicate
with banks to meet financial transactions as well. As for illustration, if a consumer
needs to pay for any registrations or services, merchant can inform the bank that the
specific user wants to accomplish the financial transaction. Afterward, the bank will
conclude the financial transaction with client and provider and inform the merchant
in order to grant the specific service to the consumer. Next, merchant will inform
the consumer. To be noted, the healthcare centers sustain an interaction with both
service provider and merchant in order to confirm the legitimacy of the client. At the
end, the demanded services will be received to the consumer through the merchant.
To be noted, we considered 2FA in these proposed models which is proven to be
insufficient in terms of security. However, phishing attack is one of the main concerns
for these models which need to be addressed and mitigated. In addition, additional
cost to ensure authentication will affect the utilization of these models.

[email protected]
A Multifactor Authentication Model to Mitigate … 81

4 Proposed Multifactor Model

Most of the e-services in Bangladesh use 2FA by combining knowledge base and one-
time password (OTP) base together. Here, we have combined three factors together
to authenticate a user as shown in Fig. 3. An extra factor, security image with caption
for defending the phishing attack was included.
Figure 3 describes the authentication procedure between a user and authentication
server. At first, user puts the email ID on the site. According to user ID, server replies
his/her security image and caption. If the client verifies the image and caption as given
through registration time, password is put. Phishers can send similar login page via
email, SMS, or other media, but they cannot send security image and caption. When
server confirms a valid password, OTP is provided to the client. After the successful
OTP reply, authentication process will be completed.
Figure 4 shows the process flow of the authentication approach. When a user
needs to login into the site, he/she enters the username at first. According to the
username, the site displays the predefined image and caption which was set by the
user during the registration phase. If the image and caption match with previous one,
the password is put afterward.

Fig. 3 Authentication
procedure flowchart

[email protected]
82 Md. Zahid Hasan et al.

Fig. 4 Authentication
procedure

5 Result

As a part of our research, we made the survey on two types of users: technical and
nontechnical.
• Technical: The students who have knowledge on computer applications
• Nontechnical: The students who do not have much knowledge on computer appli-
cations.
52 students were randomly selected from different faculties where 27 of them were
technical students and 25 nontechnical students. Two different look-a-like sites for
login were provided where one is valid site and another is phishing site with different
URLs. After that, data were collected from user who logged into the phishing site
or who denied to login. We made the survey in two different phases: (1) Before
registration and (2) after registration. Before registration, as can be seen from Fig. 5,
72% of the total students were unable to identify the phishing site which comprises
62% technical and 80% nontechnical students.
After that, users have to complete the registration. From Fig. 6, we can see users
have to register themselves with their name, email ID, and password. Most impor-
tantly, users have to select their own image and caption which will be used for
authentication purpose.

[email protected]
A Multifactor Authentication Model to Mitigate … 83

Fig. 5 Phishing result with 2FA model

Fig. 6 Registration
procedure

Thereafter, the security image is verified in three steps.

Step 1: The registered user gives his/her user ID, as shown in Fig. 7.
Step 2: Server returns his/her security image and caption (Fig. 8) according to the
email ID.
Step 3: The user confirms security image and password will be required to enter the
desired site.
After registration, it can be seen from Fig. 9 that only 13% of total students failed
to identify the phishing which includes 7% technical and 16% nontechnical students.

[email protected]
84 Md. Zahid Hasan et al.

Fig. 7 Enter only user ID

Fig. 8 Security image

If we compare our multifactor authentication model with tradition 2FA model, the
result is improved by 59% points for total students, 55% points for technical and
64% points for nontechnical students.
Next, we made an another survey based on two factors:
i. Agreeable: Users find this model effective and would like to continue.
ii. Annoying: Users find this model complicated and denied to continue.
Based on Fig. 10, it can be stated that most of the users find this model an efficient
one to prevent phishing attack. 88% of total students agreed to continue where only
11% denied. In addition, the difference between agreeable and annoying is 84% for
technical and 66% for nontechnical students.

[email protected]
A Multifactor Authentication Model to Mitigate … 85

Fig. 9 Comparison between 2FA and MFA model

Fig. 10 Comparison between agreeable and annoying

6 Discussion and Conclusion

This paper is developed on previously developed e-service model which can extend
the effort on enlightening depiction of ICT setup. Taking the business into account,
this paper will assist in providing a system to categorize detailed collection of ICT
infrastructural features to achieve a secured authentication. On contrary, from the
social standpoint, this paper will offer a collective framework with a goal to obtain
secured citizen concerned services. We used 2FA in our earlier proposed e-service
model. To be noted, password is the key element for single-factor authentication
which cannot be considered as secures in the present e-commerce and Internet world.

[email protected]
86 Md. Zahid Hasan et al.

Attackers can use sophisticated applications to find out easy guessing password
like name, birthdate, phone number, etc. 2FA techniques are currently being used
by most of the e-services in Bangladesh. The combination of knowledge base and
possession base factor like one-time password (OTP) also cannot avoid phishing
attack. In this paper, by carrying out survey, we collect login behavior of 52 students
with phishing site. Using survey data, we have found that 72% of total students
compromise their password with phishing site and then we propose three-factor
authentication mechanisms by combining existing 2FA with security image.
The proposed multifactor model is experienced to be better than traditional 2FA
model by 55 and 64% points in case of technical and nontechnical students. Impor-
tantly, most users have found this model effective and have decided to continue with
the suggested approach. In short, this paper aimed to validate previously proposed
models in terms of authentication without adding any extra cost taking developing
country like Bangladesh in concern. Besides, we also addressed the user partici-
pation in accepting the multifactor authentication model. Therefore, surveys were
conducted to compare the traditional 2FA models with the proposed multifactor
authentication model. The outcomes and suggestions are set in the designed model
where the cost-effectiveness was one of the primary concerns. Besides, the execution
of this model can play noteworthy part in m-commerce that will have an enormous
success in terms of authentication of e-service system in Bangladesh.

References

1. Mahmud, A., & Sattar, A. (2013). ‘ICTization framework’: A conceptual development model
through ICT modernization in Bangladesh, Published. In Advanced Computer Science Applica-
tions and Technologies (ACSAT), 2013 International Conference, Malaysia. 23–24 Dec. 2013,
19 June 2014, 978-1-4799-2758-6, Publisher: IEEE.
2. Mahmud, A., & Sattar, A. (2014). Deployment of contextual mobile payment system: A prospec-
tive e-service based on ICTization framework from Bangladesh perspective. In Proceedings
of the International Conference on Advances in Computer Science and Electronics Engineer-
ing—CSEE 2014, Copyright © Institute of Research Engineers and Doctors. All rights reserved.
ISBN: 978-1-63248-000-2.
3. Mahmud, A., & Sattar, A. (2016) Deployment of contextual E-healthcare system: A prospective
e-service based on context aware conceptual framework and ICTization framework model. In
2016 IEEE 11th Conference, Hefei, China, 5–7 June 201624 October 2016, Electronic ISBN:
978-1-4673-8644-9, 978-1-5090-2605-0, Publisher: IEEE.
4. Khonji, M., Iraqi, Y., & Jones, A. (2013). Phishing detection: A literature survey. IEEE Com-
munications Surveys & Tutorials, 15, 2091–2121.
5. HUANG, H., TAN, J. & LIU, L. (2009). Countermeasure techniques for deceptive phishing
attack. In International Conference on, 2009. New Trends in Information and Service Science,
NISS’09. IEEE, pp. 636–641.
6. Weider, et al. (2008). A phishing vulnerability analysis of web based Systems. In IEEE Sympo-
sium on Computers and Communications. ISCC 2008. 2008. IEEE, 326–331.

[email protected]
Study on Energy-Efficient
and Lifetime-Enhanced Clustering
Algorithm-Based Routing Protocols
in Wireless Sensor Network

Ishita Banerjee and P. Madhumathy

Abstract The sensor nodes in wireless sensor network (WSN) come with limited
battery life, thus energy usage in WSN is to be handled with great care. The routing
protocols are the major research area to work on the QoS of the network such as
network lifetime, scalability, energy consumption, packet overhead, etc. Since sensor
node posses limited battery life, to use nodes efficiently yet not loosing connectivity
in the network becomes a major issue in designing the routing protocols. To achieve
energy efficiency and better network lifetime, grouping of sensor nodes into small
clusters and selecting one cluster head seems to have more advantages in comparison
with other network models to get better scalability, robustness and end-to-end data
delivery. In this paper, we have discussed and studied the different energy-efficient
routing protocols for clustering of sensor nodes in wireless sensor network, its merits,
demerits and applications.

Keywords WSN · Energy-efficient routing protocols · Clustering algorithms

1 Introduction

Wireless sensor network (WSN) is a widely used area of communication in recent

days that faces few networking issues while implementation such as different appli-
cation fields, limited resources availability and functionality, packet dimensions, and
multi-hop transmissions for mobile users in dynamic scenario. Applications of WSN
are widely spread and hardly left any area untouched such as environmental and habi-
tat monitoring [1] such as water quality [2], river/flood monitoring [3], fire detection
and rainfall observation in the field of agriculture, and also extends its applications in
the field of military [4], coal mines [5], education [6], health monitoring [7] as well
as medical diagnostics, and many more. Wireless sensor network basically consists
of nodes which might have sensors to gather data. A common working of such WSN

I. Banerjee (B) · P. Madhumathy

Department of ECE, Dayananda Sagar Academy of Technology and Management,
Bangalore 560082, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 87

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_8

[email protected]
88 I. Banerjee and P. Madhumathy

is collection or generation of data in the sources and sending the data to the intended
sinks through wireless communication. This whole network is based on few inter-
mediate nodes. The wireless communication that takes place here is through radio
which is energy-expensive method. Effective use of this kind of transmission causes
the nodes to loose energy and eventually resulting in data collection loss and loss
or delayed data delivery. Though loss of few nodes is tolerable, loss of some highly
dependable or critical node may cause serious harm to the end-to-end connection.
Routing protocol classifications are based on functioning mode, participation style
and network structure. In the study of Li et al. [8], on the basis of functioning mode,
it can be subdivided into proactive routing protocols, reactive routing protocols and
hybrid routing protocols [9]. In the first case, i.e. proactive routing protocol, data is
sent to the base station through predefined path. In reactive routing protocol, path
is established on demand. Hybrid types already establish path initially but modify
according to the need for the improvement of the communication. On the basis of
network structure, it can be subdivided into data-centric routing protocol, hierarchi-
cal routing protocol and location-based routing protocols. The data-centric routing
protocols are sink initiated. Hierarchical routing protocols [10] are cluster-based,
energy-efficient routing protocols where low-energy nodes capture data and high-
energy nodes process and transmit data. Location-based routing protocols forward
data at each hop to the neighbour node which is geographically located nearest to the
sink. Here, the location of the nodes is to be known using GPS to find the optimum
path. Depending on participation style, the protocols are classified as direct commu-
nication, diffusion-based and clustering-based. In the study of Kaur and Kad [11]
and Yi [12], closely classifying, following are the protocols that are used according
to the way the nodes participate in transferring data to the destination.
• Direct communication (base station and nodes can communicate to each other
directly).
• Diffusion-based (this type of algorithm uses location data).
• E3D (majorly diffusion-based, uses location data as well as residual energy, node
utilization).
• Random clustering (chose cluster head randomly, collect all node data present in
that cluster and forward to base station).
• Optimum clustering algorithm (based on physical location and residual energy,
cluster head formation strategy changes after few rounds of communication for
optimum cluster formation).
The major issue in designing any wireless sensor architecture is limited battery life
of the sensor nodes which link the communication process. For effortless end-to-end
data delivery, increased lifetime of the network is required which could be achieved
by grouping few nodes and selecting one cluster head rather than randomly sending
the data directly through available nodes to the base station. As soon as a cluster is
formed and a cluster head is selected, the cluster head is in terms of maintaining the
data of each node in that cluster and is responsible for packet transmission to the
destination or sink node. In this paper, we shall discuss few energy-efficient routing
protocols for cluster-based network of nodes in WSN.

[email protected]
Study on Energy-Efficient and Lifetime-Enhanced… 89

2 Related Work

The few popular clustering-based routing protocols such as LEACH [13], PEGA-
SIS [14] and HEED [15] posses their own advantages and disadvantages in terms
of load balancing, lifetime of network, scalability and fields of application. LEACH
and HEED routing protocols are the clustering protocols which work in homogenous
WSN, whereas BEENISH works in heterogeneous platform. BEENISH utilizes four
different advanced energy levels improvising the previous protocols such as the two-
level heterogeneous ones like stable election protocol (SEP), DEEC [16] and DDEEC
[17], and also the three-level ones like EDEEC [18]. If network lifetime or through-
put is considered, then iBEENISH performs better than BEENISH. MBEENISH
and iMBEENISH give better results than BEENISH and iBEENISH if we consider
sink mobility model facilities [19]. An energy-efficient routing protocol (EERP) is
implemented to maximize network lifetime of WSNs using an optimal aggregated
cost function and A* algorithm [20]. In this paper, the authors have selected optimal
and shortest path between source and sink based on residual energy of sensor nodes,
free buffer of the nodes and link quality between sensor nodes. Here, A* algo-
rithm [21] is implemented to select optimal path to process the above-mentioned
parameters. Another improved routing protocol named EE-LEACH [22] yields bet-
ter performance in comparison with the LEACH by increasing packet delivery ratio,
reducing end-to-end delay, optimizing energy consumption and thus increasing net-
work lifetime. Cluster-based ACO BAN [23] creates a routing table from where
it selects minimum cost function value to find the optimal path in between source
and sink and provides less overhead, minimum number of packets delivered within
intermediate nodes and high data transmission rate. The protocol PDORP [24] uses
the characteristics of efficient gathering of sensor information system. It also uses
PEGASIS and DSR [25] routing protocols for reduction of energy consumption and
network overhead and faster response time, and ensures better connectivity of nodes.
The hybridization of genetic algorithm along with bacterial foraging optimization
[26, 27] used by PDORP proposes the optimal path. The E2HRC [28] routing proto-
col effectively balances the energy usage of wireless sensor network, thus decreasing
energy consumption of nodes as well as reduced numbers of control messages. Load-
balancing cluster-based protocol (LCP) [29] is a modified version of HEED protocol
to increase the energy efficiency of the network. A reliable, energy-balancing, multi-
group routing protocol (REM) [30] is implemented in the field, and data at critical
time is to be transmitted to the fire station real time with least delay to save the crisis
situation. Another routing protocol discussed here is PECE [31] that promises better
energy balancing of the network.

[email protected]
90 I. Banerjee and P. Madhumathy

3 Study of Different Routing Protocols

for Clustering-Based Design of Sensor Nodes

Here, we have discussed few clustering-based algorithm-based routing protocols for

wireless sensor network which all promise to be energy-efficient and also increases
network lifetime to provide better connectivity.

3.1 BEENISH

In this paper, BEENISH, iBEENISH, MBEENISH and iMBEENISH protocols for

heterogeneous wireless sensor networks (WSNs) are implemented. Balanced energy-
efficient network integrated super heterogeneous (BEENISH) selects four different
energy levels of sensor nodes and selects cluster heads (CHs) based on average energy
level of the entire network and residual energy levels of each node, whereas improved
BEENISH or iBEENISH uses different CHs selection schemes in an very efficient
manner by increasing the network lifetime. Here, a mathematical sink mobility model
is created which is implemented on BEENISH (Mobile BEENISH or MBEENISH)
as well as iBEENISH (Improved Mobile BEENISH or iMBEENISH). Finally, sim-
ulation results show that BEENISH, MBEENISH, iBEENISH and iMBEENISH
protocols perform better than other contemporary protocols for stability period, life-
time of network and throughput. The selected four energy levels of the nodes are as
follows: normal energy level, advanced energy level, super energy level and ultra-
super energy level. The initial energy levels for the nodes are least for normal and
gradually most for ultra-super level. Here, CH is not fixed but rotated amongst the
nodes by the probability, i.e. the ratio of residual energy level of each node in the
cluster to the average energy level of the network. Here, the nodes which have higher
energy levels are chosen as CHs which means the nosed with lower energy levels
are unused and ultra-super, super and advanced node have higher probability to be
chosen as CHs. Here, iBEENISH comes to an approach of solving this problem of
choosing only the higher energy nodes as CHs, and thus both BEENISH and iBEEN-
ISH together claim to improve the network lifetime. If sink mobility is considered,
then iBEENISH and BEENISH perform better rather than the non-sink mobility.

3.2 EERP Using A* Algorithm

The energy-efficient routing protocol (EERP) is implemented to increase the network

lifetime of WSNs using an optimal aggregated cost function along with A* algorithm.
A* algorithm helps to select the optimal path from the source node to the destination
node taking into consideration of residual energy of nodes, packet reception rate and
node buffer state. Here, the sink node is having awareness of criteria of each node

[email protected]
Study on Energy-Efficient and Lifetime-Enhanced… 91

to find the optimal path. Thus, at the initial state, all the nodes are supposed to send
the parameters to the sink node. Now while sending data to the sink node later, the
parameters of that particular node will be appended to data packet that is to be sent
to the sink node. Depending on the parameters that the sink node already gathers,
A* algorithm finds the optimal path. As the residual energy level of a node falls
below the predefined threshold level energy, that node can no more be considered as
a part of the optimal path, thus ensuring a better network lifetime. This node whose
energy has fallen below the threshold level need not send its parameters to the base
station/sink node anymore. Here, the network load is balanced depending on the
threshold value of energy of each node.

3.3 EE-LEACH

In this paper, EE-LEACH is implemented. Generally, the nodes which have maxi-
mum residual energy if participate in any routing protocol yield energy efficiency and
better network lifetime to the network, better packet delivery ration and less energy
utilization, better end-to-end delay. This makes only the highest residual energy
containing nodes to send their data to the base station. To implement EE-LEACH
probability of coverage is calculated from Gaussian distribution function. A list of
neighbour node is created with the residual energy data information of those neigh-
bouring nodes using one sorting algorithm. If the sink node is not near to the source
node, data ensemble source node cluster is formed in a relatively small area. Taking
into account the concentration degree of the source nodes and the residual energy,
the optimal cluster head is selected for this purpose. Though energy consumption is
reduced here, data confidentiality and integrity are not considered to greater extent.

3.4 Cluster-Based ACO BAN

In the field of health care, WSN plays a major role. The routing protocol named
cluster-based ACO BAN that is to be discussed here helps to monitor patient data
efficiently while sensitively takes care of the network lifetime and energy consump-
tion of the nodes. Here, cluster caves are formed, where nodes send their data to
the cluster head and cluster head sends its aggregated data to the net cluster head.
Sensitive and important data are passed to the destination effectively and reliably in
critical situations. Here, clusters are meant to send only critical information rather
than continuous information. At need, the critical information will be shared with the
corresponding healthcare department through only the cluster heads. A probabilistic
function is created to select the cluster heads in each level. In healthcare field since
the monitoring of physical parameters are required, link failure becomes a common
occurrence. Due to dead node in the transmission path, packet loss occurs and that
misuses the bandwidth and energy. To improve this drawback, minimum numbers of

[email protected]
92 I. Banerjee and P. Madhumathy

nodes are required to forward critical data to destination. This also reduces network
overhead. A method using CH-ACO algorithm enhances the accuracy of finding such
path. This is mainly based on ANT colony algorithm. The algorithm finds its path
from source to destination by connecting the neighbours and also updates data in the
routing table by collecting data from the nodes regarding their residual energy. To
avoid network overhead in this process, clustering concept is used which balances the
load of the network, reduces the number of intermediate nodes and finally increases
network lifetime.

3.5 PDORP

In wireless sensor network, several routing protocols are used to deal with issues
like reliability of the connection/ network, energy consumption, shortest path,
delay, network overhead, resource utilization, etc. PDORP, known as directional
transmission-based energy-aware routing protocol, is directed towards solving the
above-mentioned issues. In case of dynamic source routing (DSR) for small energy
density as the node switches from active to sleep, the data packet waits initially which
increases the end-to-end delay of packets and waiting time, thus decreasing the effi-
ciency of the network. This drawback increases energy consumption of the network.
In PDORP, the goal is to choose the dead nodes and a different path is identified by
conserving less energy. This protocol uses the characteristics of efficient gathering of
sensor information system, also PEGASIS and DSR to reduce energy consumption,
overhead, response time and connectivity issues. Hybridization of genetic algorithm
and bacterial foraging optimization helps to find the optimal path. PDORP creates a
trust list of the nodes that transmit information. This type of trust list that contains
the node parameters and which is updated after every transmission is useful at the
time of aggressive transmission and helps to prevent existing route damage. The
use of directional transmission decreases the path distance, and thus less energy is
consumed. Overall, this protocol provides less BER, less time to packet delivery,
less energy utilization, better throughput, ultimately leading to better QoS as well as
enhanced lifetime of the network.

3.6 E2HRC

There are several clustering algorithms and routing protocols in the field of wire-
less sensor network in order to encounter the problem of dying out the nodes due
to energy consumption which in turn breaks the communication channels between
the nodes. The E2HRC routing protocol effectively balances the energy usage of
wireless sensor network, thus decreasing the energy consumption of nodes as well
as reduced numbers of control messages. E2HRC is experimented in a ring-domain
heterogeneous communication topology. A topology control model is created in a

[email protected]
Study on Energy-Efficient and Lifetime-Enhanced… 93

ring domain where nodes are divided into separate levels as compared to their posi-
tions and after that ring domains are also divided. While sending and receiving data,
next hop node is selected with the optimal direction angle, thus reducing the energy
consumption. Based on the node residual energy and relative node position in the
cluster, the network is divided into different heterogeneous clusters and along with
this cluster head rotation mechanism helps to balance the node energy consumption.
E2HRC routing protocol uses two parameters as optimal direction angle along with
energy to balance energy utilization of the network.

3.7 LCP

The energy-aware distributed and dynamic load-balancing cluster-based protocol

(LCP) is a modified version of HEED protocol. HEED protocol’s clustering part is
modified to increase the energy efficiency of the network. The clustering method
has two phases known as the setup phase and the steady-state phase. Features of
LCP are as follows: The selected cluster heads advertise message within its own
cluster range. Selection of cluster heads after every round, i.e. re-clustering, is one
of the energy-consuming jobs which is eliminated in this protocol discussed here
by setting an initial time interval at the starting stage of every cycle; this helps
to delay in re-clustering message that is to be received from the base station. As
the nodes do not receive base station messages such sooner, the cluster head keeps
rotating the authority among the existing member nodes by selecting the one that
contains maximum remaining energy. LCP promises network lifetime increase in
a considerable figure in comparison with the existing protocols such as LEACH,
HEED, RHEED, etc.

3.8 REM

A variety of routing algorithms are designed in the field of wireless sensor network
that promises better performance in terms of delay, packet dropping, intrusion, energy
consumption, network lifetime, etc. A major real-life scenario arises in the field of
firefighter operation majorly known as firefighter communication where sensors are
implemented in the field and data at critical time is to be transmitted to the fire
station real time with least delay to save the crisis situation. A very fast dependable
and robust communication network is required in this situation. Design of a specific
routing protocol meets the specific need to save the crucial situation. A reliable,
energy-balancing, multi-group routing protocol (REM) is designed for reliability
and energy-balancing communication by selection metric-based CH, cluster rotation
and imposing a routing algorithm. Sensor nodes are assigned metric value, and the
nodes are chosen as cluster heads according to the highest metric value. The node
with highest residual energy and number of connections is assigned to have a higher

[email protected]
 Table 1 Analysis of literature survey of the above-mentioned protocols
94

S. No. Routing protocol Domain/class/feature Advantage Disadvantage/drawback Application

1 BEENISH, Heterogeneous wireless Better stability, Four levels of energy level Military, traffic
MBEENISH, sensor network, increased network induce more complexity; transportation,
iBEENISH and cluster-based routing lifetime, more number of ultra-super, super and environmental
iMBEENISH protocol, multi-level energy packets/messages advanced nodes are more monitoring, mobile
based transmitted to base tortured communication,
station etc.
2 EERP Multi-hop scheme Energy-efficient, Minimum hop count to find Micro-electro-
reduced packet optimal shortest path needs mechanical
retransmission, load high link quality systems (MEMS),
balanced wireless
communications,
target tracking,
environmental
monitoring and
battlefield
applications

[email protected]
3 EE-LEACH Effective data ensemble and Energy-efficient data Lack of data security and Military, object
optimal clustering gathering, reduced integrity tracking, habitat
energy consumption, monitoring
increased network
lifetime
4 Cluster-based Wireless body area network, Better network Multiple path selection Body area
ACO BAN optimal path selection, a connectivity, better algorithms are to be run for network, health
meta-heuristic search throughput, robustness backup, acknowledgement monitoring
algorithm combines ANT segment with every packet
colony optimization (ACO) caused network overhead
using clustering
(continued)
I. Banerjee and P. Madhumathy
 Table 1 (continued)
S. No. Routing protocol Domain/class/feature Advantage Disadvantage/drawback Application
5 PDORP Energy optimization using Less buffering, delay is Sink mobility is not Battlefield
hybridized algorithm, cache reduced, less energy considered, initial creation surveillance,
and directional transmission utilization and better and updation of trust list at habitat monitoring
concept throughput, consistent each round puts overhead to and underwater
QoS and more lifetime the network monitoring
of network
6 E2HRC A heterogeneous Reduced energy The packet delivery ratio IPv6 routing
ring-domain communication consumption, number of gradually decreases as protocol for low
topology, clustering control messages, number of nodes increases in power and lossy
algorithm, event-driven reduced signal wireless sensor network networks
cluster head rotation attenuation for
mechanism long-distance
communication and
decreased packet loss
Study on Energy-Efficient and Lifetime-Enhanced…

ratio
7 LCP Distributed and dynamic Energy-efficient, better Compromised end-to-end Extreme

[email protected]
clustering protocol, load balancing, delay, packet delivery is environment like
inter-cluster approach, increased network done hop by hop dense forest,
cluster head rotation lifetime earthquake zone,
approach volcanic area, etc.
8 REM Cluster-based hierarchical Reliable, less delay, Delay variation occurs with Firefighter
approach, metric-based CH energy balancing, better increased numbers of communication
selection and CH rotation packet delivery ratio firefighters network (FCN)
9 PECE Cluster formation, stable Reduced cost of When the number of clusters Industry, military,
data transfer, BCO communication, is more than 6, the environmental
energy-efficient, load communication between monitoring and
balancing, increased node and cluster both medical field
network lifetime increase in number
95
96 I. Banerjee and P. Madhumathy

metric value. When the cluster head forwards data to the base station, it also takes
account of minimum hop to the base station along with residual energy and number
of connections of the nodes. This protocol offers low latency, more reliability, good
energy balancing and longer network lifetime.

3.9 PECE

The routing protocol named predictive energy consumption efficiency (PECE) for
WSN consists of the steps: cluster formation and stable data transfer. First, cluster
formation is done based on forming an energy-saving routing algorithm which takes
into account the node degree, residual energy of the nodes and distance of the route,
i.e. path between the nodes. Cluster head is selected according to the node degree and
node relative distance to provide better coverage and to minimize communication
cost within the cluster. These formed clusters determine the number of CHs and
each cluster size, and also reduce path cost between member nodes and CH within
the same cluster, increasing the overall cluster performance. The stable data transfer
stage in PECE data transmission is designed with bee colony optimization (BCO).
Here, two different bee agents are used to predict the optimized route according to
the energy consumption, neighbour hops and delay in the network. This procedure
improves the energy balancing of the network (Table 1).

4 Conclusion

Growing demand in the field of WSN accelerates the research scope in this field. There
is hardly any domain where communication is not addressed through WSN. Since
sensor nodes contain limited battery life, proper energy balancing and utilization
is major concern in designing any routing protocol. Clustering routing protocols
show better energy efficiency, reliability, load balancing and network lifetime in
comparison with direct communication or diffusion-based routing protocols. The
logic behind the protocols is epitomized along with the advantages and areas that
could be improvised. We have also mentioned the possible application domains of
the protocols discussed.

[email protected]
Study on Energy-Efficient and Lifetime-Enhanced… 97

References

1. Ding, X., Yang, H., & Sun, S. (2014). A design of ginseng planting environment monitoring.
Sensors and Transducers, 166(3), 80–83.
2. Geetha, S., & Gouthami, S. (2017). Internet of things enabled real time water quality monitoring
system. Springer open.
3. Hughes, D., Ueyama, J., Mendiondo, E., Matthys, N., Horré, W., Michiels, S., et al. (2011). A
middleware platform to support river monitoring using wireless sensor networks. Journal of
the Brazilian Computer Society, 2011(17), 85–102.
4. Ge, Y., Wang, S. & Ma, J. (2018). Optimization on TEEN routing protocol in cognitive wireless
sensor network. EURASIP Journal on Wireless Communications and Networking.
5. Li, M., & Liu, Y. (2009). Underground coal mine monitoring with wireless sensor networks.
ACM Transactions on Sensor Networks.
6. Hao, J. (2015). Image processing and transmission scheme based on generalized Gaussian
mixture with opportunistic networking for wireless sensor networks. EURASIP Journal on
Wireless Communications and Networking.
7. Depari, A., Ferrari, P., Flammini, A., Rinaldi, S., Rizzi, M., & Sisinni, E. (2014). Development
and evaluation of a WSN for real-time structural health monitoring and testing. Eurosensors,
Procedia Engineering, 87, 680–683.
8. Li, M., Li, Z., & Vasilakos, V. (2013). A survey on topology control in wireless sensor networks:
taxonomy, comparative study, and open issues. Proceedings of the IEEE, 101(12), 2538–2557.
9. Sohrabi, K. (2000). Protocols for self-organization of a wireless sensor network. IEEE Personal
Communications, 7(5), 16–27.
10. Liu, F.A., Zhang, C.H., & Wu, N. (2014) Adaptive hierarchical routing algorithm for WSN
based on energy and distance. Appl. Res. Comput., 31(11), 3434--3437.
11. Kaur, P., & Kad, S. (2017) Energy-efficient routing protocols for wireless sensor network: a
review. International Journal of Scientific & Technolgy Research, 6(12), ISSN 2277–8616 92.
12. Yi, S. (2007). PEACH: Power-efficient and adaptive clustering hierarchy protocol for wireless
sensor networks. Computer Communications, 30, 2842–2852.
13. Heinzelman, W. R., Chandrakasan, A., & Balakrishnan, H. (2000) Energy-efficient commu-
nication protocol for wireless microsensor networks. In System sciences, Proceedings of the
33rd Annual Hawaii International Conference on, 2(4–7), 10, 2000.
14. Lindsey, S., & Raghavendra, C. S. (2002). PEGASIS: Power-efficient gathering in sensor
information systems. In Proceedings IEEE Aerospace Conference (pp. 3–1125–3–1130). MT,
USA: Big Sky.
15. Youni, O., & Sonia, F. (2004). HEED: A hybrid, energy-efficient, distributed clustering
approach for ad hoc sensor networks. Mobile Comput IEEE Trans., 3(4), 366–379.
16. Qing, L., Zhu, Q., & Wang, M. (2006). Design of a distributed energy-efficient clustering
algorithm for heterogeneous wireless sensor network. ELSEVIER Comput. Commun., 29,
2230–2237.
17. Elbhiri, B., Saadan, R., El-Fkihi, S., & Aboutajdine, D. (2010). Developed distributed energy-
efficient clustering (DDEEC) for heterogeneous wireless sensor networks. In 5th International
Symposium I/V Communications and Mobile Network (ISVC) (pp. 1–4).
18. Saini, P., & Sharma, A.K. (2010). E-DEEC—enhanced distributed energy efficient clustering
scheme for heterogeneous WSN. In Parallel Distributed and Grid Computing (PDGC), 1st
International Conference on (Iss. 28–30., pp. 205–210).
19. Akba, M., Javaid, N., Imran, Md., Amjad, N., Khan, M. I., & Guizani, M. (2016). Sink mobility
aware energy-efficient network integrated super heterogeneous protocol for WSNs. EURASIP
Journal on Wireless Communications and Networking.
20. Ghaffari, A. (2014). An energy efficient routing protocol for wireless sensor networks using
a-star algorithm. Journal of Applied Research and Technology, 12, 815–822. Yeung, K. Y., et al.
(2001, April 1). Validating clustering for gene expression data. Bioinformatics, 17, 309–318

[email protected]
98 I. Banerjee and P. Madhumathy

21. Yao, J., Lin, C., Xie, X., Wang, A., & Hung, C.-C. (2010). Path planning for virtual human
motion using improved A* star algorithm. In: Information Technology: New Generations
(ITNG), Seventh International Conference (pp. 1154–1158).
22. Arumugam, G. S., & Ponnuchamy, T. (2015) EE-LEACH: Development of energy-efficient
LEACH Protocol for data gathering in WSN. EURASIP Journal on Wireless Communications
and Networking.
23. Srinivas, M. B. (2016). Cluster based energy efficient routing protocol using ant colony opti-
mization and breadth first search. Procedia Computer Science, 89, pp. 124–133.
24. Brar, G. S., Rani, S., Chopra, V., Malhotra, R., Song, H., & Ahmed, S. H. (2016). Energy
Efficient Direction-Based PDORP Routing Protocol for WSN, Special Section on Green Com-
munications and Networking for 5G Wireless. Digital Object Identifier, 4, 3182–3194.
25. Johnson, D., Hu, Y. & Maltz, D. (2007). The dynamic source routing protocol (DSR) for mobile
Ad Hoc networks for IPv4. IETF, document RFC 4728.
26. Chen, H., Zhu, Y., & Hu, K. (2009) Cooperative bacterial foraging algorithm for global
optimization. In Proceeding IEEE Chinese Control and Decision Conference, (vol. 6,
pp. 3896–3901). Guilin, China.
27. Kim, D. H., Abraham, A., & Cho, J. H. (2007). A hybrid genetic algorithm and bacterial
foraging approach for global optimization. Information Sciences, 177(18), 3918–3937.
28. Zhang, W., Li, L., Han, G., & Zhang, L. (2017). E2HRC: An energy-efficient heterogeneous ring
clustering routing protocol for wireless sensor networks. Special Section on Future Networks:
Architectures, Protocols, and Applications, Digital Object Identifier, 5, 1702–1713.
29. Eshaftr, M., Al-Dubai, A. Y., Romdhani, I., & Yassien, M. B. (2015). A new energy efficient
cluster based protocol for wireless sensor networks. Proceedings of the Federated Conference
on Computer Science and Information Systems, 5, 1209–1214.
30. Atiq, M. K., Manzoor, K., Hasan, N. ul., & Kim, H. S. (2014). A reliable energy-balancing multi-
group (REM) routing protocol for firefighter communication networks. EURASIP Journal on
Wireless Communications and networking.
31. Zhang, D., Wang, X., Song, X., Zhang, T., & Zhu, Y. (2015). A new clustering routing method
based on PECE for WSN. EURASIP Journal on Wireless Communications and Networking.

[email protected]
Role of Fog Computing in IoT-Based
Applications

Charul Thareja and N. P. Singh

Abstract Internet of Things (IoT) as the name suggests is an interconnection of

our daily usage items like smartphones, cars, laptops, tube lights, electric fans, etc.
With the inception of IoT, each and every electronic item in this world would have
an IP address. It would transform our way of interaction, making our life more
autonomous. The essential requirement for the upcoming of this technology includes
real-time response, low latency, fast processing and computation capabilities. These
features being difficult to be handled by cloud data centres because of their remote
location leads to development of a new technology called fog computing, which
is an extension of cloud computing services from cloud data centres towards the
edge devices. Fog computing possesses some security and privacy threats which are
discussed in this paper. There is a need to resolve these threats so that people use
fog-based IoT applications without any reluctance. Some of the solutions proposed
are also mentioned in this paper.

Keywords IoT · Fog computing · Cloud computing · Applications · Challenges

1 Introduction

Internet of Things (IoT) is the interconnection of billions of devices that can interact
and share resources with the help of embedded software, sensors and other electronic
equipment [1]. The IoT possess applications in almost all domains like military for
battlefield surveillance [2], health care for patient’s health monitoring [3], environ-
ment for analysing moisture content in soil [4], etc. It is inferred that IoT is going
to make our life autonomous, which can be illustrated from one of the most com-
mon applications of IoT, i.e. Smart home application [5], enabling lighting system,
room temperature, entertainment system and other electronic devices to be controlled
automatically, with the help of sensors.

C. Thareja (B) · N. P. Singh

Department of ECE, NIT Kurukshetra, Kurukshetra 136119, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 99

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_9

[email protected]
100 C. Thareja and N. P. Singh

The above-mentioned IoT applications collect a huge amount of data, which needs
to be processed or analysed. Presently, with 23.14 billion Internet-connected devices,
the cloud data centres can handle the amount of data gathered by them, but with the
rapid growth of Internet-connected devices, which is estimated to increase up to 500
billion by 2025, the data to be processed will also increase with the approximation
of about 500 zettabytes by 2019, which is beyond the handling limit of cloud data
centres [2]. So, the need of the hour is to have a scenario in which the data is analysed
in real time and only some of the data which require huge computations, or which
are necessary for storage, need to be sent towards the cloud DCs.
IoT applications possess different characteristics, like data storage, huge com-
putations and real-time responses. For data storage and huge computational task,
cloud computing is a good solution. But for real-time response, it becomes difficult
for cloud data centres to handle IoT applications, because cloud DCs are remotely
located and leads to poor QoS. So, the need is to bring these cloud services from
far-away places to near the user devices, which is done through fog computing.
Thus, fog computing can be defined as an extension of the cloud computing
for the real-time, latency-sensitive applications, which consists of a combination
of heterogeneous, decentralized devices, communicating and cooperating among
themselves to perform data storage and various data computation tasks [6].
The remainder of the survey is organized as follows. Section 2 represents the
features of fog computing along with its comparison with cloud computing. The
architecture of fog computing has been reviewed in Sect. 3. Role of fog computing
in different IoT-based applications has been discussed in Sect. 4. Sections 5 and 6
include the various security and privacy threats of this system and measures to secure
it, respectively. Finally, the survey is concluded in Sect. 7.

2 Features of Fog Computing

Many of the features of fog computing are similar to that of the cloud computing but
some are different. Both have been described briefly. The similar features include
the following:
• Mobility Support [7]: Mobility is the ability to move. Through fog computing,
IoT devices acquire this ability without being disconnected from the fog nodes.
Fog services are ubiquitous in nature. Hence, mobility support offered by cloud
data centres is limited as compared to the support provided by fog nodes.
• Heterogeneity [8]: The data is collected from Internet-connected devices that can
be either mobile IoT devices including wearable devices (fitness tracker, smart
cameras, etc.) and mobile smart devices (vehicles, smartwatches, etc.) or fixed
IoT devices including sensors and RFID tags deployed on specific areas of prod-
uct. Collected data is then transferred to the fog nodes (basically routers, gateways
and switches) which are deployed in diverse environment like on the roof of build-
ings, in smartphones, in traffic lights and so on, making the whole architecture

[email protected]
Role of Fog Computing in IoT-Based Applications 101

heterogeneous. Same heterogeneity is seen in cloud computing but with the dif-
ference that data instead of being collected by fog nodes is collected by cloud data
centres which are remotely located.
• Online analytics and interplay with cloud [9]: Fog nodes acts as an intermediate
between the IoT devices from which the data is being collected and the clouds
where the data is to be sent for storage. This feature of fog computing can be
explained with the help of an example of smart e-healthcare centre. The data from
the sensors deployed on the patient’s body is being collected continuously by fog
nodes, which is being analysed in real time to detect emergent event. If everything
seems to be normal, and then the data is sent to the cloud for storage. Otherwise,
some necessary steps are taken to help patients like ringing alarms and sending
alert messages to the doctor. Until the doctor arrives, sensors on the patient’s body
take active actions to give them intensive care. It can be said fog layer is an interface
between the device layer and the cloud layer. Similar analysis is performed by the
cloud data centres but with some delay and with only two layers (Device layer and
cloud layer).
Some of the distinguishing features from the cloud computing are enlisted below:
• Location Awareness [10]: The location of the fog nodes can be tracked easily to
provide real-time services to the user devices. Based on the knowledge of location
of fog nodes, device region can also be identified easily.
• Geographic Distribution [10]: The fog nodes that are geographically distributed
may be on the roof of the building, on smartphones, on top of the vehicle and at
the point of interest so that it is able to provide services to IoT devices, i.e. collect
high data stream from IoT devices for processing.
• Low Latency [9]: Fog nodes being deployed nearer to the user devices and pos-
sessing the necessary capability of storing and computing data are able to provide
services with much less time as compared to that taken by the cloud.
• Large-Scale IoT Application support [8]: Large-scale IoT application includes
large number of sensors, and thus a large number of data for computation and
processing to be handled in real time, which is a cumbersome task for cloud
computing but can be easily handled by fog nodes which are ubiquitous and has
the expertise to manage huge number of IoT devices.
• Decentralization [8]: Fog computing is a decentralized network. No central server
is present to provide services and resources. Fog nodes arrange itself automatically
to provide real-time services to the IoT devices.
Table 1 shows the comparison of different features of cloud and fog.

3 Architecture

The fog computing architecture is illustrated in Fig. 1. It is a three-tier architecture.

Tier 1: The bottom tier known as the device layer consists of the different smart
devices including mobile IoT devices, i.e. the devices carried by their users like fitness

[email protected]
102 C. Thareja and N. P. Singh

Table 1 Comparison of different features of cloud or fog [2]

Features Cloud Fog
Latency High Low
Mobility N/A Yes
Architecture Centralization Decentralization
Service access Through core At the edge/on handheld device
Location awareness N/A Yes
Geographic distribution N/A Yes
Scalability Average High
Availability High High
Number of users or devices Tens or hundreds of millions Tens of billions
Content generation Central location Anywhere
Content consumption End devices Anywhere
Software virtualization Central corporate facilities User devices/network
equipment

Fig. 1 Fog computing architecture

[email protected]
Role of Fog Computing in IoT-Based Applications 103

trackers, bracelets and smart devices like smartphones, vehicles, smartwatches, etc.
and the fixed IoT devices including RFID and sensor nodes which are attached on
the items to perform some specified task. These heterogeneous devices transmit data
to the upper tier.
Tier 2: The second tier known as the fog computing layer consists of intelligent
devices called fog nodes which are routers, gateways, etc. that possess the ability of
data storage, computation, routing and packet forwarding.
Tier 3: The uppermost layer is known as the cloud computing layer, which consists
of data centres and big servers. The data which requires large historical data for
processing or huge computational tasks is sent to the cloud layer.
Different smart devices also known as the terminal nodes join together to form
virtual clusters (VCs). Different VCs together form an edge virtual private network
(EVPN). The data from the VCs is sent to the fog instances (FIs) via edge gateways for
computing, storage, etc. The fog layer is divided into two parts, one is fog abstraction
layer, which manages the fog resources, enables virtualization and provides user’s
privacy. Another part is fog orchestration layer, consisting of a software agent called
foglet, which checks the state of device. Within the fog instances, data is processed
and analysed, whether it requires real-time service, temporary storage or permanent
storage. For permanent storage or long-term storage, data is sent to the data centres
or cloud layer but for short-term evaluation, data is processed in fog layer only. Thus,
the fog computing architecture allows the efficient utilization of cloud Layer [11].

4 Applications

Fog nodes play four important roles, namely, real-time service for fog, transient
storage for fog, data dissemination for fog and decentralized computation for fog.
Based on these roles, some of the important applications have been discussed below.

4.1 Real-Time Services for Fog

Cloud data centres being far away from the IoT devices takes a lot of time in analysing
and computing the results of data, because first the data is sent to the cloud, then in
DCs it takes some processing time and after that the analysed result comes back to
the device. So, three different times are being added here: time taken by data to reach
the cloud (T c ), processing time (T p ) and time taken by the analysed data to reach
back to the device (T r ) [12]. So, the expression for total time (T t ) can be written as

Tt  Tc + T p + Tr

[email protected]
104 C. Thareja and N. P. Singh

But in case of fog computing, fog nodes being nearer to the fog devices, the data
after the analyses reaches within microseconds to the required destination. Though
all the three times which are added in cloud computing scenario are added in fog
computing also, the time taken by data to reach to fog nodes is much less as compared
to that of the cloud. Thus, for real-time applications like in Battlefield Surveillance, it
is really important for us to get the information regarding the entry of an intruder into
our land, so that some necessary steps can be taken to avoid any kind of uncertainty
in our country. But, if we depend upon cloud data centres for this information, then
the security personnel will get the data when the intruder has already entered the
land. So, for latency-sensitive applications, fog nodes act as a private cloud for the
IoT devices.
Some applications of fog nodes which offer real-time control and fast decision-
making capability are smart traffic light, decentralized vehicular navigation, home
energy management, e-healthcare services, etc.
• Smart Traffic Light [13]: In this application, the fog nodes are deployed on the
traffic light pole and takes necessary actions according to the data collected by the
sensors, which are installed on the road, on traffic signal pole and many more places
for covering the whole view of the Carrefour. The light sensors when detecting the
flashy lights of ambulance and police vans send the data to fog nodes, which then
take urgent steps to let these emergency vehicles go first. Some sensors send the
details of the speed of pedestrians and the bicyclers crossing the road towards the
fog nodes, and accordingly the traffic signals are arranged by giving them enough
time to cross the roads. Not only this but fog nodes also try to avoid collisions by
sending warning messages to the vehicles which are approaching with high speed.
The sensors deployed on roads continuously gather data regarding the speed of
vehicles or send it to fog nodes for further analyses. Thus, this application, with
the help of fog computing, is going to reduce the number of accidents.
• Healthcare and Activity Tracking [14]: The scenario can be understood in three
layers: in the first layer, all the sensors are included which are deployed on the body
of the patient for collecting data from all the ICUs. In the second layer, fog node,
i.e. Private cloud of hospital, is present to deliver fast responses after analysing
the data provided by sensors like sending messages to doctor or starting alarms
during emergency situation and giving necessary instructions to the sensors to
take preventive actions until the doctor arrives. The third layer consists of a cloud,
which is just for storing all the details of the patient. If there is a need to check
medical history of patients, then this layer comes in use.

4.2 Transient Storage for Fog

When the data is to be stored temporarily, i.e. for about 1–2 h, or if the data is to be
accessed frequently then instead of storing it in clouds (because storage space and
bandwidth would be wasted), it can be stored in fog nodes and when the necessary

[email protected]
Role of Fog Computing in IoT-Based Applications 105

updates have been done on the data then it can be either discarded or sent to the clouds
for permanent storage. When some files which need to be accessed continuously for
real-time application are stored in cloud, then delay is added to the response.
This role of fog node is useful in various applications including content caching,
shopping cart management, software and credential updating, etc. Some of them are
discussed below.
• Content Caching [8]: The data which is accessed by the user for the first time
comes from the cloud and gets stored into the fog node. The second time when
some other user wants to use the same information then there is no need to go to the
tier 3 again as the data has already been cached from cloud to the fog node. Hence,
the second user gets faster response as compared to the first one. Its working is
same as that of the cache memory in computers. This application can be used in
interstate buses by employing a fog node on the top of the bus providing free Wi-Fi
as a social service, and then all the passengers can access the data using that fog
node.
• Shopping Cart Management [8]: Traditionally, while e-shopping, the data of
e-cart needs to get updated in the cloud, which used to add delay to the services
of the customer. But with the advent of fog computing, the cart information gets
cached into the fog nodes, where the necessary updates are done by the customer
after that when the customer logs out his/her account, then the data is sent back to
the cloud for storage. This procedure reduces the amount of delay in services and
even increases customer’s satisfaction.

4.3 Data Dissemination for Fog

Being the middle tier in the fog computing architecture, fog nodes act like an interface
between the IoT devices and the cloud data centres. The fog nodes either collect data
from IoT devices and send it to cloud or collect data from cloud data centres and
send it to IoT devices. Fog node acts as a router, helps in packet forwarding, can
work as a data aggregator, and can also perform simple processing for choosing the
proper audiences to whom the information should be sent.
The applications including data dissemination for fog includes energy consump-
tion collection, local content distribution and fog-based malware defence. Some of
them are discussed below:
• Local Content Distribution [15]: Local fog nodes provide useful information to
the users passing by, e.g. users are given updates about the traffic, hotels, restau-
rants, gas stations, while they are travelling and if simple processing is done, then
the messages of nearby hotels, restaurants, etc. would be sent to a person who is
sensualist and about amusement parks to a person who is adventurous, etc. This
application can also be used in parking navigation service [16], by suggesting the
drivers the right place to park, with the help of various sensors and cameras being
deployed in the parking area.

[email protected]
106 C. Thareja and N. P. Singh

• Fog-based Malware Defence [8]: If malware defence hardware and software are
installed in each of the devices, then it would waste a lot of storage space and would
require timely processing (more battery backup). So to avoid it, if the malware
defence system is installed on the fog nodes, then it would act as shared resource
for all the IoT devices to detect the infected files on compromised devices, and
then clean them.

4.4 Decentralized Computation for Fog

Fog nodes provide decentralized computation services, which aid both the ends, i.e.
user devices and cloud centres from heavy computations. Like in mobile phones, all
the computations and processing cannot be done because it will consume battery, so
to avoid it, computational task is transferred to the fog nodes, which sends analysed
data back to device, after processing. Likewise, for every small computation, earlier
the mobile devices used to offload data to cloud DCs, adding latency to the process,
wasting the bandwidth and even wasting the computational resources, which is now
performed by fog nodes.
The applications supporting the above feature are computation offloading, aided
computation and big data analytics. Some of them are discussed below.
• Computation Offloading [17]: Though clouds possess huge computational capa-
bility, it consumes a lot of energy and adds latency in providing services to the
devices. Fog computing brings these computational services towards the user
devices. The IoT devices can fully exploit the computing resources of the fog
nodes even for simple processing and it also reduces the computational burden on
clouds.
• Big Data Analytics [18]: As the number of IoT devices is increasing, the amount
of data collected by these devices is also increasing. Analysing huge amount of
data at cloud DCs would be difficult; instead, the task can be distributed among
different fog nodes. Like, when finding a missing person, instead of analysing all
video recordings collected at cloud DCs, the task is divided into each fog node to
check the records. Thus, performing local analytics at the fog nodes makes process
fast and reduces delay.

5 Threats of Fog Computing

The threats which have restricted the growth of cloud computing and fog computing
are security and privacy threat. These are discussed below.

[email protected]
Role of Fog Computing in IoT-Based Applications 107

5.1 Security Threat

Fog computing is more secured than cloud computing because of two main reasons.
First, the data is transferred locally between the data sources and the fog nodes,
and the dependency on Internet connection is reduced. Since the data is being stored,
analysed and transferred locally, it becomes difficult for hackers to access user’s data.
Second, the information exchange between the cloud and the smart devices does not
take place in real time which does not let the tracker find critical information of the
user.
Fog computing being more secure than cloud computing cannot be said to be fully
secured; as the user’s data reaches to the employees employed at fog nodes, they may
misuse their data and this even leads to privacy leakage. Some of the security threats
are as follows:
• Forgery and Spam: Some attackers may make their fake profiles and may even add
fake information about them on it, to mislead other users. This also causes wastage
of storage, bandwidth and energy. Spam refers to the redundant or unwanted infor-
mation, which unnecessarily wastes the storage resources.
• Tampering: Some attackers try to interfere with the data sent by the user. They
can either alter the data or drop the data; the purpose is that the real data should
not reach the destination. Being wireless communication, it becomes difficult to
detect whether this drop occurred due to transmission failure or by some intruder.
• Jamming: Some attackers generate a huge amount of redundant or fake messages,
so that no other user can utilize the computation and routing services of fog. The
purpose is to jam the communication channels. It can also be said as Denial-of-
Service, when some legitimate user is not able to use the services of fog because
it is being flooded with fraudulent data. It is common in fog because of its limited
resources.
• Eavesdropping: Some attackers continuously listen to the communication chan-
nel, to get the important details. This type of attack is common when data encryp-
tion techniques are not used.
• Collusion: Two or more parties can combine together to fraud a legal user. The
two parties combined can be IoT devices, fog and IoT device or cloud and fog,
etc. This is going to double the attack capability.
• Man-in-the-Middle: There may be an attacker sitting in between the two parties,
listening to their conversations, who is secretly modifying the data being exchanged
between them and the two parties do not even get to know it.
• Impersonation: There can be a situation in which an illegal user can act as a
legitimate user to utilize the services provided by the fog nodes or it may happen
a fake fog is providing services to the legal users and maliciously utilizing the
information provided by them.

[email protected]
108 C. Thareja and N. P. Singh

5.2 Privacy Threat

It is also an important issue as the user’s data is being processed, shared and collected,
and no user wants his/her data to be leaked. There are four kinds of privacy threats
which includes identity, data, location and usage privacy.
• Identity Privacy: User’s identity includes his/her name, address, Aadhar card num-
ber and some more personal details which are easily disclosed during checking
the authenticity of a person while they are availing for the services of the fog.
• Data Privacy: When users utilize the fog services, some of the information about
user like its occupation, health status, preferences are exposed to some third party
which maintains the user’s data at the fog node.
• Usage Privacy: Usage pattern tells when the users utilize the fog node services,
which gives the information about the living pattern of the user including, when
they sleep, when they wake, at what time they go to office and many more details,
which are really dangerous if it gets into wrong hands.
• Location Privacy: Location privacy is something which every user has to sacrifice
in order to enjoy different online applications. If a user utilizes some fog node
services, his/her location can easily be accessed by finding the fog node to which
it is connected and an attacker can even detect the trajectory of a user.

6 Fog Security

The threats mentioned above leads to the unwillingness of users to utilize fog-based
IoT applications. So, this generates a need to secure fog computing. In this section,
the security and privacy challenges have been further discussed and the measure
which should be taken to avoid these challenges is also discussed. It is divided into
four parts similar to the applications studied above.

6.1 Challenges and Solutions on Real-Time Services

Fog computing provides various fog-assisted IoT real-time applications, which due
to security and privacy threats has not been utilized fully. Some of the challenges
faced in these applications with the solutions to them are enlisted in Table 2.

6.2 Challenges and Solutions on Transient Storage

Transient storage capability provided by fog allows users to maintain their data
provisionally. The users will be using these services only if they are sure that their

[email protected]
Role of Fog Computing in IoT-Based Applications 109

Table 2 Security challenges and solutions in real-time services application of fog computing
Role Security challenges Security solutions
Real-time services Identity Identity authentication [19]
authentication Cooperative authentication [20]
Anonymous authentication [21]
Access control Role-based access control policy [22]
Attribute-based access control policy [23]

Table 3 Security challenges and solutions in transient storage role of fog computing
Role Security challenges Security solutions
Transient storage Sensitive data Identification and Symmetric encryption [24]
protection Asymmetric encryption [25]
Secure data sharing Proxy re-encryption [26]
Attribute-based encryption [27]
Key-aggregate encryption [28]

Table 4 Security challenges and solutions in data dissemination role of fog computing
Role Security challenges Security solutions
Data dissemination Secure data search Symmetric searchable encryption [29]
Asymmetric searchable encryption [30]
Secure content Secure service discovery [31]
distribution Broadcast encryption [32]
Key management mechanism [33]
Anonymous broadcast encryption [34]

data is safe, and there would be no privacy leakage. The challenges faced in transient
storage and the solutions which can be used to provide security are enlisted in Table 3.

6.3 Challenges and Solutions of Data Dissemination

Fog nodes act as an intermediate between the IoT devices and the clouds. So, it
requires the information which is being passed between the two ends that should be
accurate. The challenges which are faced and the measures which can be taken for
efficient data transmission are listed in Table 4.

6.4 Challenges and Solutions of Decentralized Computation

Fog nodes instead of storage capabilities also possess computational capabilities.

These computations can be hacked by an attacker, and he/she may be controlling

[email protected]
110 C. Thareja and N. P. Singh

Table 5 Security challenges and solutions in decentralized computation role of fog computing
Role Security challenges Security solutions
Decentralized computation Verifiable computation Privately verifiable computation
[35]
Publicly verifiable computation
[36]
Secure-aided computation Server-aided verification [37]
Server-aided encryption [38]
Server-aided key exchange [39]

them and misguiding user. Not only this but there are some more problems which
can be faced by user and various solutions to resolve these challenges that are listed
in Table 5.

7 Conclusion

Fog computing is not a replacement of cloud computing, but it complements it. Both
the technologies when used together form a new breed of technology that serves
various IoT applications which may deal with computing, or storing temporary data,
or acting as an interface between two layers, or serving real-time applications. This
new technology in spite of making our life autonomous is surrounded by some
security and privacy threats which need to be resolved to enjoy fog computing.
Some of the techniques to resolve the challenges have been mentioned, but this is an
open research issue and needs more inputs.

References

1. Gubbi, J., Buyya, R., Marusic, S., & Palaniswami, M. (2013). Internet of Things (IoT): A
vision, architectural elements, and future directions. Future Generation Computer Systems,
29(7), 1645–1660.
2. Farooq, M. J., & Zhu, Q. (2017). Secure and reconfigurable network design for critical informa-
tion dissemination in the Internet of battlefield things (IoBT). In 2017 15th Int. Symp. Model.
Optim. Mobile, Ad Hoc, Wirel. Networks, WiOpt 2017 (vol. 17, no. 4, pp. 2618–2632).
3. Verma, P., & Sood, S. K. (2018). Fog assisted-IoT enabled patient health monitoring in smart
homes fog assisted-IoT enabled patient health monitoring in smart homes (vol. 5, no. February,
pp. 1789–1796).
4. Tarushi Wasson, P. K., Choudhury, T., & Sharma, S. (2017) Integration of RFID and sen-
sor in agriculture using IOT—IEEE Conference Publication. In Smart Technol. Smart Nation
(SmartTechCon), International Conference (pp. 217–222).
5. Lee, Y. T., Hsiao, W. H., Lin, Y. S., & Chou, S. C. T. (2017). Privacy-preserving data analyt-
ics in cloud-based smart home with community hierarchy. IEEE Transactions on Consumer
Electronics, 63(2), 200–207.

[email protected]
Role of Fog Computing in IoT-Based Applications 111

6. Vaquero, L. M., & Rodero-Merino, L. (2014). Finding your way in the fog: Towards a com-
prehensive definition of fog computing. ACM SIGCOMM Computer Communication Review,
44(5), 27–32.
7. Munir, A., Kansakar, P., & Khan, S. U. (2017). IFCIoT: Integrated Fog Cloud IoT: A novel
architectural paradigm for the future Internet of Things. IEEE Consumer Electronics Magazine,
6(3), 74–82.
8. Ni, J., Zhang, K., Lin, X., & Shen, X. (2017). Securing fog computing for internet of things
applications: Challenges and solutions. IEEE Commun. Surv. Tutorials, (no. c, pp. 1–1).
9. Bonomi, F., Milito, R., Zhu, J., & Addepalli, S. (2012). Fog computing and its role in the
internet of things. In Proceedings of the first edition of the MCC workshop on Mobile cloud
computing—MCC’12, (p. 13).
10. Shropshire, J. (2014). Extending the cloud with fog : Security challenges & opportunities.
Americas Conference on Information Systems pp. 1–10.
11. Sarkar, S., Chatterjee, S., & Misra, S. (2015). Assessment of the suitability of fog computing
in the context of internet of things. IEEE Transactions on Cloud Computing, 7161(c), pp. 1–1.
12. Liu, L., Chang, Z., Guo, X., Mao, S., & Ristaniemi, T. (2017). Multi-objective optimization
for computation offloading in fog computing. IEEE Internet of Things Journal, 4662(c), 1–12.
13. Liu, J., et al. (2018). Secure intelligent traffic light control using fog computing. Future Gen-
eration Computer Systems, 78, 817–824.
14. Akrivopoulos, O., Chatzigiannakis, I., Tselios, C., & Antoniou, A. (2017). On the deployment of
healthcare applications over fog computing infrastructure. In 2017 IEEE 41st Annual Computer
Software and Applications Conference, pp. 288–293.
15. Oteafy, S. M. A., & Hassanein, H. S. (2018). IoT in the fog: A roadmap for data-centric IoT
development. IEEE Communications Magazine, 56(3), 157–163.
16. Lu, R., Lin, X., Zhu, H., & Shen, X. (2010). An intelligent secure and privacy-preserving parking
scheme through vehicular communications. IEEE Transactions on Vehicular Technology, 59(6),
2772–2785.
17. Wang, C., Liang, C., Yu, F. R., Chen, Q., & Tang, L. (2017). Computation offloading and
resource allocation in wireless cellular networks with mobile edge computing. IEEE Transac-
tions on Wireless Communications, 16(8), 4924–4938.
18. Shi, W., Cao, J., Zhang, Q., Li, Y., & Xu, L. (2016). Edge computing: Vision and challenges.
IEEE Internet Things Journal, 3(5), 637–646.
19. Chandrasekhar, S., & Singhal, M. (2015). Efficient and scalable query authentication for cloud-
based storage systems with multiple data sources. IEEE Transactions on Services Computing,
1374(c), 1–1.
20. Lin, X., & Li, X. (2013). Achieving efficient cooperative message authentication in vehicular
ad hoc networks. IEEE Transactions on Vehicular Technology, 62(7), 3339–3348.
21. Lu, R., Lin, X., Luan, T. H., Liang, X., Member, S., & Shen, X. S. (2012). Pseudonym changing
at social spot: An effective strategy for location privacy in VANETs. IEEE Transactions on
Vehicular Technology, 61(1), 1–11.
22. Salonikias, S., Mavridis, I., & Gritzalis, D. (2016). Critical information infrastructures security
(vol. 9578, pp. 15–26).
23. Lewko, A., & Waters, B. (2011). Decentralizing attribute-based encryption. In Lecture Notes
in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture
Notes in Bioinformatics) (vol. 6632 LNCS, pp. 568–588).
24. Agrawal, M., & Mishra, P. (2012). A comparative survey on symmetric key encryption tech-
niques. International Journal of Computer Science and Engineering, 4(5), 877–882.
25. Tarakdjian, G. (1993) A Survey (pp. 74–81).
26. Ateniese, G., Fu, K., Green, M., & Hohenberger, S. (2006). Improved proxy re-encryption
schemes with applications to secure distributed storage. ACM Transactions on Information
and System Security, 9(1), 1–30.
27. Alrawais, A., Alhothaily, A., Hu, C., Xing, X., & Cheng, X. (2017). An attribute-based encryp-
tion scheme to secure fog communications. IEEE Access, 5, 9131–9138.

[email protected]
112 C. Thareja and N. P. Singh

28. Chu, C., Chow, S. S. M., Tzeng, W., Zhou, J., Deng, R. H., & Member, S. (2014). Supplemen-
tary material for key-aggregate cryptosystem for scalable data sharing in cloud storage. IEEE
Transactions on Parallel and Distributed Systems, 25(2), 1–4.
29. Rizomiliotis, P., & Gritzalis, S. (2015). ORAM based forward privacy preserving dynamic
searchable symmetric encryption schemes. In Proceedings of the 2015 ACM Workshop on
Cloud Computing Security Workshop—CCSW’ 15 (pp. 65–76).
30. Hwang, Y. H., & Lee, P. J. (2007). Public key encryption with conjunctive keyword search and
its extention to multi-user system. In Proceeding of Pairing (pp. 2–22).
31. Wu, D. J., Taly, A., Shankar, A., & Boneh, D. (2016). Privacy, discovery, and authentication
for the internet of things. In Proceedings of ESORICS (pp. 301–319).
32. Park, J. H., Kim, H. J., Sung, M. H., & Lee, D. H. (2008). Public key broadcast encryption
schemes with shorter transmissions. IEEE Transactions on Broadcasting, 54(3), 401–411.
33. Boneh, D., Gentry, C., & Waters, B. (2005). Collusion resistant broadcast encryption with short
ciphertexts and private keys. In Proceedings of CRYPTO (pp. 258–275).
34. Libert, B., Paterson, K. G., & Quaglia, E. A. (2012). Anonymous broadcast encryption:
Adaptive security and efficient constructions in the standard model. In Proceedings of PKC
(pp. 206–224).
35. Gordon, S. D., Katz, J., Liu, F.-H., Shi, E., & Zhou, H.-S. (2015). Multi-client verifiable
computation with stronger security guarantees. In Proceedings of TCC (pp. 144–168).
36. Elkhiyaoui, K., Önen, M., Azraoui, M., & Molva, R. (2016). Efficient techniques for publicly
Verifiable delegation of computation. In Proceedings of ACM CCS (pp. 19–128).
37. Wang, T., Zeng, J., Bhuiyan, M. Z. A., Tian, H., Cai, Y., Chen, Y., et al. (2017). Trajectory privacy
preservation based on a fog structure in cloud location services. IEEE Access, 5, 7692–7701.
38. Bellare, M., Keelveedhi, S., & Ristenpart, T. (2014). DupLESS: Server-aided encryption for
deduplicated storage. In Proceedings of Usenix Security (pp. 179–194).
39. Kamara, S., Mohassel, P., & Riva, B. (2012). Salus: A system for server aided secure function
evaluation. In Proceedings of ACM CCS (pp. 797–808).

[email protected]
Single Horizontal Camera-Based Object
Tracking Quadcopter Using StaGaus
Algorithm

Lakshmi Shrinivasan and N. R. Prasad

Abstract This research is a solution to the problem of a hardware platform of object

tracking drones. This platform can be used to build further advancements such as
selfie drones, follow-me drones for adventure sports and robot pets. StaGaus algo-
rithm has been derived using first principles and is compared against standard algo-
rithms like SRDCF. The algorithm is tested on, an on-board two Android phones for
real-time telemetry. This paper demonstrates that StaGaus works even on memory-
and performance-constrained devices. In order to standardise the development of
object tracking drones’ algorithms, we have built a customised ROS-based Gazebo
simulator from scratch. This simulator is capable of simulating multiple robots of
multiple types. This uses actual physics Open Dynamics Engine which has been com-
pared against the existing simulators. Finally, as a by-product, this design proved that
the cost of the proposed physical quadcopter is low.

Keywords Drone · Image processing · Android · Machine learning · Simulator

using ROS and Gazebo

1 Introduction

A quadrotor is a rotorcraft propelled by four motors typically using a separate flight

controller. One of the non-trivial research problems of machine learning-based image
processing that is being tackled with our novelty is the development of object tracking
algorithms for an on-board Android controller using a single camera. A single camera
is only 2D information unlike the 3D information that is provided by stereo-vision,
RADAR and LIDAR. We demonstrate that the accuracy of this 2D information is
sufficient enough to track the object without the need of expensive 3D sensors.
Another problem is the standardisation of results of object tracking drones’ algo-
rithms, because algorithms like StaGaus employ both image processing statistics
and statistics of the controller algorithms. It is essential to simulate both controllers

L. Shrinivasan (B) · N. R. Prasad

Department of ECE, Ramaiah Institute of Technology, Bengaluru 560054, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 113

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_10

[email protected]
114 L. Shrinivasan and N. R. Prasad

with the physics of the quadrotor while maintaining a GUI for image processing.
Instead of relying on weather conditions of the locality and relying on hardware
which may have to be re-built if they crash, the simulator we have built from scratch
using robotics operating system and Gazebo provides a single simulation interface
for both controller design and image processing. This simulator can run multiple
vehicles of multiple types, and we have made this run on a single Lenovo Y50-70
laptop having Ubuntu Operating System.

2 Problem Definition and Contributions

For both physical and software-based validations (using simulation) of single camera-
based object tracking quadcopter, we have come up with an online training StaGaus
algorithm and a custom simulator which runs in Gazebo.
Novelty is based on three aspects:

1. The proposed statistical Gaussian algorithm that has high tracking efficiency
with a single camera (unlike the additional GPS sensor, LIDAR, RADAR and
stereo-vision laden drones).
2. Performance of the algorithm is based on the performance-constrained and
battery-optimised devices (as a by-product, this is useful for cost-effective
drones).
3. The designed customised Robotics Operating System-based Gazebo simulator
that uses actual physics engine to simulate multiple vehicles of multiple types.

3 Related Work

There has been a lot of research work done on object tracking algorithms both in
spatial and in frequency domains. But, for a single camera mounted on a quadrotor, the
additional problem was with the dynamic backgrounds. In literature, a lot of generic
algorithms are available for object tracking using Android phone. Alper et al. describe
direct object detection, point-based matching and feature-based and boosting-based
machine learning tracking algorithms (like features fed to AdaBoost) [1]. Arnold et al.
describe object tracking in videos (mainly with discriminative training) [2]. Hanxuan
et al. describe feature matching and online training extensively for algorithms (with
performance optimisations like Monte Carlo sampling) [3]. Rui et al. describe an
algorithm that uses both spatial- and frequency-domain-based tracking systems [4].
But we found that the next advancement was to avoid position error that would
occur in long-term tracking. Also, there were very less or almost no literature for
an Android-controlled quadrotor design. Hence, the authors proposed new design
to build and validate system, in a standard robotic simulator with accurate physics

[email protected]
Single Horizontal Camera-Based Object Tracking Quadcopter … 115

and visuals as StaGaus algorithm which uses control along with image processing
algorithms. Since many of these surveys use OpenCV, OpenCV is used for simulation
of the system.
Literature survey was conducted to test our drone in the simulator that researchers
would use and that would simulate both physics and image processing in real time.
Aaron et al. describe the popular simulators of the time with the considerations
of ease of programming, popularity and the physics of the software simulations
[5]. Pablo et al. describe a large number of simulators for robots on the bases of
price, ease of development and usage, architectures, distributed computing and the
accuracy of the simulations [6]. From these researches, we concluded that ROS and
Gazebo provide stable simulations while still letting image processing algorithm
work along with LIDAR mounted on quadrotor. While we found simulators built on
different software like Unreal Engine, we did not find a stable object tracking-specific
simulator on ROS and Gazebo [7]. Also, researches have indicated that researchers
and users are very likely to use ROS- and Gazebo-based simulators [8]. We decided
to build multi-type, multi-robot simulator for object tracking.

4 Proposed Methodology for Physical Validation

Figure 1 describes the overall architecture. Ultrasonics sensor is used for the obstacle
of avoidance of obstacles. (HCSR04 has a range of 4 metres with a time limit of
60 milliseconds.) Ardupilot APM 2.6 flight controller, 2200 mAh lithium polymer
battery, four 850 kV brushless motors, four electronic speed controller cum battery
elimination circuit modules, propellers and two Android phones are used.
Ardupilot APM 2.6 supports USB communication, pulse width modulation of
50 Hz, micro-aerial link commands via telemetry port. It has provisions for external
sensors. It has a lot of internal sensors like accelerometer, gyroscope, compass,
internal barometer and 16 MB internal memory. It supports up to 18 V and 90 A
ESCs [9].

4.1 Proposed OpenCV-Based Image Processing Algorithm

The authors have used Open Computer Vision library because Java is well supported
in Android through OpenCV Manager. We use a user interface in Android in which
the user just draws a rectangular bounding box by drawing diagonally (user touches
on one corner of rectangle, triggers Action Down event, swipes till the diagonally
opposite corner and releases thereby triggering Action Up event of Motion Event
of Android). For this, GrabCut’s filtering is employed for initial image to segregate
foreground and background [10]. The posterior Bayesian smoothing provides filter-
ing around the local maximum [11]. If the image does not look fine, the user can just
repeat initialisation.

[email protected]
116 L. Shrinivasan and N. R. Prasad

Fig. 1 System architecture. Architecture of the proposed physical quadrotor system

As the processors’ performance is generally tuned to work well for hexagonal

approximation for colour spaces, the standard hexagonal approximation has been
used [12]. The centroid of the blob is calculated on the basis of confidence of the
features of the image. The centroid and size of the blob constitute the input vector
for the online training algorithm StaGaus. The blob would be derived out of the
filtered image of the object of interest derived out of rectangular area enclosed by
the rectangular bounding box. During live tracking, feature matching is used against
this reference blob. The main advantage of this method is that the users need not
enter parameters of the online training algorithm StaGaus. The parameters would be
optimised automatically over time. ORB is used for feature matching as it provides
the confidence of the features and also the feature matching. It is real time and is
suitable for Android phones [13].
A certain number of features has to be detected within the object (so that the
centroid would not shift). If the occlusion is higher and the number of features detects
gets lesser than the Hamming matcher threshold, then the image would be discarded.
For tracking, the position of centroid is used for X- and Z-axes (represented by yaw
and throttle, respectively). The root mean square of the distances of the features from
the centroid would determine the size of the blob. This size of the blob is used for
the third dimension Y -axis (represented by pitch).

[email protected]
Single Horizontal Camera-Based Object Tracking Quadcopter … 117

4.2 Proposed Linear Control Algorithm

Velocity is chosen as a linear function of the position of the centroid. (Experimental

setup proved that choosing acceleration as a linear function of the position of the
centroid would lead to transients and object tracking percentage drops drastically.)
Width and height represent the width and height of the image respectively. ωyaw
represents the angular velocity of the yaw. vthrottle represents the velocity of the
throttle. vpitch represents the horizontal and forward velocity.
 
−t |x < 0 0, t < 0
∀ mod (t)  step(t) 
t |x ≥ 0 t, |t ≥ 0

x and z are the horizontal and vertical distances of the left-top corner of the image.
    
width 3 × width
vyaw  1 − step x − − step x −
4 4
    
4 width
× 1− × mod x − (1)
width 2
    
height 3 × height
vthrottle  1 − step z − − step z −
4 4
    
4 height
× 1− × mod z − (2)
height 2

Equations (1) and (2) represent the yaw and throttle. For pitch control, the lesser
σ (Standard Deviation) means higher speeds. So,

− f (x)xdx − f (x)xdx
d f (x)  → f (x)  ce σ 2 (3)
σ2

ηindividual × centroidindividual (x, z)
μcentroid (x, z)   (4)
ηindividual
  
 width height 
∀distanceimage    , − μcentroid (x, z) (5)
2 2

The distance determined in Eq. (5) is used for the determination of statistics of
the object being tracked.

 2
distanceimage
σcentroid  (6)
n

Using Eqs. (3), (5) and (6), the equation deduced is used for pitch.

vforwardMax − 21 distanceimage 2

vforward  √ e σcentroid
(7)
σcentroid (2π )

[email protected]
118 L. Shrinivasan and N. R. Prasad

Equation (7) represents the third dimension derived out of two dimensions using
statistical analyses. Equations (1) and (2) account for the changing bounding box
size. On-field test demonstrated that full linearization of (1) and (2) results in yaw
oscillations. So, we have restricted the linearization to 50% exactly.
In StaGaus algorithm, if distanceimage > σ centroid , the existing curve of StaGaus
is used. But, the distanceimage would be used for the updation of the curve. Else,
curve is updated and centroid is used as per Eqs. (1), (2) and (7). This StaGaus
algorithm avoids position error that usually occurs in long-term learning. By using
the template and re-centering using the control algorithms, the target is not lost until
there are occlusions.

4.3 Proposed User Interface

The user interface determined the most of the development of autonomous algorithm.
Applications typically use a Ground Control Station (GPS) and are usually partially
autonomous (like the usage of GPS waypoints). But, with a single-camera on-board,
the quadrotor has an interface which just streams videos back along with the detected
object marked with a circle. The only control the user has is the emergency button
for an emergency joypad control. Pro Joypad and simple joypad modes are just for
initialisation. However, they can also be used for the manual testing of object tracking
algorithm. In Pro Joypad mode and simple joypad mode, several manoeuvres can be
tried with manual input with detected object while bypassing autonomous control.
As mentioned in the left side of Fig. 2, Pro Joypad mode would stream the video
while still maintaining a four-channel control. The four channels values: roll, pitch,
yaw and throttle values, are sent. The value of each channel would vary 1000–2000
which corresponds to the RC input values of the flight controller. This mode made
StaGaus compute the object in real time as the quadrotor also had to accept input
from Pro Joypad mode. Right side of Fig. 2 displays the simple joypad mode. This
mode is similar to Pro Joypad. But the controls are simple and are not meant for
manoeuvres.

Fig. 2 The Pro Joypad mode in the left and simple Joypad mode in the right

[email protected]
Single Horizontal Camera-Based Object Tracking Quadcopter … 119

Fig. 3 Object tracking mode indication in the Android application

User swipes across the screen once. User can select an image already stored on
phone or use camera (managed through Android intent) to capture a new image and
that can be loaded in the setup screen. The object is filtered using the bounding box
as the limit and background is removed. ORB feature matching starts.
Figure 3 demonstrates the tracking scenario. The feature matching is done using
ORB feature matching using brute-force Hamming matcher. Note that the algorithm
is very robust against rotations (both in-plane and out-of-plane). So, a circular pattern
reinforced the StaGaus algorithm regarding rotational invariance which boosted the
tracking percentage. GrabCut consumes about 0.6 s. This is sufficiently enough as
the features would be stored in global class of Android.

5 Proposed ROS-Based Gazebo Simulator

The authors have chosen to build Robotics Operating System-based Gazebo simulator
after literature survey indicated that Open Dynamics Engine (ODE) provides accurate
physics library and people are likely to use it [14, 6, 8]. We have built our custom
Robotics Operating System and Gazebo simulator from scratch using ODE. We used
ROS Indigo, Gazebo7 and Ubuntu 14.04 operating system. Pioneer 3DX is used as a
rover with a 64 × 48 resolution running at 5 Hz. It has an angular velocity of 1 rad/s
and has a velocity of 0.5 m/s. Copter is the quadrotor using 10 Hz LIDAR (only for
validation) and 320 × 240 resolution camera refreshed at 5 Hz. We have successfully
simulated the linear dynamics of the quadrotor with observable dynamics which is
as shown in Eq. (8).
      
F m0 a ω × mv
 + (8)
τ 0 I α ω × Iω

[email protected]
120 L. Shrinivasan and N. R. Prasad

Fig. 4 Object tracking initialisation in ROS Gazebo software in the loop simulation

F Net force of the quadrotor (thrust of the rotors)

τ Net torque of the quadrotor (yaw)
a Acceleration of the quadrotor
ω Angular velocity vector
I Inertia
α Angular acceleration
m Mass of the quadrotor
v Velocity of the quadrotor.

Figure 4 demonstrates the initialisation of object tracking using the camera that
is also simulated. The height is adjusted, and LIDAR is turned on so as to account
for the calculation of tracking percentage and detection percentage. This method
helps us test for conditions like occlusion without risking the breakage of physical
hardware. Both rover and the quadrotor are being simulated in a single computer.
Figure 5 depicts the physics tests of our simulator. It may be easy to get visually
nice looking simulator in Unity or Unreal. The control dynamics of quadrotor landing
on Pioneer 3DX has been simulated with aerodynamics.

6 Parameters Setup

Android phone’s NV21 images are converted to RGB. MJPEG streaming is achieved
by integrating our custom OpenCV components built from scratch (with the help of
Seed Studio’s Webcam Android application) [15].

[email protected]
Single Horizontal Camera-Based Object Tracking Quadcopter … 121

Fig. 5 Physics tests in ROS Gazebo software in the loop simulation

Arduino’s servo maps the range of 45°–135° to the range of 1000–2000 μs 50 Hz

PWM. PWM values of 1400–1600 are used with 1500 as the mid-value. ORB feature
matching is used along with brute-force Hamming matcher in Android phone for
10 frames of 9 × 9 kernel images. To avoid the frame rate dropping down to 1 frame
per second when internal memory or microSD memory is used, cache variables
are used in global class. Sony XPERIA Dual M2 mobile phone is being used with
a resolution of 320 × 240 for input image. This resulted in processing at about
150 frames per second for colour thresholding range of 1/6. Threshold for the features
of 10 frames of 9 × 9 kernel images is 2 frames.
If γ represents the Hamming matcher threshold,

γ + 2, n new < n template
y
γ − 1, n new ≥ n template

7 Results

For 640 × 480 image, detection rate has been close to 97% and tracking rate has
been close to 79% in the practical tracking validation (89% in SITL). Our StaGaus
algorithm is robust in rotational invariance. It handles both in-plane and out-of-plane
rotations of up to 30° on each side as demonstrated in Fig. 6.
The comparison of StaGaus with SRDCF and colour-filtered ORB tracking is
shown in Fig. 7. Note that we have also included plain RGB-filtered tracking for
reference (it loses track within first 20 s). SRDCF and StaGaus are close in both
detection and tracking percentages. The main advantage is that SRDCF achieves

[email protected]
122 L. Shrinivasan and N. R. Prasad

Fig. 6 Detection and tracking % of StaGaus algorithm observed with physical Android setup

detection and percentage through 2D images only without StaGaus controller. But,
when control algorithm is used, StaGaus performs much better in real-time perfor-
mance (by Big O considerations, by several times). Also, StaGaus does not have the
position drift that SRDCF suffers. StaGaus uses the template as the reference so that
the centroid is re-centred.
As mentioned in Fig. 7, the new designed algorithm is compared against three
more algorithms. Some of the earliest attempts included primitive form of colour
thresholding followed by blob detection by dilating and eroding (as in RGB thresh-
olding with 1/6 thresholding range without feature matching or using separate con-
troller algorithm that overrides FCU control algorithm). This method results in loss
of tracking when the background has a similar colour to that of the object being
tracked.
Similarly, for RGB thresholding with 1/6 colour range that is coupled with ORB
feature matching with the default controller algorithm that FCU offers, the tracking
is significantly better. But, yaw oscillations are observed when the object is directly
in front of the quadrotor, thus decreasing the tracking rate slightly. When StaGaus
controller algorithm is used, yaw oscillations are dampened making the flight stable.
Thus, experiments with both frequency domain and spatial domains lead to StaGaus.
% tracking
We have defined a metric ψ  −log10 memory×% CPUusage
to assess the algorithms
on time complexity, space complexity and tracking % with the same metric.
Table 1 contains the reported time complexity and space complexity of algo-
rithms. For StaGaus and RGB, feature detection requires minimum of 9 pixels and
n represents the number of features. Also, as StaGaus requires 10 frames of 9 × 9
image kernels, the appropriate offset is added. We consider 81 features, S  1, N ne 
1 and K  0. Rest of the parameters are as they are (values authored in their paper).

[email protected]
Single Horizontal Camera-Based Object Tracking Quadcopter … 123

Fig. 7 Detection and tracking rates of various algorithms with reported rates and SITL

Table 1 Comparison of ψ for various algorithms (worst case for StaGaus)

Algorithms Time complexity Space Time Space ψ
complexity
StaGaus Width × height + Width × height + 308,010 308,820 9.03
n + 9n n + 9n + 10 ×
9×9
SRDCF dSMNlog MN + l 2 + d + width × 66,300,583 307,282 11.4
SMNNNe +(d + height
K 2 )dMNNGS
RGB (1/6) ORB Width × height + Width × height + 308,010 308,820 9.11
(without StaGaus n + 9n n + 9n + 10 ×
Controller) 9×9
RGB (1/6) Width × height Width × height 307,200 307,200 9.21
(without StaGaus
Controller)

MN  50 × 50  2500 and N GS  4. We ignore trivial constants and variables that

do not significantly contribute to performance or memory.
It can be observed that higher ψ means that the algorithm does not adhere to
memory and performance constraints of mobile phones. We can see that SRDCF is
more than two orders higher than StaGaus on a log scale.

[email protected]
124 L. Shrinivasan and N. R. Prasad

Fig. 8 a Length of kernel image, b image resolution results in right

As shown in Fig. 8a when the size of the kernel is too small, the features are
mapped to the same cell resulting in loss of features. When the size of kernel is too
large, noisy features will also be mapped resulting in the deterioration of tracking rate.
When the image resolution gets higher, the detection rate and tracking rate get higher
as depicted in Fig. 8b. However, when Android phone processor’s performance hits
the peak, if the image resolution is increased, real-time processing no longer occurs
and this results in the slight reduction of tracking percentage.
The authors have relied on Android phone for most of the processing and for
real-time control, Arduino has been used. APM 2.6 flight controller has been used
for low-level quadrotor dynamics.
Regarding the simulations, Table 2 compares the various simulators meant for
object tracking drones. It is observed that our ROS and Gazebo simulator is capable
of simulating high-quality visuals while still maintaining physics. We have used the
metrics of evaluations as that of used in the popular literature [7, 6, 8]. “~” represents
the dependence of implementation (For example, high-speed simulations are possible
in V-REP at the cost of approximations in physics). “OS” in the cost’s column stands
for open source.
As can be observed from Table 3, the cost of proposed and designed drone has
been low (a by-product aspect) due to the StaGaus algorithm that has been used.
It reduces the need of expensive sensors, and it is able to run on normal Android
phones.

8 Conclusions and Future Work

This drone has been used with a horizontally placed camera. We have compared
various image processing algorithms and demonstrated that our algorithm works on
memory and performance-constrained devices like Android phones while maintain-
ing a higher tracking efficiency than other popular algorithms. We have also built

[email protected]
Single Horizontal Camera-Based Object Tracking Quadcopter … 125

Table 2 Comparison of various simulators meant for object tracking drones

Simulator Cost Physics Testing Simulation CPU % Memory Visuals
condi- pace %
tions
V-REP OS ~ ~ Real Low Low High
time
MATLAB/Simulink High ~ ~ Slow High High ~
Unreal engine Free ~ High Real Moderate Moderate High
(including time
Airsim)
ROS Gazebo Free ✓ Moderate Slow High High High
(TurtleBot)
ROS Gazebo OS ✓ High Real High High High
(Ours) time

Table 3 Costs of various object tracking drones

Drone Tracking principle Marketing stage Cost
TraQuad Imaging based – 17,000 Rs
AirDog Bluetooth based Existing 107,312 Rs
Solo FPV drone Shut down 53,656 Rs
DJI GPS based FPV Existing 37,420 Rs (min)
Xiaomi FPV drone Existing 34,812 Rs
Parrot FPV drone Existing 18,700 Rs (min)
Nixie Proprietary Pre-order –
Lily GPS based Shut down 61,704 Rs
Zero Zero robotics Proprietary Pre-order –
GoPro FPV drone Pre-order 59,872 Rs

a ROS Gazebo simulator and have efficiently built a unified software platform for
object tracking drone researchers.
Future work would be the addition of sensors (like one vertical camera) and
performing sensor fusion on it using deep learning. Multi-terrain quadrotor would
be a future work: The wheels would be used on roads (to save power) and the
quadrotor would be flown when necessary (to save time). Swarm object tracking
quadrotors for a single-object and multi-object tracking with swarm planning are
possible research aspects. Another future work would be an ISO packaged with native
GPU drivers of NVIDIA (the setup time would be reduced and the performance of the
simulations would be much higher). To enable these developments, we have included
the simulator’s setup and usage, placed the algorithms in the paper and open-sourced
the Android apps’ code [16].

[email protected]
126 L. Shrinivasan and N. R. Prasad

Acknowledgements We thank Vladimir Ermakov, leading developer of MAVROS for assistance

regarding MAVROS. We thank Venkat, Edall Systems, for assistance regarding PWM RC values
and also Harsha H N and Amrinder S R for their support in programming.

References

1. Yilmaz, A., Javed, O., & Shah, M. (2006). Object tracking: A survey. ACM Computing Surveys,
38(4), Article Number 13. https://doi.org/10.1145/1177352.1177355.
2. Smeulders, A. W. M, Chu, D. M., Cucchiara, R., Calderara, S., Dehghan, A., & Shah, M.
(2014). Visual tracking: an experimental survey. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 36(7). https://doi.org/10.1109/tpami.2013.230.
3. Yang, H., Shao, L., Zheng, F., Wang, L., & Song, Z. (2011). Recent advances and trends in
visual tracking: A review. Neurocomputing, 74(18). https://doi.org/10.1016/j.neucom.2011.07.
024.
4. Li, R., Pang, M., Zhao, C., Zhou, G., & Fang, L. (2016). Monocular long-term target following
on UAVs. CVPRW. https://doi.org/10.1109/cvprw.2016.11.
5. Staranowicz, A., & Mariottini, G. L. (2011). A survey and comparison of commercial and open-
source robotic simulator software. PETRA 11. In Proceedings of the 4th International Confer-
ence on Pervasive Technologies Related to Assistive Environments, Article No. 56. https://doi.
org/10.1145/2141622.2141689.
6. Blasco, P. I., Diaz, F., del Rio, M., Ternero, C. R., Muniz, D. C., & Diaz, S. V. (2012). Robotics
software frameworks for multi-agent robotic systems development. Robotics and Autonomous
Systems, 60(6), 803–821. https://doi.org/10.1016/j.robot.2012.02.004.
7. Mueller, M., Smith, N., & Ghanem, B. (2016). A benchmark and simulator for UAV tracking.
In Computer vision, ECCV. https://doi.org/10.1007/978-3-319-46448-0_27.
8. Ivaldi, S., Padois, V., & Nori, F. (2014). Tools for dynamics simulation of robots: a survey
based on user feedback, Humanoid Robots (Humanoids). In 2014 14th IEEE-RAS International
Conference. https://doi.org/10.1109/humanoids.2014.7041462.
9. APM 2.6 Documentation. http://ardupilot.org/copter/docs/common-apm25-and-26-overview.
html.
10. Rother, C., Kolmogorov, V., & Blake, A. (2004). GrabCut interactive foreground extraction
using iterated graph cuts. ACM Transactions on Graphics, SIGGRAPH. https://doi.org/10.
1145/1015706.1015720.
11. Greig, D. M., Porteus, B. T., & Seheult, A. H. (1989). Exact maximum a posteriori estimation
for binary images. Wiley for Royal Statistical Society, 51(2), 271–279.
12. Color Models, Image Color Conversion, Volume 2, Image processing, Intel Integrated Perfor-
mance Primitives for Intel Architecture Developer Reference. https://software.intel.com/en-us/
node/503873.
13. Rublee, E., Rabaud, V., Konolige, K., & Bradski, G. (2011). ORB: An efficient alternative to
SIFT and SURF, Willow Garage. In ICCV Proceedings of the 2011 International Conference
on Computer Vision, pp. 2564–2571. https://doi.org/10.1109/iccv.2011.6126544.
14. Configuring your Environment, Erle Robotics. http://docs.erlerobotics.com/simulation/
configuring_your_environment.
15. Seeed Studio’s Webcam Android Application. https://github.com/xiongyihui/Webcam.
16. Our core softwares for physical quadrotor. https://github.com/traquad.

[email protected]
IoT-Enabled Medicine Bottle

A. R. Shreyas, Soumya Sharma, H. Shivani and C. N. Sowmyarani

Abstract Internet of Things (IoT) has facilitated a reassuring opportunity to develop

powerful and dynamic industrial systems and applications by making use of uni-
versality of sensors, actuators and network connectivity. IoT can immensely benefit
people when used in the medical field. Non-adherence to a medical routine is a major
problem faced by patients. According to the World Health Organization, increasing
the effectiveness of adherence interventions may have a far greater impact on the
health of the population than any improvement in specific medical treatments. In an
attempt to address this problem, in this paper, a special system that monitors and
tracks the consumption of medicine of any patient is presented. Adherence needs are
met through the implementation of a mobile application. The result of this will be
timely consumption of the medicine by the patients.

Keywords Internet of things (IoT) · NodeMCU · Infrared sensor · Ultrasonic

sensor · Firebase cloud

1 Introduction

The Internet of Things (IoT) has facilitated the development of dynamic and power-
ful industrial systems. Network connectivity and integration of individual physical
devices being the foundation of IoT have found use in multitudinous fields, with each
field constructing its own platform on this very foundation. Currently, the agricul-
tural industry, automobile industry, oil and gas industry and of course, the medical
field heavily depend on IoT [1–4]. Harnessing the power of IoT leads to efficient
solutions that cater to diverse needs, and present a complete package that serves a
singular aim.
Technical advancements in modern medicine have been rampant, providing waves
of solutions to diseases that were considered incurable a couple of decades back.
However, at the most basic level, the efficacy of those advancements comes about

A. R. Shreyas (B) · S. Sharma · H. Shivani · C. N. Sowmyarani

Department of CSE, R.V College of Engineering, Bangalore 560059, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 127

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_11

[email protected]
128 A. R. Shreyas et al.

only when the medicines are consumed at the right time, in the right quantity. Not
adhering to a routine might lead to catastrophic results, like severe degradation in a
patient’s health, rendering the use of modern medicine futile.
Currently, the solution in use place to the aforementioned problem is the employ-
ing of helpers or caretakers to constantly monitor the patients or the use of alarm
clocks. The former case maybe expensive and unreliable, while the latter has several
restrictions, as it is not integrated medicinal bottles, to be able to gauge the changes
in quantities consumed by the patient.
With the creation of IoT-enabled medicine bottle, and the support of an application,
not only can timely reminders be sent to the patient on his mobile phone, but also the
quantity consumed can be monitored as well. Through this implementation, several
other features can be reinforced.
Such a system is efficient and systematic, as is explained through the course of the
paper. The use of sensors coupled with a WiFi (Wireless Fidelity) module provides
the hardware support. Data analysis and integration is done on the firebase platform,
the result of which is elegantly presented in a simple to use application. Hence an
IoT based system has been successfully implemented.

2 Motivation

The treatment of most diseases invariably includes the use of pharmacotherapy, which
refers to therapy using pharmaceutical drugs. However useful these medicines maybe,
adherence is the key to make them effective. According to a WHO (World Health
Organization) report, adherence rates in developed countries is a mere 50% and worse
in developing countries. Among patients with chronic illness, approximately 50%
do not take medications as prescribed [5]. Thus, there is an urging need to oversee
the medical consumption on a daily basis. A severe example is of hypertension (high
blood pressure), where according to WHO, up to 80% patients suffer because of non-
adherence. Specifically, they claimed that improving adherence to medical therapy for
conditions of hypertension, hyperlipidemia, and diabetes would yield very substantial
health and economic benefits. In another bold statement, they stated that a rise in
adherence may have a greater effect on health than improvement in specific medical
treatments. Effective solutions to combat this problem are not in place currently
either. Thus, motivated by the multiple ramifications of non-adherence, and a lack
of an existing efficient system, the IoT-enabled medicine bottle was developed.

3 Related Work

In this section, we will briefly highlight the related researches of this paper. There are
multiple researches that have proposed various models with respect to IoT-enabled
medicine applications. First, a medication adherence monitoring system was devel-

[email protected]
IoT-Enabled Medicine Bottle 129

oped that is based on a wearable sensor [6]. Authors of this paper focus on two
actions of twist-cap and hand-to-mouth and the act of pill intake is then identified
by performing a moving window dynamic time warping in real time between sig-
nal templates and the signals acquired by the wearable inertial sensor. In another
research, the authors have developed an alarm system for the community health care
[7]. They track state changes of medicine bottles by using weight sensors. Accord-
ing to the tracking record, their proposed system will decide whether or not to alarm
the patient. Rohokal et al. proposed a feasible IoT approach for the better health
surveillance of poor and rural human being’s health issues related to any part of
the body [8]. In another system, an IoT framework was developed where medicine
intake was ensured using activity recognition and an activity classification scheme
[9]. Other work has been done to make intelligent medication boxes [10–12], and
also to improve medicine adherence [13].

4 Design

This section proposed the system architecture. The system is composed of three
modules, where the first one is the alarm module, which notifies the user when it is
time to consume medicine. The second module is the sensor module that continuously
gathers sensor outputs and conveys them to the third module, which is the classifier
module. This module determines whether medicine has been consumed or not.
The various steps are illustrated on a flow diagram in Fig. 1.
The IoT has facilitated the development of dynamic and powerful the system can
be divided into various steps
1. The user sets the times at which medicines have to be consumed. The user does
this on the mobile application that has been specially designed for the system.
2. When it is time to consume the medicine, the mobile application notifies this to
the user.
3. The system waits for a certain time duration, after which the classifier module
decides whether the user has actually consumed the medicine or not. The classifier
module does this based on the results obtained from the sensor module.
4. If the medicine has not been consumed, the user is notified another time.
5. The system waits for another time duration, after which the system once again
checks the status of the medicine consumption. The result obtained is added to
the weekly report that can be viewed in the application.
6. In case the amount of medicine in the bottle is less than a certain threshold, the
user is notified.

[email protected]
130 A. R. Shreyas et al.

Fig. 1 Flow diagram for the

system: after the medicine
timing is set the above steps
are followed whenever it is
time to consume medicine

5 Implementation

In order to achieve the medication monitoring system, an IoT-enabled medicine bottle

is developed. The user sets reminders on the mobile application, for the tablets he
needs to consume as prescribed. Our surveillance approach includes the detection
of whether the bottle is opened or not and also the difference in heights before and
after the medicine consumption. The decision for the second action is taken only
after the first action, i.e., only after the bottle cap is opened. The decisions taken are
based on the sensor readings sent to the cloud database. Figure 2 shows the various
components of the system.
FC-51 (Infrared Sensor), Infrared Obstacle Avoidance Proximity Sensors Module
has a built-in Infrared transmitter and receiver that emits IR (InfraRed) energy and
checks for a reflected IR energy in order to detect any obstacle in front of the sensor.
This inexpensive sensor is used to detect if the bottle cap is opened or not.
Further, HC-SR04 (Ultrasonic Sensor) can be used to measure small distances [14,
15]. It is used here to measure the difference between the height of medicine bottle
before and after medical consumption. This economical sensor provides 2–400 cm
of noncontact measurement functionality with a ranging accuracy that can reach up
to 3 mm.

[email protected]
IoT-Enabled Medicine Bottle 131

Fig. 2 Implementation of the proposed system: readings from the infrared sensor (FC-51) and ultra-
sonic sensor(HC-SR04) are monitored by the NodeMCU, and data is sent to the mobile application
via firebase

The data from these two sensors is collected by a versatile open source IoT plat-
form, NodeMCU 1.0 (ESP8266-12E). Figure 3 shows the hardware components
wired together which act like the bottle cap.
This sensor is initially programmed once using Arduino IDE 1.8.4 to perform
the function of a microprocessor for the sensors and as well as a WiFi module to
transmit the data to Firebase which is a mobile and web application development
platform. The NOSQL (non-Structured Query Language) feature of Firebase real-
time database makes it extremely flexible to store the information from hardware
[16]. It also permits the user to view the information obtained from the hardware
components in a real time from any device with permissions that has access to the
network. It stores data in key-value pairs. The necessary information that the database
needs to store is whether the bottle is opened or not, which can be indentified using
Boolean values and the height of the medicine stack in the bottle, which can be stored
using floating point values.
Figure 4 shows the IoT-enabled medicine bottle.
The mobile application is developed with the help of Android Studio. This appli-
cation collects the data from firebase and takes required decisions accordingly.

[email protected]
132 A. R. Shreyas et al.

Fig. 3 Medicine bottle cap-

infrared and ultrasonic
sensors attached to the cap
along with the NodeMCU

Fig. 4 Cap being placed on

the medicine bottle

[email protected]
IoT-Enabled Medicine Bottle 133

6 Results

The system is developed to detect changes in the contents of the medicine bottle in
real time and respond to the changes appropriately by notifying the user. The system
also notifies the user in case the contents of the bottle are going empty.
The results of the experiments are shown in the following section. Figure 5 shows
the current contents of the medicine bottle. In Fig. 6a, the user sets a reminder on
the mobile application. The user gets a notification (as shown in Fig. 6b at the time
in which he sets the reminder, to consume the medicine. In case the user does not
consume medicine at the right time, the android application reminds the user on
another occasion as shown in Fig. 6c. A weekly report is also generated as shown in

Fig. 5 Page on the

application showing contents
of medicine bottle

[email protected]
134 A. R. Shreyas et al.

(a) Setting reminder (b) Viewing the set reminders

(d) Second notification when

(c) First notification
pills were not consumed

Fig. 6 Performing an experiment using the proposed implementation

Fig. 6d. In the same figure, the Yes or No signifies if the pills were consumed when
the reminder went off, or/if they were not consumed.
The remainder of the section presents statistics obtained after performing trials.
In Table 1, the first column indicates if the bottle was actually opened or not. The
next two columns show a comparison between the depths of the top of the medicine
from the ultrasonic sensor. The fourth column indicates whether the medicine was

[email protected]
IoT-Enabled Medicine Bottle 135

Table 1 Results after 6 trials

Bottle opened? Initial Final depth Medicine Report Report
depth (cm) (cm) actually (shown in correct?
consumed? app)
No 4.56 4.56 No No Correct
No 3.86 3.86 No No Correct
Yes 6.79 7.02 Yes Yes Correct
Yes 5.23 5.51 Yes Yes Correct
Yes 4.56 4.54 No No Correct
Yes 3.86 4.10 No Yes Incorrect

actually consumed or not. The next column shows the report generated by the mobile
application. Lastly, the correctness of each report is mentioned.
There are three outcomes to the experiment.
1. The bottle is not opened.
2. The bottle is opened and medicine is consumed.
3. The bottle is opened but medicine is not consumed.
The bar charts in Figs. 7 and 8 show the results obtained after performing 60 trials
using syrup and pills as the medicine. The x-axis shows the various outcomes to the
experiment. The y-axis shows the count obtained in each case as mentioned in the
chart.
From these results, we can estimate the accuracy of the system.

Accuracy  Number of Correct Results/Number of Trials. (1)

For syrup as medicine,

Accuracy  (20 + 18 + 19)/(20 + 20 + 20)  0.95

So the system has an accuracy of around 95% with syrup as the medicine.
For pills as medicine,

Accuracy  (20 + 18 + 15)/(20 + 20 + 20)  0.883

So the system has an accuracy of around 88.3% with pills as the medicine.
From these results, we can conclude that the system works more accurately when
syrup is stored in the bottle rather than pills.

[email protected]
136 A. R. Shreyas et al.

Fig. 7 Accuracy evaluation of the proposed system with syrup as the medicine

7 Limitations

The proposed model suffers from three main limitations. First, the system is unable
to guarantee that the patient actually consumes the medicine, ensuring only the
removal of the medicine from the bottle. Second, a majority of the patients that are
unable to adhere to their medical routine include the elderly who may not possess
sufficient knowledge on how to use the application and to utilize its benefits to
the fullest extent. Helping them understand the system maybe the first hurdle, and
the elderly maybe be unwilling and unresponsive to such a method of improving
adherence. Thirdly, since the model monitors variations in height to predict when
to send notifications, insignificant changes in the height may go undetected by the
sensors, leading to erroneous results. This, however, can be eliminated by using more
powerful hardware support. Also, the bottle must be kept away from sunlight at all
times, as the sensors are sensitive to sun rays. However, most medicinal bottles are
generally stored in dark and cool places so this limitation can be eliminated with
ease too. Additionally, trivial reasons like the phone being out of battery or switched
off may cause hindrances in the efficient results that the system aims to provide.

[email protected]
IoT-Enabled Medicine Bottle 137

Fig. 8 Accuracy evaluation of the proposed system with pills as the medicine

8 Future Work

A potent method to make the system more streamlined and effectively monitored is
by extending the application to provide doctors a view of their respective patients’
routines, and view their adherence levels regularly. This can help the patients and
their doctors to work towards recovery in a more steady and harmonious manner.
Also, to make the medicinal bottles easy to carry and use, the model can be
applied to a smaller medicinal bottle, which requires the assistance of more powerful
hardware support. By improving the strength of the hardware used, the same model
maybe applied to medicines of a relatively smaller dimension too, which as earlier
mentioned is a limitation of the current model. The IoT-enabled bottle may also be
presented in varying shapes, to suit the needs of different medicines. Encasing the
hardware over the bottle, in order to provide it protection from any liquid medicine
or any powdered medicine is another requirement, which might otherwise spoil the
hardware.
Another technique to improve the current model, maybe to provide it in multiple
languages. As already stated, the elderly forms a major consumer base for the model,
and being provided of a solution in their native language might prove to be a relief
to many, and increase the reach of the model.

[email protected]
138 A. R. Shreyas et al.

9 Conclusion

The medicine bottle developed by incorporating the principles of the IoT aims at
helping patients recover speedily and in a steady manner. On entering the quantity to
be consumed, and setting reminders on when to consume the medicines, the patients
can remain assured that they need not keep track of the number of times they have
consumed in the day, the application will take care of that by sending notifications
when it is time to consume. Weekly reports of consumption and adherence can be
shared with the doctor or family members too.

References

1. Uddin, M. A., et al. (2017). Agriculture internet of things-Ag-IoT. In 27th International

Telecommunication Networks and Applications Conference (ITNAC) (pp. 1–6).
2. Zhang, X., et al. (2017). Monitoring citrus soil moisture and nutrients using an IoT based
system. IEEE J. Sens., 17(3), 430–447.
3. Kabir, M. H., et al. (2017). A low cost sensor based agriculture monitoring system using
polymeric hydrogel. J. Electrochem. Soc., 164(5), 107–112.
4. Mahendra, B. M., et al. (2017). IoT based sensor enabled smart car parking for advanced driver
assistance system. In 2017 2nd IEEE International Conference on Recent Trends in Electronics,
Information & Communication Technology (RTEICT) (pp. 2188–2193).
5. Brown, M. T., & Bussell, J. K. (2011). Medical adherence: WHO cares? Mayo Clinic Proceed-
ings, 86(4), 304–314.
6. Chen, C., et al. (2014). A medication adherence monitoring system for pill bottles based on
a wearable inertial sensor. In 2014 36th Annual International Conference on Medicine and
Biology Society (EMBC). IEEE.
7. Sohn, S. Y., et al. (2014). Alarm system for elder patients medication with IoT-enabled pill
bottle. In 2015 International Conference on Information and Communication Technology Con-
vergence (ICTC) (pp. 60–61). IEEE.
8. Rohokale, V. M., et al. (2011). A cooperative internet of things (iot) for rural healthcare monitor-
ing and control. In 2011 2nd International Conference on Wireless Communication, Vehicular
Technology, Information Theory and Aerospace & Electronic Systems Technology (Wireless
VITAE) (pp. 1–6). IEEE.
9. Serdaroglu, K., et al. (2015). Medication intake adherence with real time activity recognition
on IoT. In 2015 IEEE 11th International Conference on Wireless and Mobile Computing,
Networking and Communications (WiMob) (pp. 230–237).
10. Gomes, C. E. M., et al. (2015). Extending an intelligent medicine cabinet through the use
of consumer electronic devices in order to increase the medication adherence. In 2016 6th
International Conference on Information and Communication Technology for the Muslim World
(ICT4M) (pp. 98–102).
11. Xu, L., et al. (2014). A health-IoT platform based on the integration of intelligent packag-
ing, unobtrusive bio-sensor, and intelligent medicine box. IEEE Transactions on Industrial
Informatics, 10(4).
12. Magalhães, L., et al. (2017). A three-staged approach to medicine box recognition. 2017 24º
Encontro Português de Computação Gráfica e Interação (EPCGI), pp. 1–7.
13. Rodrigues, M. A. S., et al. (2015). An intelligent medication system designed to improve the
medication adherence. In 2015 IEEE 5th International Conference on Consumer Electronics -
Berlin (ICCE-Berlin) (pp. 46–49).

[email protected]
IoT-Enabled Medicine Bottle 139

14. Zahir, E., et al. (2017). Implementing and testing an ultrasonic based mobility aid for a visually
impaired person. In 2017 IEEE Region 10 Humanitarian Technology Conference (R10-HTC)
(pp. 453–456).
15. Mantoro, T., & Istiono, W. (2017). Saving water with water level detection in a smart home
bathtub using ultrasonic sensor and Fuzzy logic. In 2017 Second International Conference on
Informatics and Computing (ICIC) (pp. 1–5).
16. Alsalemi, A., et al. (2017). Real-time communication using firebase cloud IoT platform for
ECMO simulation. In 2017 IEEE International Conference on Internet of Things (iThings)
and IEEE Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and
Social Computing (CPSCom) and IEEE Smart Data (SmartData) (pp. 178–182).

[email protected]
Revamp Perception of Bitcoin Using
Cognizant Merkle

J. Vijayalakshmi and A. Murugan

Abstract The bitcoin network adopts the regulation of peer-to-peer network for stor-
ing and retrieving data. Data synchronization and consistent verification is important
in peer-to-peer networks for proving truth of data. Proof of existence of data is impor-
tant for many applications and it can be verified by means of searching process. The
goal of this paper is to minimize the time of verifying the transaction residence in
the shared ledger of bitcoin network. In global ledger management like blockchain,
the bitcoin transaction verification and validation is fundamental which is mainly
used by miners for providing the proof-of-work for the transactions to achieve block
reward and providing trust among the peers. In this paper we have provided a new
way of verifying the transaction existence in blockchain by means of altering the data
structure of Bitcoin Merkle into a new form of Cognizant Merkle which modifies
the structure of existing bitcoin system into a new form which uses less memory and
achieve more speed compared to Bitcoin.

Keywords Bitcoin · Mining · Bitcoin Merkle · Cognizant Merkle · Proof-of-Work

1 Introduction

The evolution of financial crisis in the year 2008 gives rise to the landing of cryptocur-
rency like bitcoin in the business world. This cryptocurrency provides secure digital
transmission, reduces the transaction complexity, and prevents the seeking of inter-
mediaries like security brokers, insurance agents, financial lawyers and credit card
companies [1]. The need for electronic commerce payment system based on lexicog-
raphy proof instead of centralized acception is that the transactions are irreversible,
preventing from fraudulent transaction and protecting the buyers from deceptive per-
son across the network [2]. Bitcoin is a first decentralized cryptocurrency which was

J. Vijayalakshmi (B) · A. Murugan

Research Department of CS, Dr. Ambedkar Government Arts College, Chennai 600039, India
e-mail: [email protected]
A. Murugan
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2019 141
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_12

[email protected]
142 J. Vijayalakshmi and A. Murugan

introduced by Satoshi Nakamato in the year 2009 for implementing virtual commod-
ity. This cryptocurrency is treated as “Gold standard” which is widely accepted for
providing services by various retailers like Amazon, Subway and Victoria Secret [3].
Bitcoin is the basis of digital money and it act as currency unit that is used to
cache and broadcast the value among the members across the blockchain. Bitcoin
network is available as free source software which can be run on multiple computing
devices like laptops, smart phones and it can be communicated through internet and
other transport networks. This can be acquisted, sold and transferred at specialized
currency exchanges in the perfect form of money for fast, secure and border less
in internet. Bitcoin is flip-side of virtual money which is fully distributed across
peer-to-peer network [4]. The following few features make the bitcoin transaction
more appealing for vendors, they are bitcoin transactions are publicly available.
They can have multiple inputs and multiple outputs and the consumer and merchant
are identified through public keys and public–private key pairs [5]. The motivation
behind the usage of bitcoin is anyone can use bitcoin anywhere around the world.
Even novice user can use this bitcoin for buying and selling their products like any
other fiat currencies [6].

2 Bitcoin Mining

Bitcoin system is connected with users along with wallet containing key functions
which are propagated across the network. The miners are responsible for produc-
ing the consensus blockchain, which is the authentic ledger of all the transactions.
Mining process discover new bitcoins through a solution to the associated difficult
problem. Any colleague in the bitcoin network running full bitcoin node may act as a
miner. For every ten minutes a new quick fix is found by someone who then confirm
transactions for past ten minutes and compensated with fresh unused 25 bitcoins
[2]. The bitcoin contract includes implicit algorithms that coordinate the mining task
across the bitcoin network which prevents the central repository need in financial
institutions. The key innovation of a miner [7] is the use of distributed computation
algorithm (Proof-of-Work) which is used to conduct a global election for every ten
minutes for arriving the consensus about the state of transactions [4].
Transferring of bitcoins from one or numerous source accounts to one or differ-
ent destination accounts is called transaction [7]. A bitcoin transaction reveals the
blockchain network that the bitcoin owner is authorized to transfer some bitcoins
to another owner. Another transaction is created by the endorsed owner who can
now spend these collected bitcoins and by licensing the ownership transfer of bit-
coins to another in a chain of ownership. Transactions are analogous to double entry
bookkeeping ledger. The summation of input and output may not be same. The trans-
action productivity is always less than total aid where the change is paid as transaction
reward for miners for including the transaction in the shared ledger. Not all existing
transactions in the network are included in the blockchain only those transactions
that are verified by miners are included in the blockchain. Miners do the transaction

[email protected]
Revamp Perception of Bitcoin Using Cognizant Merkle 143

proven process by accumulating transactions into blocks which require huge amount
of computation to prove but only a less amount of computation to justify.
The principle of mining is to create new bitcoins in each block and to create
trust by ensuring the authorization of transaction. The transactions are organized as
a ordered back linked list of blocks which is used as representation of blockchain
data structure. Each blocks are linked back to the previous block in the chain. Block
height is calculated from recently added block to the initial block. Each block is
identified through hash on the block header and the sequence of hashes looking
each block to its parents creates a chain to the genesis block. The structure of block
includes blocksize, block header, transaction counter and transactions. The block
header format includes software version number, hash of previous block, merkle
root, blocktime stamp, target difficulty and Nonce value. Blocksize denotes the total
bytes of blocksize. Transaction counter indicates the statistics of transactions where
transactions indicates recorded agreement in this block.
The Merkle root point out the root hash that comprises of all subordinated transac-
tion hashes within the current block. The timestamp denotes the approximate block
creation time. The target value gives the block Proof-of-Work. The nonce value is a
counter used for the proof-of-work. Each block contains a merkle tree which con-
tains the summary of all the transactions in the block. The proposed work is focused
on the revision of the merkle tree data structure which achieves less memory and
more speed for verification of transaction survival. The transaction data existence
in current block is identified using merkle tree and its proof is verified using pre-
vious blockhash. Merkle trees are exercised in distinct fields like Digital Currency
(Bitcoin, Ethereum), Global Supply Chain (IBM and Laersk), Health Care Indus-
try (Deep Mind Health), Capital Markets, Git and Mercurial platforms and Apache
cassandra [8].
Merkle [9], patented the concept of merkle tree in the year 1979. In peer-to-peer
network, data verification is important because the cloned data may exist in multiple
locations. Using merkle tree, data verification, synchronization and consistency can
be easily achieved. Merkle trees ensures the authenticity of the data by verifying the
data before downloading a file from peer-to-peer nodes and it can also be used to
verify the information coming from the untrusted resource. The root hash is obtained
from the trusted source and then lower nodes of the merkle tree which is a part of the
same tree can be obtained from the untrusted peers. The nodes from the untrusted
sources are checked against the trusted hash for matching process. If they match
then the nodes are accepted and the process continues if not, they are discarded and
searched again with the different source.

3 Literature Survey

Current bitcoin system uses merkle tree implementation based on Binary Hash Tree
(BHT) which occupies more memory and more time for searching and data retrieval.
For this, alternative data structures which supports string data operations based on

[email protected]
144 J. Vijayalakshmi and A. Murugan

tree is analyzed. One of the alternative data structure for BHT is a trie data structure.
The other alternatives are hash table which is efficient and relatively fast but unsorted.
A trie is a data structure which is an alternate to BHT, which is used for storing strings
in sorted order which facilitate the dynamic support of data structure [10]. Searching
in the trie is faster which require separate pointer traversal for each character in the
query string.
Alternatively Heinz et al. [11] proposes a Burst-trie which is considered as effi-
cient for small set of keys or values which is stored in container providing faster
retrieval than a conventional tree. A Burst-trie is a data structure similar to trie but
the leafs of the trie are replaced with container which is able to store a small number
of string efficiently. Primary goal of Burst-trie is to lessen the moderate the number of
string comparisons during a process of searching a string but it follows storing more
frequent terms than less frequent terms [10]. Our target is to analyze the performance
of different prototypes such as hash table, tries and others on Resource Description
Framework (RDF) [12].
Of this a hybrid data structure of HAT-trie [11] which was proposed by Nikolas
et al. is considered for implementation. The reason for this is, cache-attentive data
structure which supports the combination of trie along with hash table and Burst-trie.
HAT-trie is an extension scheme of Burst-trie which replaces the linked list containers
with cache aware based hash tables [13]. Ruslan Mavlyutov et al. proposed that HAT-
trie is best option for memory utilization, packing time and look up for string in the
tree data structure.

4 Bitcoin Merkle

In Bitcoin, merkle trees are used to encode blockchain data in a secured manner.
Merkle trees are used to produce a Digital Finger-print of the entire set of trans-
actions. This can also be used to verify the existence of transactions in the block.
Simplified Payment Verification (SPV) [2] nodes in the bitcoin network uses merkle
trees. SPV nodes can be used to verify the payments without running a full network
node. Using the merkle path, the SPV node connect the bitcoin transaction to the
appropriate block and also verify its inclusion within that block. The transaction
existence within the blockchain or not is identified by the combination of the links
between transaction and block, and the link between block and the blockchain. In
Bitcoin, a merkle tree is constructed by recurrent hashing the nodes pair until there is
only one hash called merkle root which is the public key of merkle signature system
which requires estimation of 2h leaves and 2h − 1 hash computations where h denotes
the merkle tree height [14].
A BHT is a self balancing tree whose frequently accessed nodes are moved close
to the root and every terminal node is labeled with a data block and every non terminal
node is labeled with the cryptographic hash of the labels of its child nodes. Double
SHA256 cryptographic hash is used for increasing the efficiency and improving the
security of bitcoin transaction. The merkle tree structure is based on BHT, the tree

[email protected]
Revamp Perception of Bitcoin Using Cognizant Merkle 145

Merkle Root
H0123=HASH(H01+H23)

H01=HASH(H0+H1) H23=HASH(H2+H3)

H0=HASH(T0) H1=HASH(T1) H2=HASH(T2) H3=HASH(T3)

Fig. 1 Merkle tree construction based on even number of nodes

Merkle Root
H0122=HASH(H01+H22)

H01=HASH(H0+H1) H22=HASH(H2+H2)

H0=HASH(T0) H1=HASH(T1) H2=HASH(T2) H2=HASH(T2)

Fig. 2 Merkle tree construction based on odd number of nodes

must follow the representation of even number of transactions. If the leaf contains
even number of transaction it is used as it is for merkle tree which is showed in Fig. 1.
If number of transaction is odd sequence, the last transaction will be duplicated to
achieve counterbalanced which is depicted in Fig. 2. Look at an example of merkle
tree development in bitcoin transactions T 0, T 1, T 2, T 3 which is represented as
leaves of the merkle tree. The hashes of the transactions are stored directly in merkle
trees in each terminal node like H0, H1, H2, H3. The hashes are determined by
H0  SHA256(SHA256(T 0)) and H01  SHA256(SHA256(H0 + H1)). In Fig. 1
merkle tree is produced for even number of nodes. Initially transaction hashes for
T 0, T 1, T 2 and T 3 is calculated as H0, H1, H2 and H3. Then intermediate hashes
are computed at each level to achieve the final hash merkle root as H0123. In Fig. 2
merkle tree is established based on odd number of nodes.
In a blockchain, every block contains few hundreds of transactions. If someone
wants to verify the transaction existence in a specific block then he needs to download
the entire blockchain. Instead, the proposed approach can download only a branch
of the merkle tree in the blockchain. The blockchain contains those transactions
obtained using Light Client or Simplified Payment Verification (SPV). This can check

[email protected]
146 J. Vijayalakshmi and A. Murugan

Merkle Root
H0123=HASH(H01+H23)

H01=HASH(H0+H1) H23=HASH(H2+H3)

H0=HASH(T0) H1=HASH(T1) H2=HASH(T2) H3=HASH(T3)

Fig. 3 Merkle tree path used to justify the incorporation of data

the hashes on traversing up the tree branch. If these hashes are equal then we know
that this particular transaction exists in this block. When a SPV node inspect the
particular transaction say T 2 in the merkle tree, then it can prove the inclusion of
transaction T 2 in the block by producing a merkle path which consists of hashes
H2, H23 and H0123 from the merkle tree and find out whether it is included in that
block or not. With these hashes a SPV node can confirm that this transaction T 2 is
included in the merkle tree. The above searching process is depicted in the following
Fig. 3 which states whether the transaction T 2 is included or not.

5 Cognizant Merkle

The proposed system aims to achieve faster searching of transactions available in

the merkle tree structure for its existence and its nonexistence. The recommended
system combines the features of merkle tree along with HAT-trie to achieve more
speed compared to Bitcoin Merkle. HAT-trie can be represented in two alternatives
namely hybrid and pure HAT-trie which is exhibited in Fig. 4. HAT-trie variants
are based on the technique of bucket management and splitting [15]. Pure HAT-
trie employs the approach of burst-trie, which is faster but requires extra space for
application. In pure HAT-trie, a single parent pointer is used as reference when the
buckets contain strings that share only one common prefix. In a pure HAT-trie, a
full bucket is bursted into small buckets of at most n buckets where n is the size of
alphabet that are parented by a new trie. The removal of leading character in strings
that are stored in the pure buckets are depicted in Fig. 5. Original strings can be
consumed during the splitting procedure.
The hybrid HAT-trie employs B-trie splitting algorithm which reduces the buckets
number with low cost and support several pointers to a bucket for supporting fast
access. In case of hybrid HAT-trie, the last character part of prefix is stored along with
string and the buckets also share a single prefix, but it is not removed. This has more
than one parent pointer. This may also contains pure buckets. HAT tries are assembled

[email protected]
Revamp Perception of Bitcoin Using Cognizant Merkle 147

A B C . . . . X Y Z

A B C . . . . X Y Z

A-M N-Z 4FTER\0

8IGHT5LEVEN\0

3OUR4IVE6IFTEEN\0

(Hybrid node) (pure node)

Fig. 4 HAT-trie with hybrid HAT-trie node representation for values 8, 11, 4, 15, 6, 7 and water

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

3REE\0

6IRTEEN 4IRTY 8IRTYFIVE\0

(Array Hash table)

Fig. 5 HAT-trie with pure HAT-trie node representation for values of 3, 13, 30 and 35

and inspected in a top-down fashion. Initially it starts as a single unfilled hybrid bucket
which is occupied until full. Buckets wont record duplicate values hence when a
bucket is full, it is bursted based on the type of HAT-trie. After bursting, a new parent
trie along with pointers are created where pure bucket is converted into a hybrid.
The previous trie is moved as grandparent and the bursting procedure continues
as a hybrid. To burst into a hybrid HAT-trie, a suitable split-point is chosen which
achieves even distribution and is not applicable for all cases. String can be accessed by
accumulating the occurrence of every leading character whose reoccurrence are then
traversed in a linguist order to determine the number of movements of string. Once
obtained, the split-point is elected as current occurrence counter alphabet. Based on
split-point buckets are classified as pure or hybrid. Buckets of empty type are deleted
and their corresponding parent pointers are set to empty.

[email protected]
148 J. Vijayalakshmi and A. Murugan

6 Results and Discussion

In Bitcoin, merkle root is composed by applying SHA256 as binate for each and
every transactions which is depicted in Fig. 6. Merkle root is constructed based
on cryptographic hashes which is already discussed in Bitcoin Merkle. The planned
system construct the merkle root based on HAT-trie which is both memory and speed
improvement cache keen data structure. This uses base 64 encoding for preventing
data loss across communication. This uses bit-wise hash to reduce the memory for
storing and retrieval which is depicted in Fig. 7.
The proposed merkle tree serve buckets as an array of n + 1 word-length pointers,
which are vacant or point to their respective slot entries. A dynamic array or bucket is
used for storing base-64 encoded transactions. Here Base-64 encoding [16] is chosen
for encoding the transaction because it can map all type of characters like ASCII,
UTF8, UTF16 and other types and this can be transported to any systems in the
network without data loss or modification of the content. The required slot is found
by using the combination of base-64 encoding and fast bit-wise hash function [17]
for performing insertion or searching of a transaction on a particular bucket. This
is implemented based on shift-add-xor string hashing functions which are resistant
of adversarial attack and the merkle root is determined based on the first encoded
character. In some cases, space is wasted due to few string insertions for this; we
implement the bucket as array hash table which makes changes from simple array
when it is overflowing.
The projected system is compared with the speed of construction, searching,
and memory requirements of Bitcoin Merkle and the conclusion is portrayed in
Table 1. Our measurements of various factors indicate that the memory size is less
and the searching speed is fast in Cognizant Merkle compared to Bitcoin Merkle.
The searching speed of transaction in Bitcoin Merkle is very slow compared to
Cognizant Merkle. Our dataset consists of a finite number (20 and 40 data sets) of
null terminated strings that are unsorted. The above table shows for 20 inputs the
memory size required by Bitcoin Merkle is approximately 6.33 MB and for Cognizant
Merkle it requires only 3.33 MB and the searching time of Bitcoin Merkle is 0.0024 s
whereas Cognizant Merkle requires only 0.0001 s. From this outcome, our research

Transaction ID SHA256 SHA256 Merkle Root

(TXID)

Fig. 6 Bitcoin Merkle Root

Transaction ID Fast bitwise

Base 64 Merkle Root
(TXID) hash

Fig. 7 Cognizant Merkle Root

[email protected]
Revamp Perception of Bitcoin Using Cognizant Merkle 149

Table 1 Comparison of Bitcoin Merkle with Cognizant Merkle

Factor/Tree type Bitcoin Merkle Cognizant Merkle
Total inputs 20 40 20 40
Virtual memory size (MB) 6.33242 6.33242 3.33 3.33
Insertion time (s) 0.000206 0.000403 0.000645 0.001078
Searching time (s) 0.00248 0.000865 0.000179 0.000230

shows that the development of merkle tree like this structure can reduce the cost of
searching time which yields rapid validation of transaction existence.

7 Conclusion

Currently, Bitcoin Merkle is a widely used data structure in blockchain for storing
and verifying the transaction survival in the database. Still, memory requirement of
Bitcoin Merkle is huge compared to the Cognizant Merkle and the depth of traversing
the tree for searching of the transaction existence is high in Bitcoin Merkle compared
to Cognizant Merkle model. The existing system of bitcoin is not affable for memory
management and genial support for fast verification of transaction existence. Con-
sequently the Cognizant Merkle, a variant of merkle tree that modifies the existence
property of merkle tree data structure to yield, fast, adaptable, cache sensible and
compressed data structure that maintains transaction id in memory which is stored
in sorted order. The proposed new model for Cognizant Merkle tree is improved
than 60% of existing Merkle tree searching. However, the proposed one provide
appropriate speed and space efficiency of hash tables which can be further improved
using cryptographic hashing without loss of transmission data across the peer-to-peer
network which also prevent the mishandling of data in future occurrence of bitcoin
transactions.

References

1. Vasek, M. (2015). The age of cryptocurrency. Science, 6241, 1308–1309.

2. Nakamoto, S. (2008). Bitcoin: A peer-to-peer electronic cash system. Journal of Cryptology.
3. Vijayalakshmi, J., & Murugan, A. (2017). Crypto coin overview of basic transaction. Interna-
tional Journal of Applied Research on Information Technology and Computing, 2, 113–120.
4. Antonopoulos, A. M. (2014). Mastering Bitcoin: Unlocking digital cryptocurrencies. O’Reilly
Media, Inc.
5. Reid, F., & Harrigan, M. (2013). An analysis of anonymity in the bitcoin system. In Security
and privacy in social networks (pp. 197–223). Berlin: Springer.
6. Oliveira S., Soares, F., Flach, G., Johann, M., & Reis, R. (2012). Building a bitcoin miner on
an FPGA. In South Symposium on Microelectronics.

[email protected]
150 J. Vijayalakshmi and A. Murugan

7. Decker, C., & Wattenhofer, R. (2013). Information propagation in the bitcoin network. In IEEE
Thirteenth International Conference Peer-to-Peer Computing (P2P) (pp. 1–10).
8. Investopedia Academy, https://www.investopedia.com/terms/m/Merkle-tree.asp.
9. Merkle, R. C. (1987). A digital signature based on a conventional encryption function. In
Conference on the theory and application of cryptographic techniques (pp. 369–378). Berlin:
Springer.
10. Hanandeh, F., Alsmadi, I., & Kwafha, M. M. (2014). Evaluating alternative structures for prefix
trees. In Proceedings of the World Congress on Engineering and Computer Science (Vol. 1).
11. Heinz, S., Justin, Z., & Williams, H. E. (2002). Burst-tries: a fast, efficient data structure for
string keys. ACM Transactions on Information Systems (TOIS), 2, 192–223.
12. Bazoobandi, H. R., de Rooij, S., Urbani, J., & Bal, H. (2015). A compact in-memory dictionary
for RDF data. In European Semantic Web Conference (pp. 205–220). Berlin: Springer.
13. Mavlyutov, R., Marcin, W., & Cudre-Mauroux, P. (2015). A comparison of data structures
to manage URIs on the web of data. In European Semantic Web Conference (pp. 137–151).
Berlin: Springer.
14. Knecht, M., Meier, W., & Nicola, C. U. (2014). A space-and time-efficient implementation of
the Merkle Tree Traversal Algorithm. arXiv, Vol. No. 1409.4081.
15. Askitis, N., & Sinha, R. (2007). HAT-trie: a cache-conscious trie-based data structure for
strings. In Proceedings of the thirtieth Australasian conference on Computer science (Vol. 62,
pp. 97–105).
16. Stack Overflow Questions https://stackoverflow.com/questions/4070693/what-is-the-purpose-
of-base-64-encoding-and-why-it-used-in-http-basic-authentica.
17. Ramakrishna, M. V., & Zobel, J. (1997). Performance in practice of string hashing functions.
In DASFAA (pp. 215–224).

[email protected]
A Novel Algorithm for DNA Sequence
Compression

K. Punitha and A. Murugan

Abstract Deoxyribonucleic Acid (DNA) sequences vary in terms of their size and
fall within the range of billions of nucleotides. The value increases to twice or thrice
its original value annually. Techniques of data compression and related methods that
originate from the information theory are frequently understood as relevant for the
field of data communication, exploration, and storage. In the present situation, it is
vital to store data for biological sequences. The compressBest algorithm proposed in
the paper for the compression of DNA sequences helps attain a better compression
ratio and is much faster when compared to the existing compression techniques.
compressBest algorithm is applicable to compression of DNA sequences with a
reduction in storage space. The proposed algorithm is tested over the data from the
UCI repository.

Keywords DNA sequences · Data compression · Deoxyribonucleic acid ·

Dynamic programming

1 Introduction

DNA databases are found and large in size [1, 2], where storage of them is a major
issue. They contain considerable complexity and feature logical organization. There-
fore, the data structure for storing, accessing, and processing this data in an efficient
manner is a challenging and problematic task [3, 4]. So, an efficient algorithm for
compression is required for storing huge data masses. The standard methods for
compression [5, 6] are not relevant when it comes to biological sequences [7] owing
to the subtle regularities in sequences of DNA [8, 9]. Biological sequences cannot

K. Punitha (B)
Department of Computer Science, Agurchand Manmull Jain College, Chennai, Tamil Nadu, India
e-mail: [email protected]
A. Murugan
Department of Computer Science, Dr. Ambedkar Govt. Arts College (Autonomous), Chennai,
Tamil Nadu, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 151

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_13

[email protected]
152 K. Punitha and A. Murugan

be compressed in a competent manner by standard algorithms used for compression.

The compressBest algorithm for DNA is introduced which is based on the precise
matching and gives the best results for comparison of the standard benchmarks in
the context of DNA sequences.
The four nucleotide bases are Cytosine, Adenine, Thymine, and Guanine, depicted
using their first character, which is C, A, T , and G. The meaning of compression is the
reduction of data size by way of modifying it to assume a format which necessitates
less number of bits than the ones required in the original format. Considerable effort
has been expanded for applying compression techniques on textual data for the
accomplishment of a number of tasks in computational biology, from storing and
indexing large sets of data to comparing sequences databases of DNA.
This paper proposed compressBest algorithm which can be employed for the
compression of DNA sequence data, which in turn results in a reduction of storage
space through the use of Dynamic programming. The mechanism which is proposed
for the compression of the DNA sequence through the use of the 2-bits encoding
method includes division into segments, exact matching, and the method of decom-
pression matching. When bases are distributed in a random manner in a sequence,
2-bit encoding method serves as an efficient method. However, nonrandomness exists
in an organism’s life and the DNA sequences which are evident in the living organ-
ism are nonrandom in nature. They also possess certain constraints. Our algorithm
resembles all other compression algorithms of the DNA sequence and is based on
creating partitions for a sequence into diverse segments, which include the repeat
segments which are also the copied segments, and the non-repeat segments.

2 Related Work

In [5], a two-stage algorithm suggested the combination of the features of the Huff-
man coding algorithm and Lempel–Ziv–Welch (LZW) algorithm. LZW is an algo-
rithm which is based on a dictionary. It is a lossless algorithm and is unified into
the standard pertaining to the consultative committee on telephony and international
telegraphy [10]. In this case, the code for every character is accessible in the dictio-
nary [11] which uses reduced bits when compared to the ASCII code (less than 5
bits).
A novel approach is proposed in [12], the compression of genomes which has
the ability to modify successions in the genome to double grouping which uses the
Genbit algorithm and the development of a grid takes place from the arrangement
which has been encoded. In case the encoding group length is not a perfect square,
a few bits are attached to the arrangement to create a flawless square with respect
to its length. When both sides record the same number of times, a ‘0’ is picked and
the bit is connected again at the end to make the length a flawless square for the
development of a framework.
Many algorithms have been recently introduced for the purpose which frequently
uses the recognition of lengthy and approximate repeats. Another algorithm is known

[email protected]
A Novel Algorithm for DNA Sequence Compression 153

as the DNAPack, which use dynamic programming as its basis. When compared to
the earlier programs, the DNA compression is marginally improved while its cost
can be ignored. A new lossless comparison algorithm is presented which enables
the vertical and horizontal compression of the data. It is based on the statistical
and substitution methods [13]. The representation of hiding data through the use of
compression of DNA sequences is attempted. In the initial step, data is converted to
a DNA sequence and the DNA sequence is compressed later on. The techniques of
four compressions are used in this case, where better compression is produced by one
of the options, which depends on the DNA sequence. The decompression algorithm
which is developed also enables the procurement of the original data [14].
A lossless compression algorithm which is seed-based is developed in the
paper to compress the DNA sequence which utilizes the method of substitu-
tion. This method has a similarity to the LempelZiv scheme of compression.
The repetition of structures is exploited in the method proposed here, which
are intrinsic to DNA sequences. The process is accomplished through the cre-
ation of an offline dictionary containing repeats accompanied by mismatch details.
When it is ensured that only promising mismatches are allowed [15], a com-
pression ratio is achieved by the method that is better than or in line with
the existing compression algorithms pertaining to lossless DNA compression
[16].
A DNA compression algorithm is introduced [17], which originated on the
basis of exact reverse matching which allows best results for compression for
the benchmarks of standard DNA sequences. When the DNA sequence is enor-
mously long, it is easy to search for exact reverses. The approximate reverses
which are optimal for the compression to take place can be found by the algo-
rithm, but the task takes considerable amount of time to complete. The time taken
is a quadratic time search or greater than that [18, 19]. Compression with high
speed and the best ratio can be achieved, but it is challenging to do so. The com-
pression for proposed DNA sequences achieves a better ratio for compression and
it is running faster when compared to the existing algorithms [20] for compres-
sion.

3 Proposed Work

This is a form of motivator for the advancement of compression tools with high
performance whose design is targeted at genomic data. The DNA compression
methods used earlier were either statistical or dictionary-based in nature. The 2-
bit encoding methods in use recently have become noticeable where the bases with
4-nucleotides in the DNA sequence {A, G, C, T } are assigned the values 00, 10,
01, and 11 respectively. The technique proposed here is used for the compression of
large DNA sequence bytes which have average ratio for compression. The decom-
pression time initially needs an input which is available at the time of sequence
compression. This value is used along with the compression input strings where the

[email protected]
154 K. Punitha and A. Murugan

Compression Decompression

DNA sequence
Compressed DNA
sequence
DNA sequence
segmentation

Decoding method
Encoding methods

DNA sequence data Original DNA

compressing Sequence

Compressed DNA
sequence

Fig. 1 Overall architecture of the proposed method

decompression takes place and the output is produced in the form of the original
sequence. The decompressed input is obtained from the original sequence as per
the segmented values. The DNA sequence algorithm for compression attains a com-
pression ratio which is moderate and runs considerably faster when compared to
present compression programs. The overall process of the proposed method is given
in Fig. 1.

3.1 Segmented in Dynamic Programming Techniques

Different symbols are in use with respect to the occurrences of different DNA
sequences which are chosen for the compression of the sequence. This means
that the DNA sequence data is divided into a number of parts throughout
the process. A technique used for solving the optimization of a sequence
of decisions is the dynamic programming technique. The underlying rationale
behind this is the representation of the problem by the way of a process
which is an evolution from one state to another in reaction to specific deci-
sions.

[email protected]
A Novel Algorithm for DNA Sequence Compression 155

3.2 Encoding of Numbers

The proposed algorithm needs the encoding of diverse integers. Let, A  00, C
 01, G  10, T  11 is the segmentation which is accompanied by a segmented
number. The process initiates with the alignment of each sequence in the dataset with
respect to the reference sequence through the use of the local sequence alignment.
The objective behind sequence alignment is the placement of homogenous segments
in the same column through the insertion of blank space. Further, aligning similar
sequences can assist in the discovery of patterns and sequence relationships, which
can improve the ratio for compression.
As an example, the segments are not fixed in length, and the encoding of their
length is important. For a repeated segment, the values which are segmented with
respect to the reference substring of the input are taken into consideration, and
the segment is copied from here, after which encoding takes place. In the case
of approximate copies, rather than exact ones, it is important to encode the value
which has been segmented with respect to the modifications. Any of these num-
bers do not feature any bounds, and the integers must be encoded in a way which
is self-delimited as opposed to encoding with respect to a fixed number of bits.
The reference segmented values are encoded by encoding the relative difference
with respect to the reference segmentation and it is preferred to make use of its
copy.
The encoding method which involves 2-bits can make use of 2 bits for the
encoding of every character which means that 00 is for A, 10 is for G, 11 is
for T , and 01 is for C. Therefore, “gaccgtca” can be encoded by “10 00 01
01 10 11 01 00”. This requires 16 bits in all. The exact matching method can
make use of repeat length and repeat position for the representation of an exact
repeat. In this way, three bits are used to encode an integer and two bits are
used to encode a character, and a single bit is used to indicate whether the
next part contains a pair, which also signifies an exact repeat or plain charac-
ter.

3.3 Compression

In experimental research work which involves the text file with a dot txt file extension,
a series of four successive base pairs are present, which are a, g, t, and c. This ends
in a blank space as a terminal character. The basic element, in this case, is the text
file which is used to consider compression as well as decompression. The output file
is also a text file containing information about four base pairs which are unmatched
and an ASCII character which has a coded value.

[email protected]
156 K. Punitha and A. Murugan

3.4 CompressBest Algorithm

Compression Algorithm
Input: DNA Sequence
Output: Compressed data

Input: Compressed text file

Output: original DNA Sequence

1: Repeat the following steps 2 to 3 until the codes are converted into bases.
2: According to the segmentation number, replace all the codes by the corresponding bases.
3: do
{
Read the character one by one from the DNA sequence;
If char = $ followed by the 2-bit based sequence base then
Assign 3 times the base
If char = # followed by the 2-bit based sequence base then
Assign 4 times the base
Else
Read two characters and
Assign A if char = 00;
Assign C if char = 01;
Assign G if char = 10;
Assign T if char = 11;

} until char = NULL;

4: The original DNA sequence is obtained.
5: End

3.5 Decompression

The input string used for compression of a particular value is decompressed to pro-
duce the original file in the form of the output. The lookup table is created from the
structure of the compressed file by finding repeat patterns present in the source. The
size of the lookup table is extracted from bit blocks which represent the pattern. The
blocks of the compressed pattern are extracted and the pattern type is recollected in
the form of pattern ID followed by patterns. Decompression takes place from the
beginning of the file to the end of the file for obtaining the original sequence of
DNA.

[email protected]
A Novel Algorithm for DNA Sequence Compression 157

Decompression Algorithm

Input: DNA Sequence

Output: Compressed data

1: Divide the sequence into segments with four nucleotides each in DNA sequences.
2: Find the matched segments in the DNA sequence and assign the numbers with symbol ‘S’ followed by corresponding numbers.
3: Repeat the following steps for each in the segment.
(i). If there is three repeat bases, then assign ‘$’ as a suffix and assign two bits based on DNA sequence values (A=00, C=01,
G=10, T=11).
(ii). If there is four repeat bases then assign ‘#’ as a suffix and assign two bites based on DNA sequence values
(A=00,C=01,G=10,T=11).
4: Repeat the steps from 2 to 3 until there are no repeat bases.
5: Assign two bits code for remaining bases.
6: End

The compression ratio can be calculated as follows:

compressed file size

Compression ratio:
original file size

4 Results and Discussions

The program code of the compressBest algorithm is executed Java 1.7.0 JDK and
tested on a dataset of DNA sequences which is from Genbank through UCI repository.
Genbank is a popular public database where the human genomic data is available. The
datasets include six DNA sequences such as two are HUMAN GENES [HUMDYS-
TROP, HUMHDABCD], two from chloroplast genomes (CHMPXX AND CHN-
TXX), and two are Viruses (HEHCMVCG AND VACCG).
The storage of the DNA sequence and its accuracy must be considered since even
a single mutation in the base, deletion, or insertion would reflect in terms of a major
change in its phenotype. For the purpose of accuracy, file compression is developed
one after the other through the use of the compressBest algorithm.
The efficient result level in terms of the performance analysis of the compression
ratio percentage of DNA sequence data is represented in Fig. 2 along with other
techniques such as Lossless compression algorithm and BDNAS compression algo-
rithm. The compressBest algorithm proposed here produces efficient output when
compared to the other techniques in existence.
The compression of the original sequence size before compression is depicted
in Table 1, and compression ratio after compression is also depicted in Table 1.
The advantages of the compressBest algorithm are, it is a high-speed algorithm
with minimized compression ratio, minimized execution time and it gives lossless
compression data.

[email protected]
158 K. Punitha and A. Murugan

28.6

compression ratio percentage(%)

28.4
28.2
28
27.8
27.6
27.4
27.2
27
BDNAS algorithm Lossless compression CompressBest
algorithm

Fig. 2 Comparison of compression techniques

Table 1 Experimental analysis

S. No. Type of DNA Original size Compression ratio percentage (%) after
sequences of sequence compression
before BDNAS Lossless CompressBest
compression algorithm compression algorithm
(Bytes)
1 HUMHDABCD 58,864 33.33 32.61 32.53
(Human Gene)
2 HUMDYSTROP 105,265 33.33 32.62 32.59
(Human Gene)
3 CHMPXX 121,024 4.18 4.02 3.66
(Chloroplast
Genome)
4 CHNTXX 155,844 33.33 32.42 32.35
(Chloroplast
Genome)
5 HEHCMVCG 229,354 33.33 32.58 32.14
(Virus Genome)
6 VACCG (Virus 47,912 33.33 32.08 31.86
Genome)

5 Conclusion

The experimental result shows that CompressBest algorithm gives better compression
ratio for mostly repeated sequences. It provides a reduction in file size without losses
of clear data. The results from the simulator help to achieve a minimization in the
time required for compression, high speed, efficiency, reduction in file size, and
accuracy with respect to the original file. The UCI repository was accessed to collect
the datasets and the Java environment is used for the development.

[email protected]
A Novel Algorithm for DNA Sequence Compression 159

References

1. International nucleotide sequence database collaboration, http://www.insdc.org. (2013).

2. Karsch-Mizrachi, I., Nakamura, Y., & Cochrane, G. (2012). The international nucleotide
sequence database collaboration. Nucleic Acids Research, 40(1), 33–37.
3. Deorowicz, S., & Grabowski, S. (2011). Robust relative compression of genomes with random
access. Bioinformatics, 27(21), 2979–2986.
4. Brooksbank, C., Cameron, G., & Thornton, J. (2010). The European Bioinformatics Institute‘s
data resources. Nucleic Acids Research, 38, 17–25.
5. Shumway, M., Cochrane, G., & Sugawara, H. (2010). Archiving next generation sequencing
data. Nucleic Acids Research, 38, 870–871.
6. Kapushesky, M., Emam, I., & Holloway, E. (2010). Gene expression atlas at the European
bioinformatics institute. Nucleic Acids Research, 38(1), 690–698.
7. Ahmed, A., Hisham, G., & Moustafa, G. (2010). EGEPT: Monitoring Middle East genomic
data. In: Proceedings of 5th Cairo International Biomedical Engineering Conference
(pp. 133–137). Egypt.
8. Korodi, G., Tabus, I., & Rissanen, J. (2007). DNA sequence compression based on the normal-
ized maximum likelihood model. Signal Processing Magazine, IEEE, 24(1), 47–53.
9. Deepak, H. (2013). State of the art: DNA compression algorithms. International Journal of
Advanced Research in Computer Science and Software Engineering, 397–400.
10. Kaur, S., & Verma, V. S. (2012). Design and Implementation of Lzw data compression algo-
rithm. International Journal of Information Sciences and Techniques (IJIST), 2(4). https://doi.
org/10.5121/ijist.2012.240771.
11. Shravan Kulkarni, S., & Kini, Y. (2017). Pre equal architecture for lossless data compression
and decompression using hybrid algorithm. International Journal of Advance Electrical and
Electronics Engineering (IJAEEE), 6(1_2). ISSN (Print): 2278-8948
12. Bhukya, R., Viswanath, B. V., Mahendra Kumar, D., Swathi Kiran, D. S., & Bagdia, P. (2017).
DNA sequence decompression using bitmap matrix & wavelet transformation in image pro-
cessing. Received: 20th April 2017, Accepted: 28th April 2017, Published: 1st May 2017,
Copyright © 2016 Helix ISSN 2319– 5592 (Online)- Helix Vol. 8: 1491–1497.
13. Mishra, K. N., & Aaggarwal, A. (2010). An efficient horizontal and vertical method for
online DNA sequence compression. International Journal of Computer Applications, 3(1),
0975–8887.
14. Bandyopadhya, S. K., & Chakraborty, S. (2011). Data hiding using DNA sequence compres-
sion. Journal of Global Research in Computer Science, 2 (1), Vol. 27–33. © JGRCS 2010, All
Rights Reserved.
15. Cover, T. M., & Thomas, J. A. (2012). Elements of information theory. New York: Wiley.
16. Eric, P. V., Gopalakrishnan, G., & Karunakaran, M. (2015). An optimal seed based compression
algorithm for DNA sequences. Correspondence should be addressed to Pamela Vinitha Eric;
[email protected] Received 28 November 2015; Revised 9 May 2016; Accepted 19
June 2016.
17. Mukherjee, R., Mandal, S., & Mandal, B. (2016). Reverse sequencing based genome sequence
using lossless compression algorithm. International Research Journal of Engineering and
Technology (IRJET), 3(5). e-ISSN: 2395-0056. www.irjet.net p-ISSN: 2395-0072.
18. Jacob, G., & Murugan, A. (2013). DNA based cryptography an overview & analysis. Interna-
tional Journal of emerging science, 3(1), 36–42.
19. Chouhan, D. S., & Mahajan, R. P. An architectural framework for encryption and generation
of digital signature using DNA cryptography. In International Conference on Computing for
Sustainable Global Development (INDIACom).
20. Jahaan, A., Ravi, T. N., & Panneer Arokiaraj, S. (2017). Bit DNA squeezer (BDNAS): A unique
techniques for data compression. International Journal of Scientific Research in Computer
Science, Engineering and Information Technology ©, 2 (4). ISSN: 2456-3307.

[email protected]
DigiPen: An Intelligent Pen Using
Accelerometer for Character Recognition

Ankur Agarwal and Swaroop Sudhanva Belur

Abstract The idea is to identify the character being written by a user on a piece
of paper by observing the acceleration of the pen while writing these characters
and passing this acceleration data to an Artificial Intelligence (AI) algorithm. Using
an off-the-shelf accelerometer paired with a microcontroller using Inter-Integrated
Circuit (i2c) protocol, we get sequential acceleration values in three directions. Upon
observing this with this data, we noticed the same patterns when writing a specific
character and decided to use efficient Long Short-Term Memory (LSTM) cells to
recognize the patterns. After collecting some data for training the neural network
and doing some preprocessing of the data, we used Basic LSTM cells in the models
to recognize the patterns from the sequences of these values. LSTM cells were
preferred over regular Recurrent Neural Network cells (RNN) due to LSTMs ability
to remember longer sequences. Multiple LSTM models with a different number of
layers and sizes with different activation functions and dropout values were trained
and tested for performance and we were able to achieve a test accuracy of 47% on
a fairly small dataset which far exceeds the 10% accuracy benchmark which would
have been simple guesswork. With some more optimization of the hyperparameters
of the neural network and training with a larger dataset, we believe better performance
can be achieved. For the purpose of this paper, we have used numerical digits (0–9)
as the characters to be classified.

Keywords Smart pen · Artificial intelligence · LSTM · IoT · Cloud · Supervised

learning

A. Agarwal (B)
Department of CSE, SRM Institute of Science and Technology, Kattankulathur
603203, Tamil Nadu, India
e-mail: [email protected]
S. S. Belur
Department of ECE, SRM Institute of Science and Technology, Kattankulathur
603203, Tamil Nadu, India

© Springer Nature Singapore Pte Ltd. 2019 161

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_14

[email protected]
162 A. Agarwal and S. S. Belur

1 Introduction

Despite the technological advances in the twenty-first century, the act of taking notes
is still relevant. This explains the plethora of note-taking applications on phones
and computers. Current techniques of digital note-taking are sub-optimal for reasons
which include but are not limited to—small touch keyboard interfaces, heavy and
uncomfortable devices, etc. We want to bring back the usage of notebooks and thus
conceived of a device that uses twenty-first century technology but has the feel of
a regular notebook [1]. This is how we arrived at the DigiPen. When trained with a
proper amount of data, this pen was able to accurately detect the characters being
written without the need for any other special type of book or pen. Using the analytical
power of AI [2], it can even understand characters of different alphabets and symbols.
The days are not far off when professionals like doctors writing prescriptions and
students writing notes can refer to and search within the digital copy instead of using
their regular books [3]. This brings out the real power of the digital revolution to
something as banal as note-taking.

2 Similar Technologies

Although devices that have a similar functionality exist, none of them use the same
technology as DigiPen and usually require extra peripherals like a special book or the
use of a mobile or a computer unlike DigiPen. Multiple products have come up with
a pen that uses a dot pattern on an accompanying book which is used to recognize
the characters being written using the location of the pen’s tip to trace out what
has been written. Moreover, these products require a mobile phone to work. This
is innovative technology but the problem is that it requires the purchasing of a new
book every time it runs out of pages. However, the advantage these products have
over the DigiPen is that they are much easier to implement due to the availability
of datasets like MNIST. In case of the DigiPen, it requires a completely different
type of dataset which uses accelerometer values. Due to the same reason, DigiPen
has a relatively lower accuracy than the others. However, with the addition of more
data into the training dataset, comparable or even better accuracy can be achieved
by DigiPen.

3 The Analogy and Inside the Accelerometer

The accelerometer [4] contains a surface machined polysilicon structure which is

positioned on top of a silicon wafer. The polysilicon structure is suspended using
springs and provides resistance against acceleration forces. Differential capacitors,
consisting of independent fixed plates, are used to measure deflection. The deflection

[email protected]
DigiPen: An Intelligent Pen Using … 163

Fig. 1 (a) The accelerometer, (b) The pen [5]

of the beam causes an unbalanced capacitor output which, in turn, results in a sensor
output whose amplitude is proportional to the acceleration. Phase-sensitive demod-
ulation is then applied to determine the polarity and magnitude of the acceleration.
Simply put, there is a polysilicon ball placed inside a hollow cube whose surfaces
correspond to the differential capacitors. The movement of the ball, due to acceler-
ative forces, causes its movement inside the cube, which results in an unbalanced
capacitor output.
We use these acceleration values as our parameter for recognizing the written
character. Since the accelerometer is placed on the flat surface of the pen, both the
pen and accelerometer become parallel to each other. Now as they are parallel and
connected, the accelerating forces applied on the pen while writing are also applied
on the accelerometer. The lines of accelerating forces on both the objects at the
point of connections have the same magnitude and direction. Due to this reason, the
accelerometer is able to capture the true forces of accelerations acting on the pen. As
the gripping point is the pivot point during writing, the accelerometer is attached at a
location from where maximum motion can be transferred on to the sensor effectively
(Fig. 1a).

4 Data Collection

The contraption that was built is an accelerometer attached near the writing tip of a
regular pen and the accelerometer data is collected using an microcontroller using the
inter-integrated circuit (i2c) communication protocol. The arduino microcontroller
is used in conjunction with an open source software called Processing [6] to take the
serially printed data into a comma separated variable (CSV) file format. We obtain x,
y, and z coordinate accelerometer data separately. These constitute the three comma
separated variables which are fed into a neural network. The Microcontroller used
is an ATMega 328P which is programmed and used by means of the Arduino Uno
development board which has general purpose input output (GPIO) pins, an oscilla-
tor crystal, a USB connection, support for i2c and SPI (serial peripheral interface)

[email protected]
164 A. Agarwal and S. S. Belur

Fig. 2 Similar acceleration patterns for a given character written

communication protocols and has an accompanying integrated development envi-

ronment for programming and flashing. The Arduino is programmed to collect data
for a short amount of time which we believe is sufficiently long enough to write a
character. After some experimentation, the time we arrived at was 1200 ms. When
the user wants to write a character, he/she has to just double tap the pen which will
trigger the accelerometer and it will start collecting data for a period of 1200 ms
(Fig. 2).
Due to resource constraints, we were only able to collect a limited amount of data
to train the LSTM. We have 200 data points for each numerical digit from 0 through
9. This adds up to 2000 data points in total. Although this is not ideal for the proposed
LSTM model, we observed that the model does fairly well as discussed above and
far exceeds the 10% accuracy benchmark which would have been simple guesswork
by the LSTM. While collecting data, we tried to ensure that the size of the character
written is small and about the same size as that written in a regular notebook. It is
important because first, that is how we believe people will write with the digipen, but
more importantly, the acceleration forces when writing small are more pronounced
and are easier to patternize. Moreover, as mentioned, we used a constant amount of
time of 1200 ms to collect data for each datapoint.
Another aspect we considered while collecting the data is the orientation in which
the data was collected. We tried to take the readings in various different orientations
to ensure that DigiPen can predict whatever has been written even if it was held in
a different position. The dataset consists of comma separated variables (CSV) files
for each datapoint. Each data point consists of x, y, z followed by the data collected
by the accelerometer for 1200 ms corresponding to each of the three axes. Due to
small, insuppressible errors in the sensor, the number of values corresponding to x, y,
and z axes are slightly variable despite taking to same amount of time to record. We
found that they range from 75 to 85. To standardize the data going into the LSTM,
we employ zero-padding. This ensures that the LSTM receives a uniform number of
values for each datapoint. We used an 80:20 split for training and test data for the
LSTM.

[email protected]
DigiPen: An Intelligent Pen Using … 165

Fig. 3 Preparing the dataset

5 Preparing the Dataset

The input data is a sequence of acceleration values in the x, y, and z axes. Each
sampling is completed within t secs and each sample recording gives three column
time series sequences of acceleration in each axis direction. Hence, each sample
recording is of shape (t, 3). Each sample recording is saved as comma separated
value sequence. Initially while creating the model input dataset, each file is read
as a single matrix having 3 columns (x, y, z) & ‘t’ rows and the complete dataset
becomes a 3D matrix in which many such 2D matrices are stacked. The label y is
taken from the filename, i.e., the digit it stores. The label is further encoded into one
hot encoding of 10 classes, i.e., for digits between 0 and 9. The dataset is normalized
and shuffled randomly. The dataset is split into training and test dataset in the ratio
80:20. The dataset is split into three 2D matrices each representing the sequential
values of the samples in different axis directions. This data is given as input to the
neural network for further classification model (Fig. 3).

6 The Different Models

The input values to the neural network model are time series based. When writing a
digit there is a particular pattern which cannot be skipped no matter which font is used
as the digit is predefined. Example, the digit 1 will always have a straight line, the
different fonts will always have this straight line as the major part. The accelerometer
catches this pattern in terms of accelerations in different directions and the pattern of
this series is important to the model. The output is to be given as a label from given
classes of 10 digits. This clearly becomes a sequential and classification problem.
The different models used have RNN layer with LSTM cells [7] common in all to
find out the important sequential patterns for a longer time. The sequential patterns
from each direction (i.e., x, y, and z) are important, and so LSTM cells are used. The
following three different models have been designed and implemented (Fig. 4).

[email protected]
166 A. Agarwal and S. S. Belur

Fig. 4 The different model designs

6.1 Model M1

A basic simple model is created. The input of sequence of values in x, y, and z

directions are directly given as input to the channelized LSTM cells. The outputs
from the LSTM cells are sent to a softmax layer of 10 classes which gives the output
of the model in the form of one hot encoding.

6.2 Model M2

Using the similar design as that of model M1, model M2 also has three LSTM cells
channelized with their inputs as the input to the model. The next layer has seven
neurons, all together having every possible input combination from the output of the
three LSTM cells. This layer is added with the intuition that all possible combination
of probabilities from the LSTM outputs will be generated, which might become
valuable to help find out the classification. Each of the seven neuron units has the
same size as that of the single output of a LSTM cell. The outputs from these neuron
units are sent to a layer having 100 neuron units which is then connected to a reduced
10 neuron unit softmax layer. This softmax layer gives the output of the model in the
form of one hot encodings.

6.3 Model M3

This model is similar to the previous model but with a more intuitive design. In
this model, the sequential inputs x, y, and z are sent to a layer of seven neurons, all
together having all possible input combinations of x, y, and z. Here, all the possible
combination of the sequences in different directions are generated. The size of each

[email protected]
DigiPen: An Intelligent Pen Using … 167

of these seven neuron unit is same as that of a single direction sequence input to
the model. The output of these seven neuron units is sent to a layer having seven
channels of LSTM cells. Passing all the possible combinations of sequences through
LSTM cells might help find more hidden sequences. The output from these seven
LSTM cells is sent to a layer having 100 neuron units which is then connected to
a reduced 10 neuron unit softmax layer. This softmax layer gives the output of the
model in one hot encoding form.

6.4 Hyperparameters, Regularization, Optimizer

and Activation Function

In the above three models, each LSTM cell has 32 units as the dataset is small. Also
in above three models, a dropout layer is attached at the end of each LSTM cell. The
keep probability of LSTM dropout layer having z direction as an element is kept
lower than others as the data samples where created with characters written on a
sheet of paper (viz. two-dimensional) with very less variation in the z direction. In
M1 and M2, dropout layers for LSTM cells L1 and L2 have a keep probability of 0.75
and for L3 it is 0.65. In model M3, LSTM cells L1 & L2 have keep probability of
0.75, L3 has 0.65, L4 and L7 have 0.74 and L5 and L6 have 0.70 as keep probability.
Apart from the dropout layers, exponentially decaying learning rate is used to help
the models converge gradually with a lower learning rate near the convergence point.
ELU activation function was used to take care of the vanishing gradient problem.
For optimizing the backpropagation cost of training, the Adam optimizer was used
(Fig. 5).

(a) With Elu activation (b) Without Elu activation

Fig. 5 Purpose of using Elu as activation

[email protected]
168 A. Agarwal and S. S. Belur

7 Performance
The models were trained after tuning the hyperparameters. A few hyperparameters
were set with common values so as to judge the three models. The values were set
keeping in mind the small size of the dataset available. The few common parameters
and hyperparameters which were set with same values are
– learning rate  0.001
– decay rate  0.05
– decay step  5
– mini-batch size  96
– iterations  1000
– number of units in each LSTM cells  32
– common random seed (Table 1).

Table 1 The performance of the three models

Model Training accuracy Test accuracy
M1 64.93 34.25
M2 56.35 32.75
M3 74.76 42.25

(a) Model 1 (b) Model 2

(c) Model 3

Fig. 6 Performance of the three models

[email protected]
DigiPen: An Intelligent Pen Using … 169

The model M3 seems to work best compared to the others. The model was run for
higher number of iterations to get better test accuracy results (Figs. 6 and 7; Table 2).
It is clearly visible from the confusion matrix that the resulting model has a good
precision and recall as the frequency of the actual digits being predicted is far greater
than the frequency of any other wrong digit being predicted.

8 Future Development

The future development from this concept is the addition of various features by
integrating with various office suites using the internet of things (IoT). This will be

Fig. 7 Normalized confusion matrix (Test set)

Table 2 Performance of Model Iterations Training Test accuracy

Model 3 with different accuracy
iterations
M3 1000 74.76 42.25
M3 2500 96.43 43.00
M3 3000 96.64 47.00

[email protected]
170 A. Agarwal and S. S. Belur

especially handy when the user wants to use the device completely independent of a
mobile phone or a computer. We could send raw sensor data, preferably encrypted,
to the cloud where a trained LSTM network will be deployed. The output generated
can be then be saved as a text document, which can then be shared. As mentioned,
this results in DigiPen being able to be used as a standalone device.

9 Conclusion

We have found an innovative way to recognize characters from the fundamental basis
of the method of writing, which is the movement of the pen and data collection is
very simple. We have found how to process the data in order to classify them into
characters using an intuitive architecture model and achieve a test accuracy of 47%.

Acknowledgements The authors would like to thank Next Tech Lab and SRM Institute of Science
and Technology for their continued guidance and support. We would also like to thank Nvidia for
their generous grant of an Nvidia Titan X which was vital for us to be able to test the various models.

References

1. Reynaerts, D., & Van Brussel, H. (1995). Design of an advanced computer writing tool. In
Proceedings of the Sixth International Symposium on Micro Machine and Human Science
(pp. 229–234). MHS ’95. Nagoya.
2. Wang, J. S., & Chuang, F. C. (2012). An accelerometer-based digital pen with a trajectory recog-
nition algorithm for handwritten digit and gesture recognition. IEEE Transactions on Industrial
Electronics, 59(7), 2998–3007.
3. Toyozumi, N., Takahashi, J., & Lopez, G. (2015). Development of pen based interface system
for digital operation. In 2015 IEEE/SICE International Symposium on System Integration (SII),
Nagoya (pp. 593–598).
4. Accelerometer working https://www.sparkfun.com/datasheet/Sensors/accelerometer/
ADXL345.pdf.
5. Accelerometer Image (Fig. 1a) http://www.starlino.com/imu_guide.html.
6. Processing https://processing.org/.
7. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computing 9(8),
1735–1780.

[email protected]
Design of FPGA-Based Radar and Beam
Controller

Adesh Panwar and Neha Goyal

Abstract Radar controller is vital component in a Radar system. It enables transmis-

sion of RF energy, generation of complex waveform schemes and precision timing to
gate the receiver. The paper presents architecture for FPGA-based Radar controller. It
includes beam controller to schedule dwell requests among all tasks. Proposed design
is robust, flexible, and uses VHDL-based modules to promote reuse and cost opti-
mization. The paper depicts detailed hardware and software methodology adopted
during the development.

Keywords FPGA · Controller · Phase gradient

1 Introduction

Radar system provides situation awareness to the operator. It assists in understanding

the nature, action and intent of the “targets of interest”. It provides such information
in a complex environment, in all weather conditions, day and night. In such systems,
electromagnetic energy is transmitted in the form of a train of pulses and the reflected
energy in its direction enables the detection and measurement of the target position.
Such a system has an antenna which concentrates the transmitted energy in a preferred
direction and also intercepts and captures the received echoes [1].
Traditionally, antennas are classified into two major categories, mechanically
steered and electronically steered antenna. The electronic scanning uses a group of
radiating elements wherein the relative phase of the respective signals is controlled
in such a way that the effective radiation pattern of the array is reinforced in the
desired direction [2]. Phased array Radar provides significant advantage such as
• Solid state technology
• Distributed architecture
• Modular design

A. Panwar (B) · N. Goyal

Bharat Electronics Limited, Bangalore 560013, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 171

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_15

[email protected]
172 A. Panwar and N. Goyal

• Beam agility
• Improved reliability.
This paper presents design of FPGA-based Radar and beam controller used in the
phased array. Section 2 introduces a typical architecture of the Radar system. The
proposed system is presented in Sect. 3. Section 4 gives details on features incorpo-
rated in the design. Sections 5 and 6 gives test results and conclusion respectively.

2 Background

The system block diagram is shown in Fig. 1. Operator selects the desired volume
to scan and communicates it to the Radar and Beam Controller (RBC) via Operator
Console (OC). RBC sets the timing parameters and provides the required waveforms.
It plans ahead the future beam position and communicates the phase values to Phase
Gradient Controller (PGC) unit.
PGC in synchronization with RBC sets Transmit Receive (TR) modules to transmit
in a specific direction upon which target-acquiring microwave beams are transmit-
ted from the antenna, which also collects the reflection. The echo radio frequency
(RF) signal is multiplied with the local oscillator (LO) signal and translated to an
intermediate frequency (IF).
The received signal is digitized and discrete values are provided to the FPGA
chip for digital pulse compression processing. It de-chirps and digital orthogonal
transformation is performed to form in-phase and quadrature signals. Each such pair
of samples is associated with a “numerical index” that corresponds to a specific time
offset from the beginning of the Radar pulse, so each of them represents the energy
reflected from a specific range.

Fig. 1 System block diagram

[email protected]
Design of FPGA-Based Radar and Beam Controller 173

The processed information is presented to the operator in the form of a bright spot
on the display along with the target characteristics.

3 Proposed System

RBC acts as a focal point for all the communications in the system. It initiates
commands to RF modules, SP module, and PGC. It provides synchronization among
subunits. Design of such a system is complex and requires deep understanding of
hardware, algorithms, data processing, and the operating environment.
It requires complexity of a CPU, memory controller, Serial Peripheral Interface
(SPI) bus interface controllers, Serial interface and Local Area Network (LAN)
module. It requires computations that can be pipelined.
To accomplish the requirements of RBC, FPGA based hardware with multi-layer
software architecture is proposed and is shown in Fig. 2. Starting at the basic level,
hardware layer provides the physical interface between the FPGA and the external
world. Programmable logics are built using VHDL on FPGA. Functionalities are
divided into number of small, well defined modules to improve code maintenance.

Fig. 2 FPGA-based
multi-layer software
architecture

[email protected]
174 A. Panwar and N. Goyal

Soft Processor IP core is interfaced with VHDL as a component. An abstraction

layer provides high level programming interface between the processor and applica-
tion specific functions.
Distinct advantages of proposed system are
• Higher flexibility and re-programmability
• Capability for parallel processing
• Small design cycle
• Lower design cost.

4 Design Methodology

4.1 Hardware Design

The proposed design is built around Stratix [3] series FPGA. It has power supply and
conditioning circuitry. As shown in Fig. 3, it has peripheral memory interface for
SRAM, SDRAM, Flash storage, and EEPROM. It contains multiple trans-receivers
and a PHY device. It uses SPI bus [4] to interface RBC over Low Voltage Differential
Signaling (LVDS) voltage levels [5].

Fig. 3 Hardware design

[email protected]
Design of FPGA-Based Radar and Beam Controller 175

4.2 Software Design

The proposed design has multiple software modules, they are as follows.

4.2.1 Phase Computation Module

Considering an array of (m, n) elements, the phase of the individual element is

represented as follows. Here (m, n) are the indices of the individual antenna channel,
PA is the phase gradient in the azimuth and PB is the phase gradient in the elevation.

Phase (m, n)  m ∗ PA + n ∗ PB (1)

The phase gradients are represented as follows. Here “az” is steering angle
azimuth, “el” is steering angle elevation, “λ” is wavelength of the EM wave being
used, “dh” is the element spacing in azimuth and “dv” is element spacing in elevation.

PA  (2π/λ) ∗ dh ∗ Sin(az) ∗ Cos(el) (2)

PB  (2π/λ) ∗ dv ∗ Sin(az) ∗ Sin(el) (3)

Once the operator selects the volume to be scanned, RBC communicates the
common phase gradient data for all the radiating elements over SPI bus. PGC receives
this data, computes the position-wise phase values for each element, caters for the
calibration error, and applies it to the individual elements on the panel.
SPI master module is located in RBC. It initiates and controls the entire commu-
nication. It derives PA values for a given azimuth from the beam control module.

4.2.2 Beam Control Module

System using electronic beam steering control can carry out the concurrent functions.
A proper scheduling algorithm is to be devised to multiplex the requests of each
function for transmission of energy. Such scheduler must be able to execute priority
functions on the expense of the less critical ones. One such scheduler is provided in
RBC and is interfaced with Radar Data Processor (RDP) over LAN interface.
As shown in Fig. 4, once the dwell requests are submitted to the scheduler, they
are maintained in one of the dictated queue. Each dwell request consists of an iden-
tification number, priority, duration, and the desired time for execution.
Each queue contains requests corresponding to the similar function. Within a
queue, dwell requests are ordered according to the desired time of transmission.
Dwell request is played within the specified time period only if any higher priority
request is not pending for the same specified period. Expired requests are unlinked
from the queue.

[email protected]
176 A. Panwar and N. Goyal

Fig. 4 Beam control module

Different type of dwell corresponds to specific waveform parameters which

include PRF, carrier frequency, etc., and are synchronously executed in all the sub-
systems via waveform and timing generation module.

4.2.3 Waveform and Timing Generation Module

Waveform and timing generation module (WTGM) provides all the timing and syn-
chronization signals in a Radar system. Timing signals like dwell, PRT (Pulse Rep-
etition Time) and cover pulse are generated and interfaced to subsystem over system
clock. During initiated Built in Test (BIT), it generates “Test Pattern” control signals
to RF subsystem and exciter unit for fault diagnostics. It facilitates the generation of
simulated targets at the input of the receiver. Such “Test Pattern” provides a com-
plete check of the total receiver chain including the exciter, down converter, SP, plot
extractor, RBC and display unit.

4.2.4 Sensitivity Time Control

The power level at the input of a Radar receiver is inversely proportional to the fourth
power of the target range. In some cases, echo signals from near range clutter saturates
the receiver and small changes in signal strength go undetected. Here, increasing the
threshold levels substantially, or reducing the receiver gain, could permit detection of
such near range targets, but the distant range targets would go undetected. Therefore
it is required to attenuate the large target returns occurring close to the trailing edge
of the transmitted pulse and gradually reduced according to the system requirement.
Such attenuation function is called as Sensitivity Time Control (STC).

[email protected]
Design of FPGA-Based Radar and Beam Controller 177

Fig. 5 LAN interface module

The control signals are synchronized with the Pulse Repetition Frequency (PRF).
Each time a pulse is transmitted, the controls set the attenuation to maximum. It
gradually reduces until a new pulse is transmitted.

4.2.5 LAN Interface Module

In the present design, an application tailored subset of UDP/IP stack is implemented

for a cost effective connection to a network. As shown in Fig. 5, it consists of four
main modules such as: receiver, packet interpreter, transmitter, and packet composer.
Both the RAMs are implemented as dual port FPGA RAM.
The receiver module manages incoming packets. Once a new packet is detected,
it is saved in one of the RX RAM bank. Each byte is sent to CRC (Cyclic Redun-
dancy Check) checker, which progressively calculates the checksum. The packets
are checked for its validity, headers, MAC address, IP address, CRC, message type,
etc., in the packet interpreter module. Receive state machine is shown in Fig. 6.
The packet sender will act after initiation from one of the packet type’s viz.
User Datagram Protocol (UDP), Address Resolution Protocol (ARP), Reversed ARP
(RARP) or Internet Control Message Protocol (ICMP). The required data is written

[email protected]
178 A. Panwar and N. Goyal

Fig. 6 Receive state machine

on one of the TX RAM bank. The transmitter module reads this data and puts out the
packet to the PHY data bus and sets control signals. CRC generator progressively
calculates check bits and is sent as the last byte. Transmit state machine is shown in
Fig. 7.

5 Test Results

To ensure the timing requirements in the proposed design, verification and validation
is performed using timing analysis tool and signal tap logic analyzer respectively.
The data exchange between Phase Computation Module in RBC and PGC over
SPI is shown in Fig. 8. Signal “SCK_RC_IN” “SDI_RC_OUT” and “SDO_RC_IN”
represents master clock, data from and data to RBC respectively. The communication
is without glitches and errors.
The trace outputs are enabled in RBC for the scheduled reports. Different priority
requests are scheduled and there timings are shown in Table 1. “Desired Time” rep-
resents the expected time for execution for the respective beam. Similarly, “Stamp”
is current RBC time, “Scheduled” is the time at which RBC executed the request and
“Delta” represents the difference between “Desired” and “Scheduled” time.

[email protected]
Design of FPGA-Based Radar and Beam Controller 179

Fig. 7 Transmit state machine

Fig. 8 Data exchange between RBC and PGC

Table 1 Timing results

ID Priority| Time (in s)
Length Absolute
Duration Stamp Desired Scheduled Delta
1 1 67 18,437 18,504 18,449 55
1 2 218 18,473 18,691 18,609 82
1 2 219 18,633 18,852 18,769 83
1 2 220 18,793 19,013 18,929 84
1 5 6022 18,953 24,975 24,887 88

[email protected]
180 A. Panwar and N. Goyal

Fig. 9 Doppler in BIT mode

Fig. 10 Data exchange with LAN transmit module

As shown in Fig. 9, Doppler is simulated in BIT mode using WTGM. It represents

the IF signal on channels and there phase difference. The target profile represents
Doppler of PRF/2.
As shown in Fig. 10, signal “tx_clk” and “stage_reg” represents the input clock and
state of LAN transmit module. During transmit initiation, signal “tx_mem_bank_sel”
is changed and “tx_mem_bank_reg” is set to zero. Number of bytes are checked and
based on “max_byte_cnt1_reg” value zero bytes are padded. Signal “gig_en” is
enabled for byte-wise transmission. Signal “byte_reg” holds transmit byte count and
once it reaches to its maximum value CRC byte transmission is enabled.
Received data is shown in Fig. 11. Signal “rx_dv” goes high and initiates data
reception. Signal “rx_data_reg” indicates the input data and is checked for preamble.
If “gig_en” is set then data is received byte wise. Each byte is stored in Rx memory

[email protected]
Design of FPGA-Based Radar and Beam Controller 181

Fig. 11 Data exchange with LAN receive module

bank and “wr_address” is incremented. Once “rx_dv” is low checksum is matched,

memory address is locked and next memory bank is selected.

6 Conclusion

The work presents FPGA-based Radar and beam controller for surveillance sys-
tem. The key modules of the design are discussed. The test results suggest that the
design meets the required functionality and the timing parameters are well within
specifications. Use of programmable device offers flexibility with high performance.

Acknowledgements Authors are very much thankful to General Manager of Military Radar SBU
in Bharat Electronics Ltd., Bangalore for giving an opportunity to publish this technical paper.
Thanks to dear colleagues whose kind support made this study and design possible.

References

1. Kraus, J. D. (1997). Antennas. New York: Tata McGraw-Hill.

2. Stimson, G. W. Introduction to airborne radar. Scitech Publishers.
3. Stratix Device Handbook, www.altera.com.
4. Chen, D. (2016). Introduction to SPI interface.
5. An Overview of LVDS Technology, www.ti.com.

[email protected]
Effect of Lattice Topologies and Distance
Measurements in Self-Organizing Map
for Better Classification

Sathiapriya Ramiah

Abstract Self-Organizing Map (SOM) is a widely used algorithm in artificial neural

network for classification. Despite the general success of this algorithm, there are
several limitations which some of them are poor classification accuracy and slow
rates of convergence when the standard lattice topology and distance measurement
are implemented. This paper investigates the performance of SOM using different
topologies and different distance measurements. The results obtained showed that
SOM with hexagonal topology and Euclidean distance measurement outperforms
other topologies and distance measurement using at any scale datasets.

Keywords Self-organizing map · Best matching unit · Topology · Distance

measurement · Accuracy

1 Introduction

Classification can be defined as assigning objects to groups on the basic criteria made
on the objects into several classes. It is mainly used to understand the existing data
pattern and predict how new instances will behave. Classification plays an important
role in many fields including medical, education, business, marketing, shipping,
transportation, and others.
Self-Organizing Map (SOM) also known as Kohonen SOM is a neural network
algorithm invented by the founder of the Neural Networks Research Centre, Profes-
sor Kohonen [1]. Similar to most artificial neural networks, SOM operates in two
modes namely training and testing. Training is a competitive process called vector
quantization to build the map using input data samples. Testing process automatically
classifies a new input sample into respective classes [2].

S. Ramiah (B)
Asia Pacific University of Technology & Innovation, Kuala Lumpur, Malaysia
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 183

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_16

[email protected]
184 S. Ramiah

SOM is called a topology preserving map because of the mapping of the nodes
from high dimensional space into low dimensional, usually two-dimensional lattice
[3]. Lattice topology specifies the arrangement of nodes on the map forming the layer.
Therefore, in this paper, three types of topologies, mainly triangular, rectangular, and
hexagonal are evaluated.
There are several ways to work out the distance between two points in multidimen-
sional space for classification. Changing the distance measure can have a major effect
on the overall performance of the classification system. The distances are measured
using Euclidean, Manhattan, Chebychev, and Canberra in this experiment.
The rest of the paper is organized as follows. In Sect. 2, experimental dataset used
for evaluation are discussed. SOM is described in detailed in Sect. 3. Sections 4 and 5
illustrate lattice topology and distance measurement. Whereby experimental results
discussed in Sect. 6. Finally Sect. 7 state the concluding remarks.

2 Experimental Data

Dataset used in this experiment is obtained from UCI machine learning repository to
evaluate the performance of SOM in classification [4]. Iris data classifies it species,
Breast Cancer Wisconsin (Diagnostic) data predict whether the cancer is benign or
malignant and Car Evaluation data evaluates acceptability of a car. Iris data represents
small-scale dataset. It contains data of 105 instances for training and 45 instances for
testing. Meanwhile, Breast Cancer Wisconsin (Diagnostic) data used to represents
medium scale dataset. 400 instances used for training and 169 instances were used for
testing purpose. Finally, Car Evaluation representing large-scale dataset. It contains
total of 1200 instances used in training and 528 instances for testing.

3 Self-organizing Map (SOM)

SOM is two layer network comprises an input layer and an output layer of neurons
in a two-dimensional lattice as is shown in Fig. 1. Each node in the network has a
specific coordinate and contains weight vectors of the same dimension as the input
vectors [5]. Each neuron in input layer is connected to all the neurons of the output
via the weight. There are main two steps: training and testing. The network needs to
be trained using input and correct output so that it can learn the pattern of output to
be produced. Once the training is done, network needs to be tested using real sample
data.

[email protected]
Effect of Lattice Topologies and Distance Measurements … 185

Fig. 1 Typical architecture of SOM. Source Lohninger [8]

A few parameters need to be specified in order to develop SOM are defined below
i. Number of input nodes
The number of input determines the number of data to be fed into the network for
training and testing.
ii. Map size
Map size determines the size of the output layer.
iii. Number of neurons
The number of neurons on the output layer is determined by the map size. For
this study, map size of 10 × 10 will be used.
iv. Number of learning iteration
The number of learning iteration is used to train the network.
v. Learning rate
Learning rate decides the speed of algorithm and the value is between 0 and 1.
The larger the value, the faster the network grows.
vi. Lattice topology
Lattice topology triangular, rectangular or hexagonal to be identified as this is one
of the comparison criteria.
vii. Distance measurement
Distance measurement method to be identified as this is one of the comparison
criteria too.
Once the parameters are defined, training process shall begin. The training pro-
cesses in steps as described below

[email protected]
186 S. Ramiah

a. Initialization of weights for output units

Weight vector (W ) for each node initialized to small standardized random values
between 0 and 1.
b. Present input
Input vector (V ) is fed into the network to train. The inputs are the training dataset.
c. Calculate the Best Matching Unit (BMU)
BMU is determined by iterating all the nodes and calculate the distance between
each node’s weight vector and input vector. The node which has weight vector closest
to the input vector is labeled as the BMU. Five different methods will be used to
calculate the distance. They are Euclidean, Manhattan, Chebychev and Canberra.
This is to identify which measurement gives better accuracy.
d. Update neighbouring nodes
Weight vector of every node within the BMU’s neighbourhood including the BMU
adjusted so that their weight gets more similar to the weight of input vector based
on following equation:

W (t + 1)  W (t) + (t)L(t)(V (t) − W (t)), (1)

where,
t time step (iteration of the loop)
L learning rate which decreases over time
Decay of learning rate calculated based on following:
 
t
L(t)  L 0 exp − (2)
λ

As training goes on, the area of the neighbourhood shrinks gradually over time to
the size of just one node which is the BMU by formula
 
dist2
(t)  exp − 2 (3)
2σ (t)

(t) influence rate

σ (t) width of the lattice at time t
e. Determine convergence
Steps b, c and d are repeated for all input data. Convergence will happen when
the training stops once an accurate topological map is obtained. Upon completion
of training process, the output vector holds the cluster for classification. Meanwhile,
weight vector defines the attributes of an element that falls into the segment. Both
this information will be used for evaluating test data.

[email protected]
Effect of Lattice Topologies and Distance Measurements … 187

In the testing phase, the dataset fed into the algorithm together with output vector
and weight vector obtained from training process. The output achieved from the
testing is compared with the expected target output to determine the ability of SOM
to classify data into correct predefined classes.

4 Lattice Topology

A topological map is a mapping that preserves the relative distance between the
points. A lattice is connected if neighbouring points in the input space are mapped
to nearby map units [6]. Neighbourhood function is applied on top of the topology.
In this study, triangular, rectangular and hexagonal topologies are experimented.
In triangular topology, neurons are arranged in a triangular lattice with each neuron
having three immediate neighbours. However, in rectangular and hexagonal topolo-
gies, neurons are arranged in a rectangular and hexagonal lattices with each neuron
having four and six immediate neighbours respectively.

5 Distance Measurement

There are several ways to work out the distance between two points in multidimen-
sional space. Changing the distance measurement can have a major impact on the
overall performance of the classification system [7]. Below are the four distance
measurements used for this experiment
(i) Euclidean Distance
Calculates the square root of the difference between coordinates of pair of
nodes.

 K
 
di j  
2
xik − x jk (4)
k1

(ii) Manhattan Distance

Calculates absolute difference between coordinates of pair of nodes.


K
di j  xik − x jk (5)
k1

(iii) Chebychev Distance

Calculate the absolute magnitude of the differences between coordinate of a
pair of nodes. Chebychev also known as maximum value distance.

[email protected]
188 S. Ramiah

Table 1 Analysis of iris dataset

Euclidean (%) Manhattan (%) Chebychev (%) Canberra (%)
Triangular 95.56 77.78 97.78 77.78
Rectangle 97.78 77.78 97.78 77.78
Hexagonal 100.00 84.44 97.78 77.78


 K

di j  
λ
λ
xik − x jk λ∞ (6)
k1

(iv) Canberra Distance

Calculate the sum of series of fraction differences between coordinates of a
pair of nodes.


K
xik − x jk
di j  , (7)
k1
|xik | + x jk

where,

xi input to the node i at time k

x j output node j at time k with weight from input node i
k no of iterations

for all of the above measurements.

6 Experimental Results

In this study, SOM is tested in different topology and distance measurement using
datasets as specified in Sect. 2.
Iris dataset represents small-scale data for this experiment. The result obtained
using triangular, rectangular and hexagonal topologies for this dataset is summarized
in Table 1.
For iris dataset, classification accuracy shows good result. All three topologies
work well with Euclidean distance whereby it gives 100% accuracy for hexagonal
and Euclidean combination. Rectangular and triangular topologies with Euclidean
also shows high accuracy percentage. Followed by Chebychev distance measurement
where all topologies give the same accuracy. Manhattan and Canberra show similar
result in term of its classification accuracy, but hexagonal topology using Manhattan
shows a slight increase in its percentage. Overall performance comparing all three
topologies confirms that hexagonal topology works better compare to triangular and
rectangular topologies using Euclidean distance measurement.

[email protected]
Effect of Lattice Topologies and Distance Measurements … 189

Table 2 Analysis of Breast Cancer Wisconsin (diagnostic) dataset

Euclidean (%) Manhattan (%) Chebychev (%) Canberra (%)
Triangle 88.76 88.76 88.76 71.00
Rectangle 88.76 88.76 88.76 71.00
Hexagonal 88.76 88.76 88.76 79.89

Table 3 Analysis of car evaluation dataset

Euclidean (%) Manhattan (%) Chebychev (%) Canberra (%)
Triangular 90.91 90.91 90.91 85.23
Rectangle 91.86 94.70 91.86 71.97
Hexagonal 96.60 90.91 85.23 75.76

For the second experiment, Breast Cancer Wisconsin (Diagnostic) dataset was
used as medium scale dataset. The result obtained is summarized in Table 2.
For breast cancer dataset, almost all three topologies show similar classification
accuracy for Euclidean, Manhattan and Chebychev distance measurement. Accuracy
percentage does not increase even if multiple testing is done. However, Canberra
distance measurement showed a slightly higher accuracy percentage for hexago-
nal topology despite the other two remain lowest among the overall testing output.
Overall performance comparing all three topologies confirms that hexagonal topol-
ogy works better compare to triangular and rectangular topologies using any distance
measurements for this dataset.
For the last experiment, Car Evaluation data used to test as large-scale data. Result
for this experiment is represented in Table 3.
Based on the above result, can be seen that hexagonal topology using Euclidean
distance gives the highest accuracy. Rectangle topology with collaboration of Man-
hattan can be said as good classification too. Results of other testing show similar
classification accuracy percentage and among all Canberra proves the poorest. Over-
all performance comparing all three topologies confirms that hexagonal topology
works better compare to triangular and rectangular topologies using Euclidean dis-
tance measurement.
Based on the above experiments, can be seen that classification accuracy differs in
each type of dataset. As described above, this study is done to address two issues. First
to investigate which lattice topology is better and second is to identify which distance
measurement proves better classification. Comparison between lattice topologies and
distance for each dataset is done as described in Table 4.
Euclidean and Chebychev proved good classification accuracy for small-scale
dataset. Euclidean, Manhattan and Chebychev work well for medium-scale dataset.
For larger scale dataset, Euclidean and Manhattan performed better.

[email protected]
190 S. Ramiah

Table 4 Summary of result analysis based on topology and distance measurement

Topology Iris Breast Cancer Car evaluation
Wisconsin (Diagnostic)
Triangular Chebychev All except Canberra All except Canberra
Rectangular Euclidean, Chebychev All except Canberra Manhattan
Hexagonal Euclidean All except Canberra Euclidean

Hexagonal topology showed higher accuracy percentage for all scale dataset fol-
lowed by rectangular. Triangular topology does not show much performance to be
stressed out. Euclidean distance measurement moves well with all topologies for all
scale dataset.

7 Conclusion

SOM is a famous algorithm for classification yet there are areas of research to be car-
ried out and one of it is classification accuracy. Therefore, in this paper lattice topolo-
gies and distance measurements are investigated to evaluate classification accuracy.
Based on experimental results, can be concluded that SOM using hexagonal topol-
ogy enhances the performance of classification accuracy mainly due to connection
between neurons in SOM map in hexagonal topology is more compared to rectan-
gular and triangular.
Euclidean distance has proven to calculate distance between nodes perfectly in
every topologies analyzed. The efficiency of classification relies on computation
complexity. Euclidean distance which requires less computation is seen to reduce
the computation complexity of classification system compared to other distance mea-
surements which are used in this experiment. Since the computation complexity is
less, it proves better performance in term of accuracy.
The size of dataset used does not reflect much difference in this study probably
due to the data used from clean dataset. Thus, network corresponds accordingly to
form good classification.

References

1. Kohonen, T. (1997). Self-organizing maps. In Springer series in information sciences. Berlin,

Heidelberg, New York: Springer.
2. Pastukhov, A., & Prokofiev, A. (2016). Kohonen self-organizing map application to representa-
tive sample formation in the training of the multilayer perceptron. St. Petersburg Polytechnical
University Journal: Physics and Mathematics 2, 134–143.
3. ChandraShekar, B. H., & Shoba, G. (2009). Classification of documents using Kohonen’s self-
organizing map. International Journal of Computer Theory and Engineering, 1(5), 610–613.

[email protected]
Effect of Lattice Topologies and Distance Measurements … 191

4. Dua, D., & Karra Taniskidou, E. (2017). UCI machine learning repository http://archive.ics.uci.
edu/ml. Irvine, CA: University of California, School of Information and Computer Science.
5. Jaramillo-Vacio, R., Ochoa-Zezzatti, A., & Figueroa-Godoy, F. (2015). Self-organizing map
improved for classification of partial discharge using desirability function. International Journal
of Combinatorial Optimization Problems and Informatics, 6(3), 49–65.
6. Chaudhary, V., Ahlawat, A. K., & Bhatia, R. S. (2013). An efficient self-organizing map learn-
ing algorithm with winning frequency of neurons for clustering application. In Proceedings of
International Advanced Computing Conference (pp. 664–668). IEEE.
7. Singh, A., Yadav, A., Rana, A. (2013). K-means with three different distance metrics. Interna-
tional Journal of Computer Applications, 67(10).
8. Lohninger, H. (2018). Kohonen Network—Background Information, http://www.lohninger.
com/helpcsuite/kohonen_network_-_background_information.htm.

[email protected]
Evaluation and Classification of Road
Accidents Using Machine Learning
Techniques

Jaspreet Singh, Gurvinder Singh, Prithvipal Singh and Mandeep Kaur

Abstract The root cause of traffic accidents is hard to determine these days due to
complex combination of characteristics like mental state of driver, road conditions,
weather conditions, traffic, and violations of traffic rules to name a few. The deploy-
ment of machine learning classifiers has replaced traditional data mining techniques
like association rule mining. Application of machine learning techniques in the field
of road accidents is gaining popularity these days. This paper utilized four machine
learning techniques viz. Naïve Bayes, k-Nearest Neighbours, Decision trees, and
Support Vector Machines for evaluation of Punjab road accidents. This work had a
challenge of performing parametric evaluation to extract highly relevant parameters
especially for Punjab. The outcome of this study yields 12 most suitable parameters
and maximum performance of 86.25% for Decision Tree classifier.

Keywords Road accidents · Machine learning · Parametric evaluation · Punjab

road safety

1 Introduction

Road accident is one of the most prominent issues in the modern times of equipped
and fast moving traffic on roads. World Health Organization has reported top ten
disastrous reasons for taking human’s life, and unfortunately road accidents come
at ninth place, where cardiac arrest sits on the top. The impact on society seems
significant when cost of casualties and injuries from road accidents is evaluated. The
young researchers these days have witnessed the increasing trend of evaluating the
causes and implementing safety measures in order to preserve human life from dan-
gerous road mishaps. Also, the trending machine learning techniques have facilitated
the overall process of understanding the features like demographic factors, physical

J. Singh (B) · P. Singh · M. Kaur

Department of CS, GNDU, Amritsar 143005, Punjab, India
e-mail: [email protected]
G. Singh
Department of CS, Faculty of Engineering and Tech, GNDU, Amritsar 143005, Punjab, India
© Springer Nature Singapore Pte Ltd. 2019 193
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_17

[email protected]
194 J. Singh et al.

factors, victim related factors, driving environment, mental factors, and other culprit
features for accident severity. This paper employs four machine learning techniques
viz. Naïve Bayes (NB), K-Nearest Neighbours (k-NN), Decision Trees (DT), and
Support Vector Machines (SVM). The dataset for the evaluation is taken from Pun-
jab government’s authentic organization named Punjab Road Safety Organization.
However, it was difficult to extract data from reports in diverse Punjabi fonts but,
the use of Google Punjabi to English translator has facilitated the data preprocess-
ing task marvelously. Section 2 of the manuscript provides literature work of some
state-of-the-art machine learning algorithms used for evaluation and classification of
road accidents. The parametric evaluation (Sect. 3 of the manuscript) is exhaustively
performed thereby examining around 71 parameters taken from 12 benchmarks pub-
lished in the same research direction. Further, the close investigation of dataset for
Punjab state of India motivated this work to consider 12 highly relevant parameters
associated with Punjab road mishaps. Section 4 gives experimental results, where the
highlights of best performed classifier (Decision Tree) are being presented in terms
of confusion matrix followed by concluding remarks in Sect. 5.

2 Related Work

Road accident is a global issue with skyrocketing trend where the immediate need
is to analyze data through some data classification algorithms which help in antic-
ipating the relevant factors influencing the problem of accident severity. Table 1
shows summary of related research work, where the well-known machine learning
classifiers are exploited for the evaluation of traffic accident analysis.

3 Parametric Evaluation

This study considers twelve benchmarks related to evaluation of road accidents. The
close investigation of 71 parameters threw light on enormous range of possibilities
where the accurate understanding of causes behind accidents is gained. However,
most of these parameters are not suitable for evaluation of road accidents in the Punjab
state of India. Table 2 gives insights to consider the 12 contributing parameters viz.
Road type, Vehicle type, Vehicle speed, Light conditions, Traffic volume, Accident
time, Accident severity, Stray animal, Alcohol consumption while driving, Age,
Gender, and mental conditions of driver.

4 Experimental Results and Discussion

This work has utilized Scikit-Learn package from SciPy library of Python 3.6 for
experimental evaluation. The classification models viz. Naïve Bayes (NB), K-nearest

[email protected]
 Table 1 Summary of related work
Author(s), publication, (Year) Description Dataset Attributes Machine learning Performance
techniques
Ramani, R. G., et al. WCECS Analyzed and Obtained data from State, Country, C4.5, CR-T, ID3, Arc-X4 meta
(2012) [1] classified road traffic Fatality Analysis month, date, CS-CRT, CS-MC4, classifier has shown
data through data Reporting System collision, seating Naïve Bayes, highest accuracy of
mining techniques (FARS) position, age, gender, Random trees, 99.73%
range, injury, severity, Arc-X4, Feature
ejection, alcohol ranking algorithm
Elfadil A. Mohamed, DAVD Proposed a Taken from police Location, vehicle Multiclass support Precision  0.767,
JCC, (2014) [2] framework for department in Dubai type, country a driver vector machine Recall  0.754, F1
predicting cause of and United Arab belongs to, age, year measure  0.752,
road traffic accidents Emirates of vehicle made, Accuracy  75.395%
gender
ARIMURA et al. Infrastructure Performed road Taken dataset from Road structure, SVM Accuracy of
Planning Review (2007) [3] accident analysis by Institute for Traffic driving environment, classification
extracting dangerous Accident Research designated maximum observed was 70%
locations with high and Data Analysis speed
Evaluation and Classification of Road Accidents Using …

[email protected]
fatality and injury (ITARDA)
Amira. A. El Tayeb., et al., Evaluation of traffic Dataset was taken Day, month, year, Apriori algorithm and Apriori algorithm was
IJSCE, (2015) [4] accidents using from Dubai traffic number of injured predictive apriori found more accurate
association rule department, UAE persons, cause of algorithm than predictive apriori
mining algorithm from 2008 to 2010 accident, type of algorithm
accident, gender, age,
vehicle type, alcohol,
seat belt, lightening,
weather, road
conditions
(continued)
195
 Table 1 (continued)
196

Author(s), publication, (Year) Description Dataset Attributes Machine learning Performance

techniques
Martin et al., Elsevier (2014) Proposed a decision Dataset was taken Layout, signaling and Data mining Performance of
[5] tree based algorithm from Spanish national marking of road techniques (Decision Decision Tree was
for detection and government obstacles, Tree) satisfactory
improvement of intersection, junction,
susceptible elements roundabout,
in Spanish roads crossings, tunnels,
road access, etc.
Fu. H., et al., IEEE, (2011) [6] Presented a The national accident Number of accidents, Backpropagation Improved LMBP
prediction technique statistics from 1985 time, death, direct neural networks performed better than
for traffic accidents to 1997, and 1995 to loss traditional BP
using neural networks 1999 network
Miao, Chong., et al. Developed a machine Obtained from Age, gender, eject, Artificial Neural DT_ANN method
Informatica, (2005) [7] learning based model National Automotive alcohol, restraining Networks (ANN) outperforms ANN
for accurate Sampling System system, body type, using hybrid learning, and DT with
classification of (NASS) and General vehicle type, rollover, Decision Trees (DT), maximum average

[email protected]
severity of injuries in Estimates System road surface Support Vector accuracies of 91.53%
road accidents (GES) condition, light Machines (SVM) and 90.00% for
condition training and testing
respectively
K. Geetha., et al., IJARCSSE, Proposed a hybrid Traffic accident Age, gender, alcohol J48, PART algorithm, DTANN outperforms
(2015) [8] machine learning dataset taken from usage, eject vehicle, Hybrid DTANN
model for analysis of Tamil Naidu body type, vehicle
traffic accidents Government role, vehicle age,
rollover, road surface
conditions, light
conditions, weather
conditions
(continued)
J. Singh et al.
 Table 1 (continued)
Author(s), publication, (Year) Description Dataset Attributes Machine learning Performance
techniques
Sohn, S. Y., et al., Elsevier, Evaluation of severity International database Road width, shape of DTNN, Dempster Clustering method
(2003) [9] of road accidents in on road traffic and car body, accident Shafer algorithm, using Decision Tree
Korea accident (IDRTA) category, speed Bayesian procedure, produced best
before accident, logistic model, accuracy of 76.10%
violent drive, and ensemble, bagging,
protective drive clustering, and
K-means algorithm
Vasavi, S., Springer Nature Proposed model Dataset was taken Traffic, time of K-mediods, and EM Precision  0.8,
Singapore, (2018) [10] analyze hidden from police accident, age, clustering algorithm recall  0.6,
patterns to identify authorities of Andhra weather, speed limit, F-measure  0.69
the root causes for Pradesh (Nunna type of accident,
road accidents Dataset) deceased emotions,
hospital reported etc.
Razzaq, S., et al. ICCEEE, Coined a system Hypothetical dataset Environmental, Used fuzzy logic’s Proposed system
(2016) [11] based on multiple obtained from virtual physical and mental modeling in Multi outperformed over
Evaluation and Classification of Road Accidents Using …

[email protected]
factors for evaluation sensors in the factors Factor Based Road traditional methods
of road accidents simulator Accident Prevention
System (MFBRAPS)
Kumar, S., ICCCS (IEEE), Performed clustering Taken data from Victim injured, age, K-mode clustering Clustering followed
(2015) [12] and association rule Emergency gender, economic and association rule by association rule
mining of road traffic Management status, time of mining algorithms mining yields
accidents Research Institute accident, day, month, satisfactory results
(EMRI) region, lightening,
road features
197
 Table 2 Parametric comparison among twelve benchmarks
198

Parameters Benchmarks
B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11 B12
Demographic Weather Winter ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✓ ✓ ✓
factors Summer ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✓ ✓ ✓
Road conditions Type ✗ ✓ ✓ ✓ ✓ ✗ ✓ ✓ ✗ ✓ ✓ ✓
Separation ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✓ ✗ ✓ ✗ ✗
Width ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✓ ✓ ✗ ✗
Orientation ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✓ ✗ ✓ ✗ ✗
Junction ✗ ✗ ✓ ✗ ✓ ✗ ✗ ✓ ✗ ✓ ✗ ✗
Lanes ✗ ✓ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗
Sidewalk ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗
Turns ✗ ✓ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗
Light conditions Crossings ✗ ✓ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Lightening ✗ ✗ ✗ ✓ ✗ ✗ ✓ ✓ ✗ ✓ ✓ ✓

[email protected]
Traffic Volume ✗ ✗ ✓ ✗ ✓ ✗ ✗ ✗ ✗ ✓ ✓ ✗
Physical factors Vehicle conditions Type ✗ ✗ ✗ ✓ ✗ ✗ ✓ ✓ ✓ ✓ ✓ ✗
No. Plate ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗
Company ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗
Speed ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗
Rollover ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗
Accident severity Fatal ✓ ✗ ✗ ✓ ✗ ✗ ✓ ✓ ✗ ✗ ✗ ✗
No Injury ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✓ ✗ ✓ ✗ ✗
Possible Injury ✓ ✗ ✗ ✓ ✗ ✗ ✓ ✓ ✗ ✗ ✗ ✓
(continued)
J. Singh et al.
 Table 2 (continued)
Parameters Benchmarks
B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11 B12
Non-incapacitating ✓ ✗ ✗ ✗ ✗ ✗ ✓ ✓ ✗ ✗ ✗ ✗
Incapacitating ✓ ✗ ✗ ✓ ✗ ✗ ✓ ✓ ✗ ✗ ✗ ✓
Accident type Collision Manner ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Seating Position ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Harmful Event ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Punctured Tire ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Reckless driving ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Wrong direction ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Incorrect pass ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
U-turn ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Stray animal ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Evaluation and Classification of Road Accidents Using …

Sleep ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗

[email protected]
Door open ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
License With ✗ ✓ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Without ✗ ✓ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Ejection Path ✓ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗
Extrication ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Destructive activity Eating ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗
Texting ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗
Overspeed ✗ ✓ ✓ ✗ ✗ ✗ ✗ ✗ ✓ ✓ ✓ ✗
(continued)
199
 Table 2 (continued)
200

Parameters Benchmarks
B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11 B12
Victim related Driver’s profile State ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Country ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Age ✓ ✗ ✗ ✓ ✗ ✗ ✓ ✓ ✗ ✓ ✓ ✓
Experience ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✓ ✗ ✗ ✗
Gender ✓ ✗ ✗ ✓ ✗ ✗ ✓ ✗ ✗ ✗ ✓ ✗
Nationality ✓ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Economic Status ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✓
Person type Driver ✓ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✓
Passenger ✓ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✓
On the spot death Year ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Month ✓ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗
Day ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗

[email protected]
Drug test Alcohol ✓ ✓ ✗ ✓ ✗ ✗ ✓ ✗ ✗ ✓ ✗ ✗
Other Drug ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Driving environment Accident location Suburb ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✓ ✗ ✗ ✗
Country side ✓ ✓ ✗ ✓ ✗ ✗ ✓ ✗ ✓ ✗ ✓ ✗
Accident time Month ✓ ✗ ✓ ✓ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✓
Date ✓ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Day ✓ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Time ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✓
(continued)
J. Singh et al.
 Table 2 (continued)
Parameters Benchmarks
B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11 B12
Year ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗
Mental factor Mental condition of Emergency ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗
driver Fear ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗
Temper ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗
Behavior ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗
Negligence ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗
Others Other parameters No. of persons injured ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗
Accident cause ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗
Signaling ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗ ✗
Ambulance used ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗
B1 Shanthi et al. [1], B2 Mohamed [2], B3 Hironobu et al. [3], B4 Amira et al. [4], B5 Luis et al. [5], B6 Fu Huilin et al. [6], B7 Chong et al. [7], B8 Geetha
et al. [8], B9 So et al. [9], B10 Vasavi [10], B11 Sheeba et al. [11], and B12 Kumar et al. [12]
Evaluation and Classification of Road Accidents Using …

[email protected]
201
202 J. Singh et al.

Table 3 Classification S. No. Classification model Accuracy (%)

accuracy
1 Naïve Bayes 79.99
2 K-Nearest neighbor 85.73
3 Decision tree 86.25
4 Support vector machine 84.33

Table 4 Top three Contributing factor Percentage of accidents (%)

contributing parameters for
accidents Mental condition of driver 24.57
Alcohol consumption 26.49
Speed of vehicle 16.92

neighbours (k-NN), Decision Trees (DT), and Support Vector Machine (SVM) were
built to analyse causes for road accidents in Punjab. The dataset consists of 38,604
accident cases reported for Punjab state on www.punjabroadsafety.org [13]. The
initial 500 cases were manually preprocessed, thereby providing average values at
missing places for collecting 12 parameters. The rest of the cases are managed by
Python code to fill missing values either with null in case of nominal factors or with
average values for ratio scale parameters. The dataset is sliced in the proportions of 80,
10, and 10% for training, testing and cross-validation purposes respectively. Table 3
gives classification accuracies of four machine learning models. The performance of
Decision Tree learning model is higher among the four classifiers.
The close investigation of dataset shows that the top three contributing parame-
ters among 12 are mental conditions of driver, alcohol consumption, and speed of
vehicle. Table 4 shows the topmost three contributing factors in terms of percent-
age of accidents [13]. The confusion matrix of 12 parametric classes classified with
Decision Tree method depicts an error rate of 13.75% thereby yielding the accuracy
of 86.25% as shown in Table 5.

5 Conclusion

This paper evaluated road accidents using four machine learning classifiers on Punjab
road safety organization’s dataset. The twelve benchmarks illuminated 71 parame-
ters extensively; however, these parameters depend upon regional factors. The keen
examination of all collected parameters (in Table 2) assisted the focus on 12 param-
eters which are most suitable for evaluation of Punjab road accidents. The results
obtained after 10-fold cross-validation show that Decision Tree classifier carries
maximum accuracy of 86.25% among four classifiers. The main causes behind the
road accidents in Punjab come from three most contributing factors viz. mental state
of driver, alcohol consumption, and speed of vehicle. The future scope of this work

[email protected]
 Table 5 Confusion matrix of decision tree classifier (Error rate: 13.75%)
Error rate: 0.1375
RT VT AT TV LC MC AC AS Age Gen Sp SA Sum
RT 1140 0 2 2 0 3 2 0 0 0 1 0 1150
VT 0 1040 2 1 2 1 2 100 0 0 0 0 1148
AT 2 0 1029 0 10 1 0 0 0 0 0 0 1042
TV 0 0 0 1230 0 0 22 208 0 0 72 1 1533
LC 0 0 0 0 2310 0 0 0 0 0 0 0 2310
MC 12 0 1 2 0 6964 2326 0 10 0 103 0 9418
AC 0 0 0 0 0 101 7937 2114 2 0 0 0 10154
AS 2 0 0 0 1 0 0 2277 2 2 12 2 2298
Evaluation and Classification of Road Accidents Using …

[email protected]
Age 0 12 10 0 0 0 0 0 1894 0 0 0 1916
Gen 0 0 0 13 0 0 1 0 0 752 0 0 766
Sp 0 0 0 0 32 105 0 0 0 0 6349 0 6486
SA 2 0 2 0 3 0 0 2 0 0 0 374 383
Sum 1158 1052 1046 1248 2358 7175 10,290 4701 1908 754 6537 377 38,604
Here RT Road Type, VT Vehicle Type, AT Accident Time, TV Traffic Volume, LC Lightening Condition, MC Mental Condition, AC Alcohol Consumption, AS
Accident Severity, Age Age of driver, Gen Gender, Sp Speed of Vehicle, SA Stray Animal
203
204 J. Singh et al.

is to consider sentiment analysis of road accident’s cases using ensemble classifiers

and deep neural network.

References

1. Ramani, R. G., & Shanthi, S. (2012). Classifier prediction evaluation in modeling road traffic
accident data. In 2012 IEEE International Conference on Computational Intelligence and
Computing Research. https://doi.org/10.1109/iccic.2012.6510289.
2. Mohamed, E. A. (2014). Predicting causes of traffic road accidents using multi-class support
vector machines. Journal of Communication and Computer, 11(5), 441–447. https://doi.org/
10.17265/1548-7709/2014.05004.
3. Arimura, M., Hasegawa, H., Fujii, M., & Tamura, T. (2007). A consideration on application
of support vector machine for non-linear optimization. Infrastructure Planning Review, 24,
421–426. https://doi.org/10.2208/journalip.24.421.
4. El Tayeb, A. A., Pareek, V., & Araar, A. (2015). Applying association rule mining algorithms
for traffic accidents in Dubai. International Journal of Soft Computing and Engineering, 5(4),
1–12.
5. Martín, L., Baena, L., Garach, L., López, G., & De Oña, J. (2014). Using data mining techniques
to road safety improvement in Spanish Roads. Procedia—Social and Behavioral Sciences, 160,
607–614. https://doi.org/10.1016/j.sbspro.2014.12.174.
6. Fu, H., & Zhou, Y. (2011). The traffic accident prediction based on neural network. In 2011
Second International Conference on Digital Manufacturing & Automation. https://doi.org/10.
1109/icdma.2011.331.
7. Chong, M., Abraham, A., & Paprzycki, M. (2005). Traffic accident analysis using machine
learning paradigms. Informatica, 29, 89–98.
8. Geetha, K., & Vaishnavi, C. (2015). Analysis on traffic accident injury level using classification.
International Journal of Advanced Research in Computer Science and Software Engineering,
5(2), 953–956.
9. Sohn, S. Y., & Lee, S. H. (2003). Data fusion, ensemble and clustering to improve the classifi-
cation accuracy for the severity of road traffic accidents in Korea. Safety Science, 41(1), 1–14.
https://doi.org/10.1016/s0925-7535(01)00032-7.
10. Vasavi, S. (2016). A survey on extracting hidden patterns within road accident data using
machine learning techniques. Communications on Applied Electronics, 6(4), 1–6. https://doi.
org/10.5120/cae2016652455.
11. Razzaq, S., Riaz, F., Mehmood, T., & Ratyal, N. I. (2016). Multi-factors based road accident
prevention system. In 2016 International Conference on Computing, Electronic and Electrical
Engineering (ICE Cube). https://doi.org/10.1109/icecube.2016.7495221.
12. Kumar, S., & Toshniwal, D. (2015). Analysing road accident data using association rule min-
ing. In 2015 International Conference on Computing, Communication and Security (ICCCS).
https://doi.org/10.1109/cccs.2015.7374211.
13. www.punjabroadsafety.org. Last Accessed January 11, 2018.

[email protected]
Multi-language Handwritten Recognition
in DWT Accuracy Analysis

T. P. Umadevi and A. Murugan

Abstract Discovery of gender orientation from penmanship of an individual shows

an intriguing exploration issue with applications in measurable record test, essay-
ist recognizable proof and mental investigations. This paper displays a compelling
strategy to anticipate the gender orientation of a person from disconnected pictures
of penmanship. The proposed strategy depends on a worldwide approach that thinks
about composing pictures as surfaces. Each written by hand picture is changed over
into a finished picture which is disintegrated into a progression of wavelet sub-groups
at various levels. The wavelet sub-bands are then stretched out into information suc-
cessions. Every datum succession is quantized to create a Discrete Wavelets Trans-
form (DWT) that produces extraction. These highlights are utilized to prepare two
classifiers, numerous occurrence learning, and multiclass bolster vector machine to
separate among male and female handwriting recognition compositions. The execu-
tion of the proposed framework was assessed on two databases, HWSC and TST1,
inside various testing exploratory situations and acknowledged arrangement rates of
up to 94.07%.

Keywords Disconnected penmanship examination · Gender location · Surface

investigation · Wavelet sub-band · Emblematic progression

1 Introduction

Examination of penmanship and hand-drawn shapes is a delightful area of research

for clinicians, report analysts, paleographers, graphologists, scientific investigators,
and pc science analysts. While the guide assessment of penmanship has been an
activity for a long time, the mechanized assessment appreciates a restored look

T. P. Umadevi (B)
Department of Computer Science, JBAS College for Women, Teynampet, India
e-mail: [email protected]
A. Murugan
Department of Computer Science, Dr. Ambedkar Government Arts College (Autonomous),
Vysarpadi, Chennai 600039, Tamil Nadu, India

© Springer Nature Singapore Pte Ltd. 2019 205

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_18

[email protected]
206 T. P. Umadevi and A. Murugan

into pastime because of the very most recent innovative and algorithmic headways
in extraordinary zones of workstation sciences. In spite of the fact that penman-
ship acknowledgment remains the greatest programming of automated penmanship
examination, a wide range of fascinating applications have likewise been explored.
These incorporate order of composing styles, watchword perceiving in written by
hand documents and, recognizable proof and confirmation of essayists from their
transcribed specimens. A nearly related inconvenience to make recognizable proof
is the order of shopper socioeconomics from penmanship. Distinguishing proof of
statistic properties which incorporates gender orientation, handedness, age, race and
soon have been researched in the writing [1, 2]. Additionally, a couple of mental
research propose the presence of connection between’s penmanship and particular
traits two of the character of the creator [3–8]. The cases of these graph logical con-
siders, nonetheless, stay subjective and are however to be approved by method for
trial ponders. As unfavorable to private traits which tend to be subjective, statistic
properties are objective and the order of these qualities from penmanship can be
tentatively approved through measured outcomes. Among an assortment of statistic
characteristics, a fairly solid relationship has been tried among gender and parts of
penmanship in a wide assortment of studies [9–14]. Mental research advocate that
the male compositions, by and large, tend to be more prominent ‘spiky’, ‘rushed’
and ‘chaotic’ while the lady compositions are likely to be additional ‘embellishing’,
‘homogenous’ and ‘reliable’ [13–15]. Programmed discovery of gender construct
absolutely in light of these highlights makes the issue of our examination.
Programmed gender orientation forecast from transcribed examples can be
inferred in a wide assortment of intriguing applications not exclusively for clinicians
however furthermore for criminological pros furthermore, record inspectors. It can
fill in as a sifting step allowing experts to diminish the inquiry space and concentrate
on a subset of compositions originating from a particular gender orientation gather-
ing. In addition, gender orientation characterization can moreover prompt expanded
impacts in many creators distinguishing proof and check purposes [16]. Notwith-
standing these used parts of gender identification, the find out about of connec-
tion among gender orientation and particular written work signs offers a fascinating
inconvenience of crucial research. Gender orientation arrangement is deliberately
connected with the creator recognizable proof issue. A fundamental assessment of
the gender orientation grouping procedures principally in light of penmanship styles
uncovers that in many examples features utilized for creator recognizable proof have
been adjusted for gender order [1, 2, 16]. This paper tends to be computerized gen-
der orientation forecast from disconnected manually written pictures. We advocate
a novel brilliant approach that thinks about works as surfaces what’s more, utilizes
wavelets to symbolize male and lady works. An outline of the means stressed in the
proposed strategy is displayed in Fig. 1. The proposed work include extraction, relies
upon wavelet change the utilization of representative dynamic separating (SDF). In
a current report, SDF-based capacity extraction from the time succession of robot
development conduct has been completed by means of Hassaine et al. [17]. More-
over, SDF-based capacity extraction from time arrangement data has been proposed
through Al-Maadeed et al. [18] for target recognition and order in outskirt locales. To

[email protected]
Multi-language Handwritten Recognition in DWT … 207

Fig. 1 Hand written proposed framework DWT

the best of our insight, this is the principal endeavor of abusing emblematic dynamic
separating to imply gender from penmanship and speaks to the essential commit-
ment of this work. In our component extraction technique, the wavelet sub-band is
stretched out into a data grouping that is symbolized to develop Discrete Wavelets
Change (DWT) which produces the component vectors.
These component vectors are marked by methods for two k-closest neighbors and
Multi-bolster vector figuring gadget (SVM). The proposed technique is assessed on
the normal HWSC also, TST1 databases with composing tests in English, Cursive,
and Arabic.

2 Related Work

In this area, we initially talk about the connection among penmanship and gender
as perceived in the manual assessment of penmanship. Gender orientation and Pen-
manship Various investigations [12–14] have inferred that gender can be anticipated
from penmanship through the precision of expectation may shift. It has likewise been
watched that people who every now and again manage manually written pictures end
up plainly gifted at separating the compositions of the two gender orientation classes
[1]. These distinctions are for the most part credited to the distinctions in engine coor-
dination among the two sexes. The designs explanatory examination of Pratikakis
[19] considered the impact of gender on penmanship. Looking at male and female
penmanship styles is a goal reconsideration of Lesters graphology discoveries that
there are contrasts between them works originating from various gender orientations
[9]. As talked about before, when all is said in done, female works have a tendency
to be slick, even, efficient, adjusted, little and symmetrical.
Male compositions, then again, are for the most part portrayed by properties like
rough, uneven, messier, inclined, and slanting [13]. An arrangement of point by point
separating factors among male and female compositions can be found in [20].

[email protected]
208 T. P. Umadevi and A. Murugan

Eames did an examination utilizing tests from 12 college understudies to decide if

penmanship has any relationship with the evaluations of the understudies and whether
the gender of an understudy could be perceived through his or her penmanship. The
consequences of the examination demonstrated that evaluations of understudies were
most certainly not impacted by the nature of penmanship. In any case, as a rule,
it was conceivable to decide the gender orientation of the understudy. In another
investigation, Agius et al. [14] performed a few analyses with male and female
understudies and revealed a 75% right distinguishing proof of gender orientation
even with exceptionally constrained written work tests. The creator closed that by
and large, one letter or image is frequently enough for distinguishing proof of gender
orientation. In a different report on manual examination of penmanship by Chahi
et al. [12], the writers investigated English and Urdu penmanship tests gathered from
30 distinct people. Every benefactor duplicated a 50-word entry (both in Urdu and
English) and the literary substance of the section was same for every one of the
people. These examples were then dissected by 25 inspectors for recognizable proof
of gender orientation. It was inferred that gender orientation distinguishing proof is
similarly dependable for content in both the dialects with a normal exactness of 68%.
The fascinating discoveries of analysts on the connection among penmanship and
gender orientation enlivened the PC researchers to apply picture investigation and
example arrangement methods to computerize this investigation. The accompanying
subsection shows a diagram of the automated grouping of gender orientation from
penmanship.
SVM proposed a gender and handedness order system from internet penmanship
that, notwithstanding the disconnected portrayal of character shapes permit misusing
the fleeting data of composing too. The creators extricate an arrangement of 29 high-
lights from the online-specimens and its disconnected portrayal what’s more, utilize
bolster vector machine and Gaussian blend models for arrangement. Exactnesses
of up to 67.06 and 84.66% are accounted for gender orientation and handedness
arrangement separately. A different examination of disconnected and online works
uncovered that the grouping rates on online penmanship styles are superior to those
on disconnected pictures of composing. In [21], the writers misuse highlights like
Fourier descriptors and shape data to distinguish gender orientation from penman-
ship. The creators be that as it may, do not present any measured outcomes and
examine the estimations of these parameters for male and female compositions.

3 Problem Analysis

While DWT technology can be effective in converting handwritten or typed charac-

ters, it does not give as high accuracy as of Document Form for reading data, where
users are actually marking forms.
Additional workload to data collectors SDF has severe limitations when it comes
to human handwriting. Characters must be handprinted with separate characters in
boxes.

[email protected]
Multi-language Handwritten Recognition in DWT … 209

Flow of work.

3.1 Dataset

The trial investigation of the proposed framework is done on two surely understood
penmanship databases, the Qatar College Essayist Recognizable proof (HWSC)
database furthermore, Multi-content Manually written Database (TST1).

3.2 HWSC Database

The HWSC database contains composing tests contributed by 1017 journalists. Every
author gave four tests, two in English and two in Arabic. Page 1 and Page 3 of
all scholars contain a content from essayist’s own creative energy in Arabic and
English individually. In like manner, page 2 (Arabic) and page 4 (English) of every
author contains the same literary substance. To permit significant examination of the
execution of our framework with different investigations, we completed the analyses
on composing tests of 475 journalists. The preparation set incorporates composing
tests of 282 authors while the test set contains 193 authors.

3.3 TST1 Database

The TST1 database contains written by hand tests gathered from an aggregate of 84
journalists. Every essayist was required to compose 12 pages, six each in Cursive and
Arabic and each page had the same printed content for all essayists. In the greater
part of the assessments, the preparation and test sets involve composing tests of 42
authors each (Fig. 2).

4 Letter Style

4.1 Offline and Online

1. Online–captured by pen-like devices, the input format is a two-dimensional signal

of pixel locations as a function of time (x(t), y(t)).
2. Offline–captured by scanning devices, the input format is a two-dimensional
image of grayscale colors as a function of location I(m*n). Strokes have signifi-
cant width (Figs. 3 and 4).

[email protected]
210 T. P. Umadevi and A. Murugan

Fig. 2 DWT using data base

analysis

Fig. 3 Handwritten offline

and online pre processing

Fig. 4 Pattern recognition and machine learning conflicts

[email protected]
Multi-language Handwritten Recognition in DWT … 211

Table 1 Database in documents framework analysis

Database HWSC TST1
Text Form English Form English
Number of writers (training test) 475 (282 + 193) 84 (42 + 42)
Distribution of writers 221 male & 254 female 37 male & 47 female
Samples per writer 4 (2 Arabic, 2 English) 12 (6 Arabic, 6 Cursive)
Total samples 1900 (4 × 475) 1008 (12 × 84)
Size of texture block 256 × 256 256 × 256
Texture blocks from n pages n  4, 52 (11 + 15 + 11 + n  12, 180 (15 + 15 + ··· +
15) (p1 + p2 + ··· + pn ) 15)
Total number of texture blocks 24,700 (52 × 475) 15,120 (180 × 84)

Table 2 DWT background on cursive handwriting

Letter and number Pos Orientation Angle
Down cusp 3.0 −90° Left
Up loop 2.0 90° Right
Down arc 2.75 180° Two hand level (Left, right)

Fig. 5 Documents form offline cursive word indication representation

It ought to be noticed that in every single test setting, the preparation and test sets
do not contain any written work test of a similar author with the goal that the issue
compares to gender orientation grouping and not essayist recognizable proof. The
division of scholars in preparing and test sets alongside other data is exhibited in
Table 1 (Table 2).
This area displays the subtle elements of the proposed approach including pre-
preparing, highlight extraction, and arrangement. The proposed portrayal of gender
orientation from pictures of penmanship principally depends on Wavelet change and
system Dynamic Separating (SDF). We display a concise outline of these ideas in the
accompanying before talking about their application to our specific issue (Fig. 5).

[email protected]
212 T. P. Umadevi and A. Murugan

5 Handprinted Implementation DWT

Wavelets are by and large named as a numerical magnifying lens in flag and picture
handling. Portraying time-subordinate information utilizing wavelet premise brings
about an intense portrayal of data that is at the same time confined in time and recur-
rence spaces. This is as opposed to the Fourier portrayal where it is not conceivable to
relate particular frequencies to particular interims of time. Wavelets are particularly
effective with regards to examination of signs and pictures with discontinuities and
sharp spikes and discover applications in a wide assortment of issues in Material
science, Arithmetic, and Electrical Designing, supplanting the traditional Fourier
changes much of the time. Run of the mill uses of wavelets in flag preparing incor-
porate picture pressure, picture denoising, discourse acknowledgment, EEG, EMG
and ECG examinations and so on. An inferable capacity Ψ ⊆ L2(R) is
• Wavelet: finite interval function with zero mean (suited to analysis transient sig-
nals)
• Utilize the combination of wavelets (basis function) to analyze arbitrary function
• Mother wavelet Ψ (t): by scaling and translating the mother wavelet, we can obtain
the rest of the function for the transformation (child wavelet, Ψ a , b(t))

 
1 t −b
ψa,b (t)  √ ψ
a a

Performing the inner product of the child wavelet and f (t), we can attain the
wavelet coefficient

∞
 
wa,b  ψa,b , f (t)  ψa,b f (t)dt
−∞

We can reconstruct f (t) with the wavelet coefficient by

∞ ∞
1 dadb
f (t)  wa,b ψa,b (t)
Cψ a2
−∞ −∞

Huge estimations of the widening parameter |s| compare to little frequencies while
the little esteems relate to high-frequencies. Changing the interpretation parameter
μ moves the time limitation focus. Each μ,s(t) is restricted around t  μ. Thus,
an impeccable time recurrence depiction of capacity f can be gotten utilizing the
wavelet change. For our situation, the written by hand pictures are disintegrated
into a progression of wavelet sub-groups utilizing the Mallat calculation [40]. This
deterioration is like the one displayed in SDF for author distinguishing proof.

[email protected]
Multi-language Handwritten Recognition in DWT … 213

Symbolic Dynamic Sifting (SDF) This area quickly shows the basic ideas of Repre-
sentative Dynamic Sifting (SDF) presented by Beam [22] for extraction of highlights.
In our execution, we have connected the SDF calculation introduced to remove gen-
der particular highlights from the written work pictures. The key advances engaged
with SDF-based component extraction are examined in the accompanying.

5.1 Symbolization/Quantization of Information

The initial phase in SDF is the quantization of information (wavelet changed pictures
for our situation). For symbolization, the given information arrangement is isolated
into a limited number of cells. Every cell is then allocated a novel name or image and
the aggregate number of remarkable images is the same as the aggregate number of
cells. Accepting the image letter set to be A  {a1 , a2 , a3 , a4 } with card (A)  4, the
information succession has been partitioned into four parcels on the y-hub. These
districts are totally unrelated and additionally comprehensive (the total information
profile is secured). Each of the four districts is allotted a mark from the letters in
order A. For an information grouping under thought, the flag an incentive at a given
point in time is doled out the image comparing to the phone in which it is found. This
permits speaking to the information grouping by a limited series of images regularly
named as the image piece. Additionally points of interest on the symbolization of
information can be found in DWT (Fig. 6).
For dividing of information, systems like uniform parceling (UP) most extreme
entropy apportioning (MEP) have been examined. Uniform dividing, as the name
recommends, creates cells of equivalent size. Most extreme entropy apportioning,
despite what might be expected, endeavors to boost the entropy of the created images
by guaranteeing that every cell contains roughly a similar number of information
focuses. Thusly, the cell measure is little in data rich locales while it is extensive in
scanty areas. In the two cases, the measure of the letter set is picked as a component
of the information under examination and the objective framework execution.


C 
Ni
W  (vij − µi ) (vij − µi )T
i1 j1

Fig. 6 Loop identification

reduce noisy in SDF

[email protected]
214 T. P. Umadevi and A. Murugan

vji : j-th data vector of class i.

Given a projection matrix W (of size n by m) and its linear transformation, the
between-class scatter in the projection space is

p  WT v


C
B  Ni (µi  − µ) (µi − µ )T
i1


C
 Ni (WT µi − WT µ) (WT µi − WT µ)T
i1


C
 Ni (WT µi − WT µ) (µiT W − µT W)
i1


C
 WT Ni (µi − µ) (µiT − µT )W
i1


C
W T
Ni (µi − µ)(µi − µ)T W
i1
 WT  B W

5.2 Discrete Wavelets Transform (DWT)

The development of Discrete Wavelets Transform (DWT) depends on the supposition

that the image age process can be demonstrated as a Markov chain, named as the D-
Markov machine [19]. Advertisement arrange Markov chain is a stochastic procedure
where the likelihood of event of an image is a capacity The states in DWT speak
to mixes of words in the grouping of images while the edges compare to progress
between various states. In our work, we take D  1, subsequently, the quantity of
states are the same as the quantity of images in the letters in order. The arrangement
of conceivable states is given by S  {s1 , s2 , …, sN } with N being the aggregate the
quantity of states. The likelihood of change from state si to state sj would then be
able to be characterized as takes after

a  a0−m , b  nb0 a0−m m, n ∈ Z,

where N(si , sj ) is the aggregate number of occasions when sj seems adjoining si .

If a0  2, b0  1, the set of the wavelet

m/2
ψm,n (t)  a0 ψ(a0m t−nb0 ) m, n ∈ Z

[email protected]
Multi-language Handwritten Recognition in DWT … 215

Table 3 DWT recognition Meaningful straight lines Meaningful Arcs

parameters SDF
Horizontal line C like

Vertical line D like

Positive slanted U like

Negative slanted A like

O like

ψm,n (t)  2m/2 ψ(2m t−n)Si × S j

When all the change probabilities P (sj |si ), ∀sj , si ∈ S have been assessed, they
have gathered the N × N state progress framework. Get the maximum coefficient 
26 (Table 3).
Pseudo code
Initialize HWSC  15, TST1  10
int[] list  new int[HWSC]
for (int index  0; index < HWSC; index++)
Round Toward zero (NIST)
if sign=positive
Else if any bits to the right of the result DWT=1
Else if sign=negative
Compute Round to nearest English
Else if sign=neutral positive
FLD X=SDF (st(0)  X)
FLD Y (st(0)  Y, st(1)  X)
FLD U (st(0)  U, st(1)  X*Y)
FLD V (st(0)  V, st(1)  U, st(2)  X*Y)
FDIV TST1 ;st(0)  U/V, st(1)  X*Y
FSUB(SVM) ;st(0)  X*Y - U / V
FSTP Z DWT ;Z  result, st(0)  empty
while(!kbhit())
for(int j=0;j<50;j++)
*(p+160*j+i+1)=2
*(p+160*j+i)=a[random(15)]
for(j=0;j<25;j++)
c=*(p+160*j+i)
d=*(p+160*(j+1)+i)
end while *(p+160*(j+1)+i)=c
end if *(p+i)=

[email protected]
216 T. P. Umadevi and A. Murugan

Regardless of the way transcribed content is masterminded on the page, the asso-
ciated parts in the composition are modified into another space keeping the first
inclination yet decreasing the spaces between lines of content and segments. This
creates surface pictures which save the overall look and feel of the written work and
permits utilizing a worldwide approach staying away from the intricacy of division.
The significant strides in this procedure are recorded in the accompanying.
1. The picture is binarized utilizing worldwide thresholding and the associated parts
are separated utilizing 8-network.
2. Small parts which are probably going to relate to accentuation stamps and com-
motion are expelled utilizing region-based separating.
3. The bouncing boxes of the rest of the segments are then used to extricate the
relating parts from the grayscale picture.
4. Components in the primary line of the first picture are adjusted in another picture
utilizing the focal point of mass of the bouncing box. The inclination of the
content lines, in this way, is standardized.
After filling the primary line, the normal stature of the parts is figured and this
esteem is utilized to characterize the y-organize of the following line, which is given
as takes after


C
B  Ni (µi − µ)(µi − µ)T
i1

C number of classes
N i number of data vectors in class i
μi mean vector of class i and μ: mean vector

6 Results and Implementation

We now show the consequences of the examinations to approve the adequacy of

the proposed method in portraying gender from penmanship. To begin with, we talk
about the grouping rates on the two databases (HWSC and TST1) and later present
the consequences of various fascinating examinations Notwithstanding the proposed
highlights, we likewise assess best in class essayist distinguishing proof highlights
for gender orientation order on the two databases. These incorporate introduction
and bend highlights, fractal measurement, neighborhood paired examples (LBP) and
auto-backward (AR) coefficients. Moreover, to think about the effect of line pressing,
we figure the grouping rates without age of finished pictures (Figs. 7, 8 and 9).
Comparison study
We have compared our proposed method of implementation with various other
methods, the below table depicts that our proposed method has more accuracy than
other compared methods.

[email protected]
Multi-language Handwritten Recognition in DWT … 217

Author Methods Accuracy (%)

G. Huang, D. Zhang, X. Zheng and X. Zhu DTW 90
H. Ding RFIPad 91
N. Sharma, B. Kumar and V. Singh, SVM 92
Gaganjot Kaur, Monika Aggarwal Levenberg–Marquardt 93
Proposed method DWT 94.07

Fig. 7 Origin pixel matrix base the recognition technique

Fig. 8 KNN designed for DWT of input and maximum symbol separation edification

[email protected]
218 T. P. Umadevi and A. Murugan

Fig. 9 Order rates as an element of SDF letter set size utilizing KNN classifier

7 Conclusion and Future Works

We displayed a novel approach for gender orientation location of essayists from

pictures of penmanship styles. The method depends on a worldwide surface based
approach by thinking about each composing as a surface. The transcribed pictures are
utilized to create surface pieces which are disintegrated into a progression of wavelet
sub-bands. Each of the sub-groups is at that point stretched out into an information
arrangement that is symbolized to build a probability state automata (DWT). The
DWT is then used to create the element vector describing gender of the essayist of a
given penmanship test. For grouping, counterfeit neural systems and bolster vector
machines are utilized. The procedure was assessed on two databases containing tran-
scribed specimens in English, Cursive, and Arabic. A progression of examinations
in various testing situations demonstrated and acknowledged arrangement rates of
up to 94.07%.

[email protected]
Multi-language Handwritten Recognition in DWT … 219

For our further examinations on this issue, we intend to upgrade the list of capa-
bilities and apply a component choice method to locate the ideal arrangement of
highlights for this issue. We additionally plan to incorporate the location of other
statistic properties of scholars from manually written pictures. These may incorpo-
rate handedness (left or right), age gathering and race and so on. Connection among
penmanship and individual and scholarly properties of essayists can likewise be
contemplated.

References

1. Almusaly, I., & Metoyer, R. (2015). A syntax-directed keyboard extension for writing source
code on touchscreen devices. In 2015 IEEE Symposium on Visual Languages and Human-
Centric Computing (VL/HCC) (pp. 195–202). IEEE.
2. Alonso, M. A. P. (2015). Metacognition and sensorimotor components underlying the process
of handwriting and keyboarding and their impact on learning. An analysis from the perspective
of embodied psychology. Procedia-Social and Behavioral Sciences, 176, 263–269.
3. Doetsch, P., & Ney, H, et al. (2013). Improvements in rwth’s system for off-line handwriting
recognition. In 2013 12th International Conference on Document Analysis and Recognition
(pp. 935–939). IEEE.
4. Keysers, D., Deselaers, T., Rowley, H. A, Wang, L.L., & Carbune, V. (2016). Multi-language
online handwriting recognition.
5. Poznanski, A., & Wolf, L. (2016). Cnn-n-gram for handwriting word recognition. In Proceed-
ings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 2305–2314).
6. Capobianco, S., & Simone, M. (2017). Deep neural networks for record counting in historical
handwritten documents, In Pattern recognition letters. ISSN 0167-8655.
7. Kumar, G., & Govindaraju, V. (2017). Bayesian background models for keyword spotting in
handwritten documents. In Pattern recognition (Vol. 64, pp. 84–91).
8. Yagoubi, M. R., Serir, A., & Beghdadi, A. (2016). Joint enhancement-compression of hand-
written document images through DjVu encoder. Journal of Visual Communication and Image
Representation, 41, 324–338. ISSN 1047-3203.
9. Akbari, Y., Jalili, M. J., Sadri, J., Nouri, K., Siddiqi, I., & Djeddi, C. (2018). A novel database for
automatic processing of Persian handwritten bank checks. Pattern Recognition, 74, 253–265.
ISSN 0031-3203.
10. Vo, Q. N., Kim, S. H., Yang, H. J., & Lee, G. (2018) Binarization of degraded document
images based on hierarchical deep supervised network. Pattern Recognition, 74, 568–586.
ISSN 0031-3203.
11. Jia, F., Shi, C., He, K., Wang, C., & Xiao, B. (2018). Degraded document image binarization
using structural symmetry of strokes. Pattern Recognition, 74, 225–240. ISSN 0031-3203.
12. Chahi, A., El khadiri, I., El merabet, Y., Ruichek, Y., & Touahni, R. (2018). Block wise local
binary count for off-Line text-independent writer identification. Expert Systems with Applica-
tions, 93, 1–14. ISSN 0957-4174.
13. Mondal, T., Ragot, N., Ramel, J.-y., & Pal, U. (2018). Comparative study of conventional time
series matching techniques for word spotting. Pattern Recognition, 73, 47–64. ISSN0031-3203.
14. Agius, A., Morelato, M., Moret, S., Chadwick, S., Jones, K., Epple, R., et al. (2018). Using
handwriting to infer a writer’s country of origin for forensic intelligence purposes. Forensic
Science International, 282, 144–156. ISSN0379-0738.
15. Siddiqi, I., Djeddi, C., Raza, A., & Souici-Meslati, A. (2014). Automatic analysis of handwriting
for gender classification. Pattern Analysis and Application.
16. Liwicki, M., Schlapbach, A., & Bunke, H. (2011). Automatic gender detection using online
and off-line information. Pattern Analysis and Application, 14, 87–92.

[email protected]
220 T. P. Umadevi and A. Murugan

17. Hassaine, A., Al-Maadeed, S., Aljaam, J., & Jaoua, A. (2013). ICDAR 2013 competition on
gender prediction from handwriting. In Proceedings of IEEE 12th International Conference
on Document Analysis and Recognition (pp. 1417–1421).
18. Al-Maadeed, S., Ayouby, W., Hassaine, A., Aljaam, & Quwi, J.M. (2012). An Arabic and
English handwriting dataset for offline writer identification. In Proceedings of 13th Interna-
tional Conference on Frontiers in Handwriting Recognition (pp. 746–751).
19. Pratikakis, I., Gatos, B., & Ntirogiannis, K.: ICFHR 2012 competition on handwritten docu-
ment image binarization (H-DIBCO 2012). In 2012 International Conference on Frontiers in
Handwriting Recognition (ICFHR) (pp. 817–822). IEEE.
20. Pratikakis, I., Zagoris, K., Barlas, G., & Gatos, B. (2016). ICFHR2016 handwritten document
image binarization contest (H-DIBCO 2016). In 15th International Conference on Frontiers
in Handwriting Recognition (ICFHR), (pp. 619–623).
21. Huang, G., Zhang, D., & Zheng, X. (2010). An EMG-based handwriting recognition through
dynamic time warping2010 Annual International Conference of the IEEE Engineering in
Medicine and Biology. Buenos Aires, 2010, 4902–4905.
22. Sokic, E., Salihbegovic, A., & Ahic-Djokic, M. (2012). Analysis of offline handwritten text
samples of different gender using shape descriptors. In Proceedings of IX International Sym-
posium on Telecommunications (BIHTEL) (pp. 1–6).

[email protected]
A Novel H-∞ Filter Based Indicator
for Health Monitoring of Components
in a Smart Grid

E. Ranjini Warrier, P. V. Sunil Nag and C. Santhosh Kumar

Abstract Health monitoring of a smart grid is very important to ensure reliability of

the grid. This can be achieved by developing various indicators for the components
of the grid. These indicators are very powerful if they are model based as they can be
used in real time without need for extra hardware provided the model of the systems
and the model parameter values are available. This work presents a novel H∞ filter
based fault indicator. The fault chosen here is the stator interturn fault of a wound
rotor synchronous generator. It will be proved that (1) the indicator sensitive to fault
and is insensitive to other kinds of spurious effects like load imbalance. (2) The
indicator can be used to find the magnitude of fault. (3) The indicator can function
irrespective of the type of uncertainties assumed in modelling the system. As far as
the knowledge of the author goes this is the first time H-∞ filter based indicators are
used for stator interturn fault of a wound rotor synchronous generator.

Keywords H-∞ filter · Fault indicators · Model based · Wound rotor

synchronous generators · Smart grid

1 Introduction

Research in Smart grids is being actively pursued in various countries to cater to the
escalating and variable demand for energy. One of the functions of a smart grid is to
monitor the health of its components by means of health/fault indicators [1]. These
indicators should flag incipient faults. They should provide rich information like the
type of malfunction/fault magnitude and location of fault. In an advanced system
these indicators can be used to obtain the remaining useful life of the component.
Various machine intelligence algorithms have been proposed to derive such indicators
[2]. These can be broadly classified as model-based- and data-based indicators. Data-
based indicators [3, 4] use information-rich historical data from the systems under

E. Ranjini Warrier · P. V. Sunil Nag (B) · C. Santhosh Kumar

Department of Electronics and Communication Engineering, Amrita School of Engineering,
Amrita Vishwa Vidyapeetham, Coimbatore 641112, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 221

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_19

[email protected]
222 E. Ranjini Warrier et al.

consideration to generate the indicators. Model-based indicators use the model of the
system for the indicators. This approach uses the physical principles governing the
system to develop a model and expressed in the form of mathematical equations. State
estimation is one of the popular methods for generating indicators in the model-based
category.
The agent which performs this estimation can be categorized as state estimator.
Several approaches are existing for state estimation of systems [5]. Stochastic estima-
tors like Kalman filters and its variants assume that the external noises, disturbances
that affect the system and the “model-real system mismatches” follow Gaussian or
normal distribution. H∞ filters on the other hand solve the state estimation problem
without assuming any statistics for the uncertainties. Hence are very powerful for
generating fault indicators. These filters are also known as minimax filters.
There are various fields where H∞ filters are utilized. In [6] this filter is used
for parameter estimation in SLAM problem and in Li–Fe batteries [7], signal recon-
struction [8] and target tracking [9]. Another such major area where this filter can
be applied is to generate health/fault indicators. In [10, 11] these filters are used
constructing fault indicators for diesel engines and induction motors respectively.
Henry [12] explains how H∞ filters are used generating fault indicators for micro
strip satellite thrusters. The present work focuses on using H∞ filters for devel-
oping fault indicators of wound rotor synchronous generator here after referred to
as synchronous generators. This work can be extended to health monitoring of any
component in a smart grid. Here synchronous generators are chosen as their func-
tioning is critical in a typical power system. Further research in applying H∞ filters
for synchronous generator fault indicators is very rare.
Synchronous generators (SG) are the most common sources of Electrical Energy.
It is the primary component in almost every power system/smart grid. Since the com-
plexity and expense of power systems has gradually increased, there is less tolerance
for performance degradation and safety issues. This leads to the necessity of detect-
ing the faults in systems at the early stage and rectify it [11]. This paper deals with
the most common and critical electrical fault known as stator interturn fault that can
occur in Synchronous Generators. The residual signal, which the difference between
the actual outputs from the system and the predicted outputs from the filter/state
estimator acts as an indicator of fault [13].
The next section explains the methodology proposed for machine fault indicators
using H∞ filter. Modelling of synchronous generator is explained in brief in third
section. Fourth section explains about the H∞ filter used in this work. Results,
analysis and comparisons with the EKF filter are presented in section five. Summary,
conclusion and the future work constitutes the final section.

[email protected]
A Novel H-∞ Filter Based Indicator for Health Monitoring … 223

2 Methodology Proposed for Machine Fault Indicators

The schematic representation of the methodology used in this work is shown in Fig. 1.
A synchronous generator contains a three phase stator and a DC excited rotor. The
synchronous generator when connected to the grid is has a fixed terminal voltage
at its stator terminals and is mechanically powered with a prime mover which is
running at a constant speed. The inputs to the generator are the terminal voltages and
the DC field excitation. The outputs are the three stator currents. The state estimator
that is the H∞ filter takes the same input as the generator and provides the estimate
of the stator currents along with a residual signal. The residual signal is obtained by
subtracting the stator currents of the synchronous generator from the estimated stator
currents of the H∞ filter. The residual has to be further processed to obtain the fault
indicator which provides full information regarding the fault like time, magnitude
and location of the fault.
Here is instead of an actual generator a fault injection capable MATLAB imple-
mentation of a generator [14] is used to get the data. The stator interturn short fault
can be introduced in the simulated generator by changing a parameter μ which is
the fraction of the total number of turns that are shorted in a given phase. In this
particular implementation the fault can be simulated only in phase A. As mentioned
earlier the H∞ filter being a model-based technique needs a model of the generator
for its operation. The model used in this work is explained in the next section.

Fault injection
capable Stator currents
MATLAB
implementation
of SG

Stator
voltages and
DC
excitation

H∞ Filter
and
residual
generator

Estimated stator currents and Residual

Fig. 1 Schematic of the proposed methodology

[email protected]
224 E. Ranjini Warrier et al.

3 Modelling of Synchronous Generator

The SG as explained in [15] can be modelled as six magnetically coupled coils (three
stator coils and three rotor coils). The coils on the rotor include a field winding and
two damper windings. The state space model of the generator is given below.

X e (k + 1)  Am X e (k) + Bm U (k) + w(k)

Ye (k)  Cm X e (k) + Dm U (k) + v(k)
X e (k)  [X 1 (k)X 2 (k)Σ(k)]T ; Ye (k)  X 1T (k)
where
 
X 1 (k)  i sd (k) − μa i fd (k)i sq (k) − μb i fq (k)i s0 (k) − μc i f0 (k)
 
X 2 (k)  i rfd (k)i rkd (k)i rkq (k) ; Σ(k)  [μa i fd μb i fq μc i f0 ], (1)

where
   
A O6×3 B    
Am  ; Bm  ; Cm  I3×3 O3×6 ; Dm  O3×3 O3×3 ;
O3×6 I3×3 O3×6

O is a null matrix with the dimension indicated in the subscript and I is identity
matrix.
 −1    −1
Z1 Z2 Y1 Y2 Z1 Z2
AI+ ∗ Ts ; B  ∗ Ts
Z3 Z4 Y3 Y4 Z3 Z4
⎡ ⎤ ⎡ ⎤
−(L md + L al ) 0 0 L afd L akd 0
Z1  ⎣ 0 −(L mq + L al ) 0 ⎦; Z 2  ⎣ 0 0 L akq ⎦;
0 0 −L al 0 0 0
⎡3 ⎤ ⎡ ⎤
L
2 afd
0 0 L fdfd L fdkd 0
Z 3  ⎣ 23 L akd 0 0 ⎦; Z 4  ⎣ L fdkd L kdkd 0 ⎦;
0 23 L akq 0 0 0 L kqkq
⎡ ⎤ ⎡ ⎤
rs −(L mq + L al ) 0 0 0 L akq
Y1  ⎣ (L md + L al ) rs 0 ⎦; Y2  ⎣ 0 −L afd −L akd ⎦;
0 0 rs 0 0 0
Y3  Z 3 ; Y4  Z 4 − [diag[ rfd rkd rkq ]];
 T
U (k)  vsd (k) vsq (k) vs0 (k) vrfd (k) vrkd (k) vrkq (k)

The definition of various constants and their values used in this work is given in
Table 1. As seen from the above state space model the state variables are not directly
the stators currents so the estimates of the H∞ filter have to be processed further to

[email protected]
A Novel H-∞ Filter Based Indicator for Health Monitoring … 225

Table 1 Definition of various constants and their values used in the model
Parameter Value (pu) Parameter Value (pu)
rs stator resistance 0.018 L mq quadrature axis 0.75
inductance
rfd field winding resistance 0.015 L fdfd field inductance 1.2
rkd  rkq rotor d-axis and 0.2 L kdkd d-axis inductance 1.32
q-axis resistance
L md direct axis inductance 1.00 L kqkq d-axis inductance 1.26
L md leakage inductance 0.15 L afd , L akd , L akq stator to rotor 1.0
field, d-axis and q-axis mutual
inductance
vf field voltage 0.018 L fdkd field to d-axis mutual 1.0
inductance

obtain the actual estimates of the stator currents. The next section describes the H∞
filter algorithm.

4 H∞ Filter

H∞ filters [5] are applicable in cases where there is lack of clarity regarding the
disturbances that acts on the system and the system uncertainties. In this work,
H∞ filter is used for generating the fault indicator. The presence of ‘γ ’ in filter
equations highlights the major difference between the H∞ filter from the traditional
Kalman Filters. The H∞ filter like the traditional Kalman filter contains two steps
the prediction step and the correction step. The flow chart shown in Fig. 2 contains
the algorithm.
The state space model of the Synchronous Generator as given in the previous
section is nonlinear but the H∞ filter formulation considers the system as linear
so the system is linearized at each time step. The states of the state space model
are different from the natural stator currents. As per the schematic shown in Fig. 1
the natural stator currents are to be estimated by the H∞ filter, hence the last three
variables of the state vector are added to the first three variables to obtain the stator
currents in the dq frame and an inverse dq transformation is applied to the currents
in the dq frame to obtain the natural stator current this is done at each time step more
details can be found in [15]. The next section describes the experiments and analyses
the results.

[email protected]
226 E. Ranjini Warrier et al.

Initialize state estimates, γ and noise

covariance matrices P, Q, V and W

L=(I-γQP+ CmTV -1CmP)-1

Stator currents in
the natural frame NO
1
Eigen value of L-1P- >0

Inverse dq YES
transformation
Calculate Jacobian matrix
for the states
Stop

Addition of first
three states to last K=Ak Pk LCmTV -1
three states

Corrected state X̂ K+1=Am XK+BmUK+K(Y-Cm X̂)

Data from the simulated
P=AKPkLkAkT+W
synchronous generator

Fig. 2 H∞ filter algorithm used in this work

5 Simulation Experiment Results and Analysis

The simulated synchronous generator was run for a duration of 20 and at 10 s an

inter turn fault with μ  0.1 (10% of the turns shorted) was introduced at phase
A at 15 s and another interturn fault with μ  0.2 (20% of the turns shorted) was
introduced in phase A. The H∞ filter was used to estimate the value of the three
stator currents at each time step. The difference between the H∞ filter estimates and
the actual stator currents from the simulation also called the residual was computed.
The residual was further processed by calculating its moving variance and the RMS
value of the resultant signal called the “RMS variance” was computed. This was done
as the variance of the residual signal was sinusoidal in shape. This RMS variance
is the fault indicator. Here the moving 10,000 time points (6 s) RMS variance of
the residual was computed and plotted. The plot obtained is shown in Fig. 3. The
plot was zoomed to the time interval of interest for better clarity. From the figure, it
can be seen that at 10 and 15 s the RMS variance of phase A is different from the
RMS variance of phase B and phase C. From this we can conclude that the fault is
in Phase A. It can be concluded from the figure that the three RMS variance signals
can indicate the location of the phase in which the fault has occurred.
Another simulation experiment was done to observe the effect of load imbalance
similar to interturn fault on the residual. This is to make sure that the indicator
does not respond to other spurious effects. The synchronous generator simulation
was modified to incorporate a load imbalance and the H infinity filter was used
to compute the fault indicator. Here a load imbalance of similar magnitude of the
interturn fault was introduced at 10 and 15 s. The plot of 10,000 point RMS variance
is shown in Fig. 4. By comparing Figs. 3 and 4, it is clear that the residual is sensitive
only to interturn fault and not to the load imbalance. In the above experiments the

[email protected]
A Novel H-∞ Filter Based Indicator for Health Monitoring … 227

Fig. 3 10,000 point RMS variance of residuals with stator interturn fault

Fig. 4 1000 point moving variance of residuals with load imbalance

value of γ was taken as 0.01. It can also be seen that the magnitude of the RMS
variance depends on the magnitude of the interturn fault. To use this phenomenon
for the obtaining the fault magnitude an experiment was conducted with different
values of fault parameters and the RMS variance values were computed. The results
are shown in the Table 2. A straight line fit was obtained for the data in the table.
The equation of the straight line is

R  0.4534 × 10−3 + μ ∗ 4.253 × 10−3 , (2)

where μ represents the percentage of fault parameter and R is the RMS variance.
This equation can be used to compute the value of the fault parameter given the value
of the RMS variance.

[email protected]
228 E. Ranjini Warrier et al.

Table 2 Different fault Fault parameter Corresponding RMS variance (10−3 )

parameter values with the
corresponding RMS variance 0.1 0.8787
values 0.125 0.9641
0.15 1.055
0.175 1.134
0.2 1.304
0.225 1.365
0.25 1.569

Fig. 5 Comparison of EKF and H∞ for different kinds of noise

An experiment was conducted to observe the effect of different types of noise

on the performance of the H∞ filter for fault indication. A similar experiment was
conducted using an Extended Kalman Filter (EKF) for the same purpose. The perfor-
mance of both the filters was compared. The results in graphical form are presented
in Fig. 5. The figure shows the RMS variance values with respect to time. Some
relevant numerical data are tabulated in Table 3. It can be observed from the figure
and the tabulated data that the change in the RMS variance values compared to the
no fault case of the H∞ filter are larger than the change in RMS variance values of
the EKF. Hence H∞ filter is better suited to be a fault indicator. Further the type of
noise has not affected the performance the filter. Hence we can conclude that H∞
filter is better suited a fault indicator as it is more sensitive to the fault and performs
independent of the type of noise affecting the system.

[email protected]
A Novel H-∞ Filter Based Indicator for Health Monitoring … 229

Table 3 Numerical results Filter RMS variance (in 10−3 )

extracted from the Fig. 5
μ  0 (no μ  0.2
fault)
H∞ filter with uniform noise 0.7188 1.39
added
H∞ filter with Gaussian noise 0.7572 1.456
added
EKF filter with uniform noise 0.6268 1.1161
added
EKF filter with Gaussian noise 0.7464 1.201
added

6 Summary, Conclusion and Future Work

The work presented tests the performance of a H∞ filter based fault indicator that
can be used on any component of a smart grid. Here a synchronous generator was
chosen as the component and a fault indicator for stator interturn fault was devel-
oped. This fault was chosen as it is the basis for almost all kinds of electrical faults
in a synchronous generator. Here a simulated fault injection capable synchronous
generator was used to test the performance of the fault indicator. The fault indicator
was the RMS variance of the residual signal obtained form the H∞ filter. It was
observed that the fault indicator will respond to the fault and will not respond to
similar effects like load imbalance. Further the indicator can also be used to com-
pute the fault parameter by using the Eq. (2) obtained from the experimental data. It
was also proved that the proposed H∞ filter based fault indicator is more sensitive
to the fault than a traditional EKF based fault indicator. It is also functions in the
same manner independent of the noise and uncertainties effecting the system. The
main contribution of this work is to use H∞ filter based residual as indicator for
faults especially for a wound rotor synchronous generator. To the best of the author’s
knowledge this approach has not been tried previously. This work can be extended
by applying to other components of the smart grid provided the model of the compo-
nent along with the model parameter values that are available. Hence this approach
provides a very reliable real-time health monitoring system for a smart grid and this
can be implemented in software without any extra hardware. As a further extension
this system can be tested on a lab-scale setup [16].

[email protected]
230 E. Ranjini Warrier et al.

References

1. Zhang, H.-T., & Lai, L.-L. (2012). Monitoring system for smart grid. In 2012 International Con-
ference on Machine Learning and Cybernetics. https://doi.org/10.1109/icmlc.2012.6359496.
2. Aubert, B., Regnier, J., Caux, S., & Alejo, D. (2015). Kalman-filter-based indicator for online
interturn short circuits detection in permanent-magnet synchronous generators. IEEE Trans-
actions on Industrial Electronics, 62, 1921–1930.
3. Gopinath, R., Santhosh Kumar, C., Ramachandran, K., Upendranath, V., & Sai Kiran, P. (2016).
Intelligent fault diagnosis of synchronous generators. Expert Systems with Applications, 45,
142–149. https://doi.org/10.1016/j.eswa.2015.09.043.
4. Gopinath, R., Kumar, C., & Ramachandran, K. (2016). Scalable fault models for diagnosis of
synchronous generators. International Journal of Intelligent Systems Technologies and Appli-
cations., 15, 35–51. https://doi.org/10.1504/IJISTA.2016.076103.
5. Simon, D. (2006). Optimal state estimation. Hoboken, NJ: Wiley.
6. Ahmad, H., & Namerikawa, T. (2011). H∞ filter-SLAM: A sufficient condition for estima-
tion. IFAC Proceedings Volumes, 44, 3159–3164. https://doi.org/10.3182/20110828-6-IT-1002.
00260.
7. Yang, W., Yu, D., & Kim, Y. (2013). Parameter estimation of lithium-ion batteries and noise
reduction using an H∞ filter. Journal of Mechanical Science and Technology, 27, 247–256.
https://doi.org/10.1007/s12206-012-1203-z.
8. Vikalo, H., Hassibi, B., Erdogan, A. T., & Kailath, T. (2005). On robust signal reconstruction in
noisy filter banks. Signal Processing, 85, 1–14. https://doi.org/10.1016/j.sigpro.2004.08.011.
9. Guo, L., Mao, Y., & Song, C. (2012). H infinity filter in maneuvering target tracking of military
guidance field. In International Conference on Automatic Control and Artificial Intelligence
(ACAI 2012). \https://doi.org/10.1049/cp.2012.1173.
10. Boulkroune, B., Pages, O., Aitouche, A., Zemouche, A., & Hajjaji, A. (2014). H∞ -based fault
diagnosis for diesel engines. In 2014 IEEE Conference on Control Applications (CCA). https://
doi.org/10.1109/cca.2014.6981414.
11. Nohra, C. (2013). Online stator and rotor fault diagnosis in induction machines by H∞ observer
and sliding mode estimator. In 2013 25th Chinese Control and Decision Conference (CCDC).
https://doi.org/10.1109/ccdc.2013.6561511.
12. Henry, D. (2008). Fault diagnosis of microscope satellite thrusters using H-infinity/H_ filters.
Journal of Guidance, Control, and Dynamics, 31, 699–711. https://doi.org/10.2514/1.31003.
13. Duvvuri, S., & Detroja, K. (2015). Model-based stator interturn short-circuit fault detection and
diagnosis in induction motors. In 2015 7th International Conference on Information Technology
and Electrical Engineering (ICITEE). https://doi.org/10.1109/iciteed.2015.7408935.
14. Rabbi, A. (2016). Detection of stator interturn fault of synchronous machine by rotor current
analysis (Master thesis project, School of Electrical Engineering, KTH Royal Institute of
Technology, Stockholm, Sweden).
15. Nag, P. V. S., Kumar, C. S., Sindhu Thampatty, K.C., & Isha T. B. (2018). A modified approach
for application of augmented extended Kalman filter for stator interturn fault diagnosis of a
synchronous generator. In 4th International Conference on Electrical Energy Systems (ICEES-
18).
16. Gopinath, R., Nambiar, T., Abhishek, S., Pramodh, S., Pushparajan, M., & Ramachandran, K.,
et al. (2013). Fault injection capable synchronous generator for condition based maintenance.
In 2013 7th International Conference on Intelligent Systems and Control (ISCO). https://doi.
org/10.1109/isco.2013.6481123.

[email protected]
A Survey on Intelligent Transportation
System Using Internet of Things

Palak Patel, Zunnun Narmawala and Ankit Thakkar

Abstract An Intelligent Transportation System (ITS) can reduce traffic congestion

on roads through reduced use of private vehicles. For the same, we need to expand
existing infrastructure for the identified region, but considerable time and resources
are required to set up state-of-the-art infrastructure from scratch. However, tech-
nologies like Internet of Things (IoT) can be used with the existing infrastructure for
the design of an efficient public transportation system. This paper surveys a set of
solutions available in the literature to design of an ITS system using IoT along with
challenges and future scope for the improvement of the existing solutions.

Keywords Internet of Things (IoT) · Intelligent Transport System (ITS) · Survey

1 Introduction

Intelligent Transportation System (ITS) is a smart way of providing transportation

with the help of various technologies [1]. It deals with various modes of transportation
and traffic management system using Internet of Things (IoT). IoT is a network which
connects people, things, application, and data through the Internet to enable remote
control and management of the devices [2]. IoT works on the concept of connecting
people to connecting everything. Basically, it collects the data from sensors, actuators
and other devices to compute the data using different algorithms. The computed data
will be sent to application layer so that it is available to the end user. This put forth
requirements to design an embedded system using sensors and mobile phones with

P. Patel (B) · Z. Narmawala · A. Thakkar

Institute of Technology, Nirma University, Ahmedabad 382481, Gujarat, India
e-mail: [email protected]
Z. Narmawala
e-mail: [email protected]
A. Thakkar
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 231

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_20

[email protected]
232 P. Patel et al.

communication and computing capabilities. Intelligent transport system is mainly

useful in Metro Cities and Urban areas.
Public Transportation should be a part of the solution for sustainable development
in transport. But most of the people prefer private vehicles to travel which leads to
traffic congestion as well as high utilization of non-renewable resources. However,
in order to keep and attract more passengers, public transport must have high service
quality to satisfy and fulfill a more wide range of different customer base as needed.

1.1 Different Application Areas of Intelligent

Transportation System

Following are various fields in ITS useful to common people in their daily lives:
Advanced Traveler Information System (ATIS) provides information regarding
arrival and departure of a transportation vehicle, predicting availability of routes,
traffic congestion, pre-trip information to reduce waiting time in traffic and frustration
due to lack of information [3].
Advanced Public Transportation System (APTS) is concerned with enhanc-
ing the efficiency of public transportation using intelligent schedule management
according to the congestion level in various areas [4].
Advanced Traffic Management System (ATMS) is used by traffic regulation
authorities to monitor and control traffic flow. It is also useful to take appropriate
decision in real-time conditions to provide effective traffic management during the
emergency situations by giving higher priority to the emergency vehicles [5].
Emergency Management System (EMS) is helpful in emergency conditions like
tracking accident spots and providing services. It is effective in providing emergency
services [6].

1.2 The Current Scenario of Public Transportation

• People are not preferring public transportation these days because of unavailabil-
ity of transportation at the scheduled time. Hence, people travel by their private
vehicles that increase number of vehicles on roads. It also causes environmental
degradation.
• Due to heavy traffic, more time is required to reach the desired place. This may
result in incomplete sleep and various kind of health issues to the human and
sometimes accidents too.
• People prefer public transportation less due to non-maintenance of the public
transportation vehicle. ITS using IoT can help in monitoring and maintaining
public transport facilities which can also improve safety.

[email protected]
A Survey on Intelligent Transportation System Using Internet of Things 233

1.3 Benefits of Using Intelligent Transportation System

Nowadays more than half of the population reside in urban areas which results in
traffic congestion and wastage of environmental resources. Awareness must be spread
among people to use public transportation that helps to provide overall sustainable
development of the country [7]. If more people start using public transportation
instead of preferring their own private vehicles, it can be helpful in reducing pollution
and indirectly helpful in saving the environment.
The following facilities can be provided to the customers using ITS.
• It introduce reliable public transportation system so that people can easily rely on
the provided services.
• It has improved safety features for the betterment of people.
• Highly efficient schedule management so that people can know the exact timings
for the transportation.
• Convenience with respect to facilities, availability of seats and other services.
• A smart system can be introduced which could generate a quicker response to user’s
query and could predict the availability of seats, whether the bus is fully occupied
or still, some seats are remaining. Keeping such records may help in building a
system which could ease lifestyle of the people and can provide a comfortable
journey for travelers.

2 Literature Survey of Intelligent Transportation System

According to the prediction of Barcelona tourist guide, IoT has potential to increase
global profits by 19% with the help of effective asset utilization, employee productiv-
ity, supply chain, customer experience, and innovations [8]. There are approximately
800 million vehicles in the world which are maintained by Vehicle Dynamic Control
(VDC) system which collects the data of vehicles related to the performance and
quality of vehicles for analyzing, processing, and remote monitoring. According to
a survey, around 40% of traffic is due to parking problem [8]. Hence a smarter way
of parking is implemented in Barcelona city which helps users to locate space for
parking by their vehicles.
There is a prediction from Cisco that the number of the connected devices will
increase to 50 billion by the end of the year 2020. This leads to salable, energy saving,
and a portable devices which consist of accurate and secured architecture [9].
In developed countries, the transportation system is well-structured and cost-
effective. All stops are clean and hygienic as resources are more compared to the
population. But in India, people have a tendency to rush towards the bus as soon as it
arrives, and as a result of this, elders, women and physically unfit people fail to catch
the required bus. India is a highly populated country resulting in lack of resources
for the increasing population. In order to serve large number of people with available
resources, an effective solution must be implemented that attracts people to use public

[email protected]
234 P. Patel et al.

transportation, e.g., current location of the bus can be provided by an ITS [10–16].
An ITS may also provide the number of seats vacant in a plying vehicle in real time.

2.1 Various Existing Intelligent Transportation System [17]

Electronic Fare Collection (EFC) that automate the ticketing system of the public
transportation network and collects electronic toll at a toll station.
Highway Data Collection (HDC) provides information of traffic scenario based
on positioning devices (Global Positioning System-GPS, Global System for Mobile
communication-GSM, General Packet Radio Service-GPRS) attached to road net-
work. It also collects wind velocity data, direction, etc., to predict weather condition
for the traveler.
Traffic Management Systems (TMS) plays a central role by gathering various infor-
mation from different hardware to improve the overall efficiency of transportation.
Vehicle Data Collection (VDC) collects parameters of vehicles for predictive main-
tenance, durability, and tracking purpose.

2.2 Challenges in Vehicle Tracking Using ITS

According to Verma et al. [10], Global Positioning System (GPS) is one of the most
emerging technology and Global System for Mobile communication (GSM) network
are easily available at most places. Hence, to get current location of any vehicle,
authors have proposed a simple design using Liquid Crystal Display (LCD), GPS,
General Packet Radio Service (GPRS) modules with ATmega16 micro-controller.
They have also designed a web-based application. Using the system, one can track
vehicle remotely. This system is mainly used to track the stolen vehicle.
Farooq et al. [11] proposed a bus tracking system using Arduino, Radio-Frequency
Identification (RFID), Infrared (IR) sensors to track the bus. This user-friendly system
consists of RFID reader which is located near the gate of the bus station. When the
bus with RFID tag enters the premise, a signal is sent to control room that confirms
the arrival of the authenticated bus to the bus station. IR sensors are attached to
parking slot for the proper parking of the authenticated buses.
Mundada et al. [12] discussed a survey done in Bangalore city that shows that
IoT is evolving continuously as Broadband connections are getting cheaper. In this
system, GPS, GSM, and cameras are used for counting vehicles to get traffic situation.
Camera is attached at various nodes in the road network and with the help of Zigbee
protocol, data of the various nodes are sent to the server which can be accessed by
the consumer.
Kumar et al. [13] proposed a low-cost design for passengers using bus trans-
portation. GPS is attached to the bus so that one can track its current location. This
system consists of four modules: BUS Station Module, In-BUS Module, BASE Sta-

[email protected]
A Survey on Intelligent Transportation System Using Internet of Things 235

tion Module and BUS Stop Module. In each module, different types of equipments
like GPS, GSM, micro-controller, etc., are used to integrate information. Matrix dis-
play is attached at the station to show the location of the bus so that passengers can
know arrival time of the bus.
Hassan et al. [14] presented a case study of Tanzania, where there is no public
transport monitoring system. It reveals that many deaths and injuries occur mainly
due to passenger overloading and not following traffic rules in this East-African
country. To overcome this problem, a system is proposed for monitoring and reporting
excessive passengers in public transportation. It includes Passive Infrared Sensor
(PIR) to count passengers and uses GPS/GSM to transmit location data. The system
sends alerts to the server room if bus carries excessive passengers. This condition
delays the departure of the bus. This system provides safety to passengers and reduces
accidents.
Tarapiah et al. [16] proposed a system for vehicle tracking using GPS/GSM mod-
ule. Vibration/shock sensors are connected in air-bags of the vehicle to send an alert
to nearby hospital and police station in case of an accident. The application is com-
posed of GPRS and Google map API with various web services using HTTP REST
protocol. “Ray” algorithm is proposed to support online and offline mode of location.

2.3 Seat Occupancy Detection

Youngki et al. [18] Lee has proposed a system in which seat sensor has been used.
The system is applied on 36 seats of the library in their university. In this system, the
presence of a human is detected by the capacitive sensor which works by sending
continuous data stream in voltages, which is further connected to micro-controller.
This system works on the fundamentals of capacitive coupling where object interrupts
proximity of capacitive foil used in seats. Further, the foils are connected to capacitive
chips and attached with Raspberry pi. When foils come in physical contact, the shape
of the foil gets changed changing the capacitance. But the system gives inaccurate
results and requires frequent recalibration of the sensor. There is a possibility of
hardware failure also (Table 1).
George et al. [19] has used capacitive and inductive proximity sensor. The system
involves capacitive sensor which is able to detect the presence of humans and the
inductive sensor is used to detect all the conductors like laptops. The drawback with
this type of hybrid system is if the laptop is kept on the seat with power supply to
the system and when a person comes in contact with this laptop, then the status of
the seat changes to occupied. Table 2 summarizes various seat occupancy detection
methods found in the literature.
Zeeman et al. [20] proposed a low-cost system to measure occupancy in a minibus
taxi. In sub-Saharan Africa, there is an informal public transportation with less pas-
senger mobility. To improve understanding among passengers, the capacitive prox-
imity sensor is used to detect occupancy of seats and it is integrated with cellular
communication to provide real-time information. Capacitive sensing is a technique to

[email protected]
236 P. Patel et al.

Table 1 Summary of ITS solutions available in the literature

Paper reference Advantages Disadvantages Suggestions
Verma et al. [10] It is mainly used for It uses CMOS 8-bit Micro-controller like
tracking purpose only micro-controller Arduino can be used
which has less features for efficient results
Farooq et al. [11] Convenient for People cannot track GSM/GPS technology
passengers as they can the location of a bus can be used
track bus, also parking when it is not in the
available for buses can range of RFID
be known
Mundada et al. [12] Overcome customers’ ZigBee is having short System collects data
difficulty by providing range and integration from every vehicle
various facilities of area wise count is terminal and
time-consuming customers can get
public transit
information which
leads security issues,
hence only the result
can be displayed
Kumar et al. [13] It has various types of SMS are sent but if Real-time application
alarms like battery there is a network can be made
availability, stoppage, problem, passengers
getting late, route are unavailable to get
deviation the message
Hassan et al. [14] Monitoring and PICAXE system is Micro-controller like
reporting excessive used as an embedded Arduino can be used
passengers in public system so advanced to implement
transportation features cannot be advanced features
implemented
Tarapiah et al. [16] Vibration sensors are Useful in particular Ultrasonic sensors can
attached to an airbag. range be used to avoid
In case of an accident, accidents
it will notify
surrounding areas
within the range of the
accident

measure the capacitance between the electrode and its surrounding. The capacitance
of unoccupied seat is less than occupied seat based on various results. A mathematical
model is proposed to measure the capacitance of sensor electrode.
In a system by ETA Info-Tech in Dubai [21], passenger seat sensor are attached
in the taxi. It is made by European Union which are very high quality, reliable and
durable. It is attached to the center of the seat where the sensor can be placed flat
and its wire is passed from the backside of a seat which is integrated with meter
connection.

[email protected]
A Survey on Intelligent Transportation System Using Internet of Things 237

Table 2 Papers related to seat occupancy detection

Paper reference Advantages Disadvantages Suggestions
Youngki et al. [18] It works on the When foils come in Mechanical support
fundamentals of physical contact, its capacitive sensor can
capacitive coupling shape gets changed be used
attaching foil in a seat and gives an
inaccurate result
George et al. [19] New approach of the Complex system Only one concept can
combined sensor is be used
used to detect
occupancy
Zeeman et al. [20] Simple and low-cost Parallel plate Small mechanical
capacitive sensor capacitive sensor is sensor can be used
system is ideal for used with diameter 22 which can be studded
occupancy detection in cm and height 10 cm in seat easily
multiple-seat vehicles
accurately
ETA Info-Tech. [21] Simple and secure Sensors need to be Better quality of seat
system placed on flat surface. sensors should be used
They get damaged like capacitive sensors
under pressure in an
uneven surface.
Further, current ratings
must not vary

2.4 Traffic Monitoring

The government of Gujarat [22] has implemented City Surveillance and Intelligent
Traffic Monitoring System wherein cameras are used to survey movement of vehi-
cles. If someone breaks the rules, the camera will detect vehicle’s number plate
and E-challan will be sent to the registered address of the vehicle. These cameras
are connected via optical fiber cables or wireless nodes which sends signals to the
monitoring room. Table 3 summarizes few traffic monitoring ITS solutions.
World Sensing System [23] focuses on city mobility management including var-
ious traffic flows, smart parking, etc. It uses real-time information to make smarter
and faster decisions. In Barcelona, smart parking system is available where an elec-
tromagnetic sensor is placed in each parking slots. When a vehicle enters that slot,
the sensor will send signals of occupancy. This status can be viewed in-app which
guides users to proper parking space.

[email protected]
238 P. Patel et al.

Table 3 Existing implementations for traffic monitoring

Paper reference Advantages Disadvantages Suggestions
The government of Data capture by the Issuing E-Challan is a By doing automatic
Gujarat [22] camera are sent manual process number plate
through fiber optics or recognition, it can be
wireless network and automated
if someone breaks
rules then E-Challan
gets generated
World sensing system Electromagnetic Electromagnetic Use of other sensors
[23] sensors are used to sensors sometimes like IR, ultrasound,
detect the arrival of miss objects like steel etc.
cars ramps and also deflect
acoustic waves

3 Future Research Directions

No research is whole and complete. There are chances to improve the existing
research work. The literature review presented so far have implemented ITS with
different features and each of them has been implemented in isolation. There is a
strong need to design an ITS by integrating all these features in a single system that
caters the need of different users. Safety, accuracy, and cost-effectiveness can be con-
sidered as parameters of interest while designing smart ITS. Also below challenges
faced by IoT can be considered:
• Security: As the number of devices will be connected, their security will be a big
challenge. Also, centralized IoT server needs to be secured.
• Connectivity: If the number of devices is moderate then the system is manageable.
But, if billions of devices are involved then managing network is a challenging
task.
• Compatibility: Every device connected for IoT is diversified in its own protocol
and operating system, which needs to be synced in order to make it usable for IoT
architecture.
For an IoT architecture, following basic hardware is required:
• Sensors: To fetch data which would be on the sensor layer of an architecture
• Micro-controller/ Microprocessor: To integrate the data. This would be in com-
munication layer of an architecture.
• Device: To display data. This would be in an application layer of an architecture.
This technology leads in the business world with major impact on the environment.
A better world can be achieved by optimizing the use of IoT technology by solving
its open challenges. Most of the people are dependent on the public transit. Hence,
a traveler should get the level of convenience same as his/her private vehicle, i.e.,
amount of waiting time should be minimal and assurance of a seat in the public

[email protected]
A Survey on Intelligent Transportation System Using Internet of Things 239

transport vehicle. Tracker sensors will be able to track the vehicle, seat sensors can
track the number of people through which prediction of the crowd can be calculated.
Gateways and mediators can be used to merge various sensors and actuators which
leads to storage of data in a particular location via signals or communicators. This
leads to a better approach for an Intelligent Transportation system.

4 Conclusion

Public transport has the potential to reduce traffic congestion if it is implemented in

intelligent and efficient manner. Existing efforts for the same lack a holistic approach.
There are numerous transit system ideas using low-cost IoT technology but each of
them lacks in consistency. A complete architecture for Public Transport System using
IoT should be devised and standardized for its wide adaptation.

References

1. van Arem, P. B. (2015). In Intelligent Transportation System. Intelligent Transportation System

Society.
2. Saleh, I. (2012). Internet of Things. Department of Civil Engineering, Wiley.
3. Khattak, A. J., Targa, F., & Yim, Y. (2004). Advanced traveler information systems.
4. Uchimura, K., Takahashi, H., & Saitoh, T. (2002). Demand responsive services in hierarchical
public transportation system (Vol. 51).
5. Hancke, G. P., & Nellore, K. (2004). Advanced traffic management system
6. Singh, A. G. B. (2015). The journal of Transport Literature. International Transport Planning
Society.
7. Das, J. H., & Tom, S. (2016). Futuristic intelligent transportation system architecture for sus-
tainable road transportation in developing countries
8. Barcelona tourist guide. 2015.05.04.
9. Yaqoob, I., Ahmed, E., Hashem, I. A. T., Ahmed, A. I. A., Gani, A., Imran, M., & Guizani, M.
(2017). Internet of things architecture: Recent advances, taxonomy, requirements, and open
challenges (Vol. 1). IEEE Wireless Communications.
10. Verma, P., & Bhatia, J. (2013). Design and Development of GPS-GSM based tracking sys-
tem with google map based monitoring system. International Journal of Computer Science,
Engineering and Application, 3.
11. Vidyasagar, K. S. K., & Farooq, M. A. (2015). Public road transport guiding system using
arduino microcontroller. International Journal of Computer Applications, 132, 12–16.
12. Mundada, M. R., & Selvapriya, P. R. (2015). IoT based bus tranport system in banglore.
International Journal of Engineering and Technical Research, 3.
13. Kiran Kumar, G., & Prasad, M. (2012). Public Transportation Management Service using
GSM-GPS. International Journal of Research in Computer and Communication Technology,
1.
14. Machuve, D., Hassan, K., & Sam, A. (2013). A system for monitoring and reporting excessive
passengers in public buses case study: Tanzania. International Journal of Engineering And
Computer Science, 1.
15. Rane, K. P., Dukare, S. S., & Patil, D. A. (2015). Vehicle tracking, monitoring and alerting
system: review. International Journal of Computer Applications, 119.

[email protected]
240 P. Patel et al.

16. AbuHania, R., Tarapiah, S., & Atalla, S. (2013). Smart on-board transportation management
system using GPS/GSM/GPRS technologies to reduce traffic violation in developing countries.
International of Digital Information Communication.
17. Abdullah, A. H., & Qureshi, K. N. (2013). A survey on intelligent transportation system.
Middle-East Journal of Scientific Research, 15.
18. Youngki, L., Rajesh Krishna, B., Nguyen Huy, H. H., & Hettiarachchi, G. (2013). Small scale
deployment of seat occupancy detectors (Vol. 1). Singapore Management University.
19. Bretterklieber, T., Brasseur, G., George, B., & Zangl, H. (2010). A combined inductive capaci-
tive proximity sensor for seat occupancy detection. IEEE Transactions on Instrumentation and
Measurement, 59.
20. Booysen, M. J., & Zeeman, A. S. (2014). Simple capacitive seat sensing for occupancy detection
and passenger counting in minibus taxis. Research Gate.
21. Eta-info tech. (2016). Retrieved from http://www.etastarinfotech.com/products/
fleetmanagement/vehicle.htm. Accessed December 03, 2015.
22. City surveillance and intelligent traffic monitoring system. No. Sep 2016. Accessed December
03, 2015.
23. World sensing system. http://www.worldsensing.com/industries/cities-governments/, 2016.
Accessed December 03, 2015.

[email protected]
CRUST: A C/C++ to Rust Transpiler
Using a “Nano-parser Methodology”
to Avoid C/C++ Safety Issues in Legacy
Code

Nishanth Shetty, Nikhil Saldanha and M. N. Thippeswamy

Abstract CRUST, a language translator (transpiler) has been developed which

converts programs in C/C++ to programs in Rust. Rust created by Mozilla has become
popular as a systems programming language as it provides constructs and tool sup-
port for developing safe and secure programs without hassles. Safe Rust ensures
that the programs developed using these programming constructs are safe and secure
by enforcing compiler restrictions, concurrency without data races, memory safety
without garbage collection, and abstractions with low overheads. CRUST is a semi-
automated transpiler, to automatically convert a subset of “C/C++” code base into
Rust without much effort. This is done using a unique Nano-Parser Methodology.
CRUST also enhances the readability and understandability of the program by adding
extensive documentation to the translated code. It yields a code which is guaranteed
to be thread safe and proven to be faster.

Keywords C++ · Memory safety · Nano-parser methodology · Rust · Transpiler ·

Type safety

1 Introduction

C and C++ programming languages provide support for easily and efficiently inter-
acting with the underlying hardware using high level program constructs, when
compared to most of the other high level programming languages [1–3]. However, it
takes a lot of effort in order to ensure that the developed programs are safe and secure
as there is no tool support to enforce this during compilation. This has to be handled

N. Shetty · N. Saldanha (B) · M. N. Thippeswamy

Department of CSE, Nitte Meenakshi Institute of Technology, Bangalore 560043, India
e-mail: [email protected]
N. Shetty
e-mail: [email protected]
M. N. Thippeswamy
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 241

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_21

[email protected]
242 N. Shetty et al.

by the programmer, which is generally a tedious task and such patches added to the
code as an afterthought are generally buggy.
Programming mistakes such as buffer overflow, null pointer dereferences, and
incorrect type casts causes the system to break down. Moreover, the lack of safety
restrictions can be exploited by an attacker to compromise the security of these sys-
tems [4, 5]. These issues that need to be addressed in order to develop safe and secure
programs have been studied extensively and numerous solutions have been proposed
to alleviate them [6–10]. However, all these require quite a bit of manual intervention.
Nowadays, there has been a dramatic rise in the popularity of functional programming
languages such as Clojure, Haskell, OCaml, Scala, and support for functional pro-
gramming principles in existing languages such as JavaScript (ECMAScript 6), Java
8. Functional Programming Languages by principle, solve concurrency and memory
issues which plague programs written in C and C++. Rust is at its heart a functional
programming language, which has mechanisms to ensure type and memory safety
at its very core.
Rust, is Systems programming language designed by Mozilla to achieve safe
systems and painless concurrency. The result is safe programming with highly
enforced compiler restrictions, memory safety without garbage collection, and zero-
cost abstraction. Rust achieves what C/C++ does not, by enforcing safety through
compiler restrictions. Due to these restrictions, it is guaranteed to be safer and more
efficient. It is a concept of borrowing and ownership that creates a well-defined object
scope, which avoids disaster.
In Rust, the problem of data race is overcome by employing mutable and
immutable references, where only mutable references can be borrowed. These along
with a few other principles have been built into Rust so that the programmer does
not have to take care of it explicitly. There is sufficient proof that Rust combines the
best of both worlds, low-level control and type/memory safety [11–13].
There is an opportunity to permanently solve problems associated with systems
programming by simply moving from C and C++ to Rust, which is a safer and
more predictable language. CRUST provides the means to do this efficiently and
effortlessly. Though we do not claim to be able to convert every construct, a fairly
large subset of code has been tried and tested. The sections that cannot be converted
currently, are well documented so that the user can manually convert these to a
Rust equivalent code with the aid of the information provided in the associated
documentation.
In the process of realizing a tool such as CRUST, we have devised an elegant
manner in which to build a Parser. Aptly named, The Nano-Parser methodology,
it involves designing several cooperatively functioning tiny parsers (nano-parsers)
each designed for very specific grammars. Using one master parser calling the nano-
parsers, it is able to set off a chain of nano-parser calls, ending in the parsing of the
complete input.
Once CRUST realizes its full potential, legacy code bases that have been aban-
doned because of their sheer size and safety may be transpiled and deployed.
In this paper, the novel contributions are as follows:

[email protected]
CRUST: A C/C++ to Rust Transpiler … 243

• Translation of unsafe legacy C/C++ code into safe and manageable Rust code with
minimal manual involvement.
• We augment the translated code with helpful documentation to ease the process
of understanding the Rust code for those who are unfamiliar with the language.
• Provision of different translation modes controlling memory safety and package
management.
There are certain assumptions made in this paper, they are the following:
• The C/C++ code is syntactically correct.
• The external API calls must be handled manually after translation.
• C/C++ code does not include any headers or modules.

2 Related Work

In this section, languages, which are considered as type safe are discussed. Type
safety is ensured to a large extent by static analyses of the source code as a part of
the compilation process, and traps are set for capturing and handling runtime type
exceptions that might occur dynamically [2]. As C/C++ compilers, in general, have
no such built-in enforcement mechanisms, the programmer has to manually ensure
that such conditions do not arise by performing code reviews before compilation and
also augment the code by enveloping the section of code with type checkers inserted
into the code to catch type mismatches dynamically. This is definitely a cumbersome
and error-prone process.

2.1 Smart Pointers

The use of ‘smart pointers’ is proposed in [6] to solve issues regarding memory
leaks associated with deallocated ‘raw pointers’ (regular pointers). The wrapper
class has attributes and methods that help to manage the lifetime of an object by
building in reference counters. Even with these mechanisms in place adoption of bad
programming practices like cyclic references could still lead to memory leaks.

2.2 Probabilistic Memory Safety

The authors in [7] have proposed a runtime system called “DieHard”, which achieves
memory safety by using probabilistic methods of randomization and replication and
creating a virtual infinite-sized heap. Randomization and Replication increase the
safety of the program as the effect of errors across the replicated versions will have
a very low probability of corruption of the same variable/entity in the replicated
versions.

[email protected]
244 N. Shetty et al.

2.3 Memory Safety

The work done in [8] proposes the memory safety for embedded software without
Garbage Collection or Runtime Checks. The authors propose a pool of free memory
from which the random allocation of objects and their replicates can be done. This
will minimize the probability of runtime errors affecting the replicated versions in
the same manner and hence enhances the safety of the system.

2.4 Manual Safe Memory Management

Gay et al. have proposed HeapSafe, which is a tool that has been developed by them
in order to trap “bad” memory free calls in programs written using “C” language.
Runtime checks are performed in order to identify dangling pointers and free memory
locations and also ensure that “bad” free function calls are treated mostly as no-ops.
Standard techniques of reference counting are adapted to verify the C programs and
also to ensure their reliability when manual memory management techniques are
adopted [9]. It relies on an extension of the malloc and free APIs. HeapSafe cannot
introduce new bugs since it only checks the existing memory management API
rather than augmenting or changing it [9]. Depending on the programs the HeapSafe
scheme can incur an execution time overhead of about 30% and designers might get
discouraged from using it for real-time systems with tight budgets.

2.5 Ironclad C++

Ironclad C++ [10] is a subset of the C++ language, which contains constructs of C++
which can be verified to be safe by static syntactic analysis of the program segments,
and also uses the concept of “smart pointers” to ensure freedom from runtime errors
that might be introduced by using simple pointers. Nearly 50,000 lines of code have
been successfully translated to Ironclad C++ [10].

3 Rust as a Successor to C/C++

In this paper, we believe that the solutions described in the previous are all valid, they
do not provide a complete solution and they each come with their own drawbacks.
Programmers, who prefer C/C++ do so because of the control it provides, which is
taken away when any of these solutions are used.
Figure 1 shows the trade-off between low-level control and safety for various
programming languages. It describes how Rust combines low-level control along

[email protected]
CRUST: A C/C++ to Rust Transpiler … 245

Fig. 1 Tradeoff between control and safety of popular languages

with safety to solve the safety implications that come with using C like languages.
Languages like Haskell tried to provide the kind of safety that is desired for hardware
applications. A Go is comparable to C and C++ in terms of low-level control and
efficiency, but is not so safe as Haskell. This is where Rust comes in with its compile-
time enforcement of safety rules and C like low-level control. It does prove that Rust
is a worthy successor to C and C++.
It is believed that these results indicate that Rust has overcome the drawbacks that
plague C/C++ while still allowing low-level control and efficiency that C is popular
for. There is nothing to lose from moving on from C/C++ except for an initial learning
curve more importantly in the context of this paper, translating an existing C code
base. CRUST solves both these problems by allowing one to translate C/C++ to safe
Rust and also provide formatted and documentation in the translated code to ease
the learning curve.
In this subsection, the design of Transpiler is discussed. Transpiler essentially is
constructed out of two functional blocks, which are executed sequentially in order
to translate the program written in “C/C” to “Rust”. Typically, like any other com-
piler, the main blocks are the Lexical Analyzer and the Parser followed by the Code
Generator. The “Match” function which is provided by Rust is used. Multiple DFAs
are launched in parallel with each constructed to recognize a template of a valid
regular expression belonging to the language. The longest lexeme will be accepted
and tokenized. It is assumed that the input source code is syntactically correct. These
identified lexemes or tokens are inserted into the Symbol Table. The Parser is mainly
composed of two modules: “Global Construct Identifier” (GCI) and “Nano-Parser”
(NP). The Lexeme streams as delivered by the Lexer are taken as input to the GCI.

[email protected]
246 N. Shetty et al.

Fig. 2 Design of the parser

with the nano-parser
methodology

Based on an initial analysis of the streams, it identifies the possible candidate gram-
matical expressions of the chosen subset of C/C++ language. Once this is identified,
the GCI triggers the appropriate NP. Each NP generates a code snippet in “Rust”
corresponding to the C/C++ source code snippet which was passed onto to it by
the GCI (see Fig. 3a, b). The Rust code snippets so generated are then collated to
compose the program in Rust which is equivalent of the input C/C++ source code.
The parser itself has been architected as a “recursive descent parser”.
We believe that this scheme called “The Nano-Parser Methodology” of using a
GCI along with specific NPs is a novel and efficient way of building compilers. Rust
inherently supports the concept of mutable and immutable variables. This allows
restrictions to be imposed on the type of functions that can be performed on the
variables by the functions to which this variable is passed. Therefore, strict control
can be enforced on the scope and the operations on variables by the functions to which
it is exposed. In order for the user to exploit this feature of Rust to enhance the safety
of the program, an option has been given where the user can tag the particular variable
as mutable or immutable. This is handled by the Mutability Analyzer (see Fig. 2). This
analyzer is invoked during the transcription process and identifies operations which
violate the immutability property tagged to a particular variable. These operations
are flagged and appropriate error messages are generated for information to the
programmer.
In this context, two modes of translation have been proposed:
• Strict mode: CRUST will treat all variables as immutable and identifies all oper-
ations which violate this criterion.
• Loose mode: CRUST will treat all variables as mutable (Fig. 3).

[email protected]
CRUST: A C/C++ to Rust Transpiler … 247

Fig. 3 a Design of the parser with the nano-parser methodology, b global construct identifier
algorithm

4 Nano-Parser Scheme

This scheme which has been defined above which comprises of a GCI and sev-
eral Nano-parsers working cooperatively is quite novel. Nesting of Nano-parsers is
allowed in order to support complex programming snippets. For example, an “if”
statement inside a “for” loop would be processed in the following manner. The GCI
based on the “for” lexeme triggers the NP associated with it. As this NP is process-
ing and translating the said input the “if” statement is encountered. This will lead
to the “for-NP” triggering the “if-NP” for further processing leading to the nesting
of the NPs (see Fig. 2). Recursive triggering can occur in case an “if” statement is
encountered inside another “if” statement.
Salient features of this methodology are listed below:
(a) As the design of the “Transpiler” is highly modular in nature, it is very easy to
implement the same.
(b) The modular design enabled test coverage of about 95% of the code in Unit
testing phase itself and the rest were checked after integration. This highly
simplified the task of testing the transpiler implementation.
(c) The modular nature of the implementation helps in easily identifying and local-
izing the errors.
(d) Upgrades and extensions to the allowed set of language constructs can eas-
ily be accommodated by building an “NP” specific to this new construct and
integrating this with the GCI.

[email protected]
248 N. Shetty et al.

5 Results and Analysis

This section provides sample results for the proposed CRUST. CRUST is able to
translate a subset of legacy C/C++ code into Rust code, with appropriate hints in
case of compile errors. The portions of the code which do not conform to the defined
translatable subset will be retained as such but flagged so that the user can translate
it manually.
Subsequently, when the translated code is passed onto the Rust compiler, it will
identify statements which are likely to violate the “safety” norms. This whole process
enables an experienced “C/C++” programmer to write “safe and secure” programs
without having to become proficient with a new programming language like “Rust”.
A large number of C/C++ programs have been translated by us using CRUST to
prove the translation process. Table 1 shows a few “C” program constructs and the
equivalent “Rust” program statements.
In Table 2, we have shown the “C” code developed for the implementation of the
matrix addition function and have also shown the translation of the same in both the
“strict” and “loose” modes. The “For loop” construct is not supported in Rust, and
must be implemented using “while loops”. The documentation that is inserted into
the code has been omitted for brevity. It can be observed that the keyword “must” is
added in the declaration of variables in the loose section. This allows modification
of the variables.
In Sl. No. 2, we see a function being converted to Rust. Sl. No. 3 shows the
conversion from C++ classes to Rust equivalent. There are no classes in Rust and
hence this must be taken care of using structures with “impl” block. The constructors
must be handled manually using static builder functions since Rust structures do not
have constructors.
The results show that very little manual translation is required in compiling the
translated code. This is very easy to do as the translated code contains appropriate
suggestions and additionally the Rust compiler generates very precise error messages.
The resultant Rust programs are very efficient and optimized. These processes gen-
erate “safe” programs without appreciably compromising on the execution speed.
With no aliasing and mutation, Rust avoids memory leaks and data races.
The code was running a “for” loop for 109 iterations. The output code was tested
by adding output statement to Rust code, I/O is a costly operation but Rust excelled.
It performed better than C++ code without any I/O doing the same task. C++ program
takes about 2.3 s to run while the translated Rust completes the same program in

Table 1 Comparison of the Timing statistic Average performance

performance of C/C++ code
and its equivalent in rust, C/C++ Rust
translated using the crust Real time (s) 2.3 0.005
User time (s) 2.3 0.002
System time (s) 0.006 0.002

[email protected]
CRUST: A C/C++ to Rust Transpiler … 249

Table 2 C/C++ code translation to Rust using CRUST in strict and loose modes
Sl. No.
C/C++ Code Rust Equivalent Code

Strict Translation Loose Translation

int main() fn main() {

{ fn main() let mut i: i32 = 0;
int i=0; { let mut a: [i32; 4] = [1, 2, 3, 4];
let i: i32 = 0;
int let a: [i32; 4] = [1, 2, 3, 4];
let mut b: [i32; 4] = [1, 2, 3,
a[4]={1,2,3,4}; let b: [i32; 4] = [1, 2, 3, 4]; 4];
int b[4]={1,2,3,4}; let c: [i32; 4]; let mut c: [i32; 4];
1. while i < 4
int c[4]; while i < 4 {
{
for(;i<4;i++) c[a]=a[i] + b[i] + 100;
c[a] = a[i] + b[i] + 100;
{ i += 1; i += 1;
c[a]=a[i]+b[i]+100; } }
} } }
}
int add_it(int a,int b) fn add_it(a: i32, b: i32) ->i32 {
{ a+b
2.
return a+b; }
}
struct A {
class A a: i32,
{ b: i32,
int a; }
int b; impl A
private: float getfloat() {
{ /** Rust structures do not support constructors
return 1.23; * Please handle them with static builder functions* >>> A ( ) { a
} =5;b=6;
3. public: int getInt(int a) }**/
{ fn getfloat(&self) -> f32
return a; {
} 1.23
A() { }
a = 5; pub fn getInt(&self, a: i32) -> i32
b = 6; {
} a
}; }
}

about a 1/2000ths of a second, a 1000 times improvement in runtime speed (see

Table 2).

6 Conclusion

The evaluation results prove that we are able to successfully translate the given C/C++
code into Rust. Rewriting existing codebase in Rust requires lots of time, money, and
effort. With the help of an automated translation tool, we can get this task done with
minimal cost and time. CRUST’s translated code is easy to debug with suggestions
and hints provided in the translated code. The code is safe and runs faster than the
original C/C++ code. Performance tests have shown that Rust runtime is very efficient
and much smaller than that of the source C/C++ code. This is yet another reason to
shift to Rust. The use of the Nano-Parser Methodology in the implementation of
parser makes it very easy to implement and test the Parser modules. This approach
may lead to significant changes in parser design and implementation.

[email protected]
250 N. Shetty et al.

References

1. Stroustrup, B. (2012). Software development for infrastructure. Computer, 45(1), 47–58. https://
doi.org/10.1109/MC.2011.353.
2. Shen, R. (2017). Why-is-C++-important. Retrieved from https://www.quora.com/. Accessed
April 11, 2017.
3. ONeal, B. (2017). What is the role of C++ today? Softwareengineering.stackexchange.com.
Retrieved from http://softwareengineering.stackexchange.com/questions/61248/what-is-the-
role-of-c-today. Accessed April 11, 2017.
4. Pincus, J., & Baker, B. (2004). Beyond stack smashing: Recent advances in exploiting buffer
overruns. IEEE Security and Privacy Magazine, 2(4), 20–27.
5. CSE341 Lecture Notes 26. (2017). Unsafe language (C). Courses.cs.washington.edu.
Retrieved from http://courses.cs.washington.edu/courses/cse341/04wi/lectures/26-unsafe-
languages.html. Accessed 11 April, 2017.
6. Colvinand Adler, D. (2012). Smart pointers—boost 1.48.0. Boost C++ libraries. Retrieved from
www.boost.org/docs/libs/1_48_0/libs/smart_ptr/smart_ptr.htm. Accessed 11 April, 2017.
7. Berger, E., & Zorn, B. (2006). DieHard. ACM SIGPLAN Notices, 41(6), 158–168.
8. Dhurjati, D., Kowshik, S., Adve, V., & Lattner, C. (2003). Memory safety without runtime
checks or garbage collection. ACM SIGPLAN Notices, 38(7), 69–80.
9. Gay, D., Ennals, R., & Brewer, E. (2007). Safe manual memory management. In International
Symposium on Memory Management.
10. DeLozier, C., Eisenberg, R., Nagarakatte, S., Osera, P.-M., Martin, M. M. K., & Zdancewic, S.
(2013). Ironclad C++: a library-augmented type-safe subset of C++. MS-CIS-13-05. University
of Pennsylvania CIS.
11. Turon, A. (2017). Fearless concurrency with rust—the rust programming language
blog. Blog.rust-lang.org. Retrieved from https://blog.rust-lang.org/2015/04/10/Fearless-
Concurrency.html. Accessed 11 April, 2017.
12. Turon, A. (2017). Abstraction without overhead: Traits in Rust—The Rust programming
language blog. Blog.rust-lang.org. Retrieved from https://blog.rust-lang.org/2015/05/11/traits.
html. Accessed 11 April, 2017.
13. Turon, A., Matsakis, N. (2017). Stability as a deliverable—The Rust programming language
blog. Blog.rust-lang.org. Retrieved from https://blog.rust-lang.org/2014/10/30/Stability.html.
Accessed 11 April, 2017.

[email protected]
Species Environmental Niche
Distribution Modeling for Panthera
Tigris Tigris ‘Royal Bengal Tiger’ Using
Machine Learning

Shaurya Bajaj and D. Geraldine Bessie Amali

Abstract Biodiversity loss due to habitat degradation, exploitation of natural

deposits, rapid change of environment and climate, and various anthropogenic phe-
nomenon throughout the last few decades in the quest of development have led to rise
in safeguarding species ecological domain. With natural habitat of the endangered
Panthera Tigris Tigris fast declining, coupled with factors such as loss in genetic
diversity and disruption of ecological corridors, there is an urgent need to conserve
and reintroduce it to newer geographic locations. The study aims to predict and model
the distribution of the species Panthera Tigris Tigris by combining various climatic,
human influence, and environmental factors so as to predict alternate ecological niche
for the already dwindling tiger habitats in India. 19 Bioclimatic variables, Elevation
level, 17 Land Cover classes, Population Density, and Human Footprint data were
taken. MAXENT, SVM, Random Forest, and Artificial Neural Networks were used
for modeling. Sampling bias on the species was removed through spatial thinning.
These variables were tested for Pearson correlation and those having coefficient
greater than 0.70 were removed. Kappa statistic and AUC were used to study the
results of the methodology implemented. Testing data comprises 25% of the presence
only points and test AUC value of MAXENT was found to be the highest at 0.963,
followed by RF at 0.931, ANN at 0.906, and lastly SVM at 0.898. These indicated
a high degree of accuracy for prediction. The most recent datasets were taken into
consideration for the above variables increasing accuracy in both time and spatial
domain.

Keywords Artificial neural networks · Bioclimatic · Elevation · Environmental

niche · Human influence · Land cover · MAXENT · MODIS · Panthera Tigris
Tigris · Random forest · Royal Bengal tigers · Species distribution modeling ·
Support vector machines

S. Bajaj (B) · D. Geraldine Bessie Amali

SCOPE, Vellore Institute of Technology, Vellore 632014, Tamil Nadu, India
e-mail: [email protected]
D. Geraldine Bessie Amali
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 251

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_22

[email protected]
252 S. Bajaj and D. Geraldine Bessie Amali

1 Introduction

As a result of various human activities, over-exploitation of natural resources and

climate change, various conservation measures need to be taken for the already dwin-
dling species of flora and fauna. The necessity of habitat suitability estimation and
species distribution in an ecological niche is closely linked to the conservation and
management of biodiversity. Ecological Niche can be described as a set of envi-
ronmental circumstances which are not hostile for species persistence and offspring
production [1]. A productive method for knowing the geographical locations so as to
reintroduce the species is species distribution models (SDMs), known by different
terms such as habitat suitability models, climate envelope models, and ecological
niche models. The main aim is to detect environmental circumstances which are
suitable and at the same time not hostile for the particular species to persist using
species occurrence (presence) data from biodiversity sources as well as several other
factors for the prediction of an output (response) variable, such as habitat suitability
for a geographical location. Satellite data with high resolution together with other
factors are seen to increase model accuracy [2].
Different SDMs have their own benefits and drawbacks. Performance varies
depending on various models, the variables chosen as well as the data chosen (pres-
ence only/presence–absence data). The behavior of different SDMs and the influence
of various predictor sets on them need to be studied. Performance of the models can
be measured through parameters such as area under the receiver operating character-
istic curve (AUC), overall accuracy, sensitivity, specificity, and Kappa statistic most
widely used being Kappa and AUC. The research uses AUC and Kappa statistic to
measure performance.

1.1 Motivation

The study to locate new geographical potential habitats is motivated by various

causes. First, declining prey bases and human activities such as deforestation,
encroachment, and development have led to a decline in Royal Bengal Tiger’s nat-
ural habitat. At geographical locations where tigers are increasing in number, there
is a tussle for territory among the tigers themselves [3]. Another motivation is to
maintain genetic diversity. As per Wildlife Institute of India and studies conducted
by the Centre for Cellular and Molecular Biology, the country is home to three
different tiger populations which are genetically connected. They thrive in different
parts of India mainly the Terai-Duar Savanna, the northeastern region of India, South
India, and parts of north India especially the Ranthambore. It was found that among
these, the Ranthambore Tiger population has highest threat of isolation and least
genetic diversity. This is verified by another study [3]. Thus, the species requires
genetic flow. The third major motivation is disrupted ecological corridors. Research
indicates that the population of Tiger species outside their natural habitat is 35%

[email protected]
Species Environmental Niche Distribution Modeling … 253

[3]. Thus, it is the ecological corridors between the natural habitats that maintain
the flow of the gene pool. Thus, it becomes necessary to link these habitats through
ecological corridors.

2 Related Works

Although the concept of SDMs has been used in modeling habitat suitability for
various species yet little work has been done for Panthera Tigris species. Kywe
[4] performed habitat suitability modeling in the Hukaung Valley Tiger Reserve,
Northern Myanmar. Covering a landscape of around 1700 km2 , the research was done
for the year 2003. Land classification dataset was developed through segmentation
based classification leading to 14 land cover classes being developed. Additional data
in the form of topographical and human influence were taken from other resources.
The research found that the species of tiger usually are found close to river beds
and grass area while avoiding evergreen closed areas and settlements. The reserve
was found to have 42% suitable areas for tiger habitats which was done with the
help of Ecological Niche Factor Analysis. Hernandez et al. [5] found Maxent to
give highly accurate results unbiased of the count of instances of species and the
extent of the study area or geographical distribution in comparison to Mahalanobis
and Random Forests models. Phillips et al. [2] made use of Maxent to model the
potential geographical locations of tiger populations during the time period of Late
Pleistocene and Holocene, thus providing new insights into the evolutionary history
and interconnectivity between populations of this endangered species. Kulloli and
Kumar [6] found out the potential habitats for Commiphora wightii (Arnt.) Bhand
in the arid zones. The SDM was implemented using MAXENT distribution taking
bioclimatic, NDVI, and elevation data in consideration. These variables were tested
for correlation and those having coefficient values >0.80 were removed. AUC was
greater than 0.9 in all the test cases with different predictor variables which indicates
that prediction precision is high. Jaisalmer and Barmer regions of Northwest India
were predicted to be geographical areas suitable for the species to be introduced.
Maxent can be modeled on both presence and pseudo-absence data, utilizing both
categorical, and continuous dataset. It uses an exponential method to predict val-
ues and thus needs additional attention when we shift to an alternate geographical
location or a time stamp. Duan et al. [7] considered five species of trees consisting
of one coniferous and four broad-leaved species. These were the Quercus wutais-
hanica, the Japanese white birch, the Mongolian oak, the Chinese cork oak, and P.
massonia which then made use of 13 different environmental variables to predict
alternate habitats. A total of six different species distribution model mainly were
then compared and their accuracy was evaluated using the coefficient of variation,
Kappa statistic, and area under the operating curve. The map closest to the mean
value of the curve was transformed into a binary map and displayed. Qing et al. [8]
implemented a Maxent model making use of presence-only records for predicting

[email protected]
254 S. Bajaj and D. Geraldine Bessie Amali

the winter ecological niche for the scaly-sided merganser in China. AUC indicated
high accuracy for both training and testing cases.

3 Methodology for Alternate Ecological Niche Prediction

3.1 Study Species and Distribution Data

The tiger species is classified in the genus Panthera with the lion, leopard, jaguar, and
snow leopard. Royal Bengal Tiger is one of the nine subspecies three of which are
extinct. An endangered species as listed by the International Union for Conservation
of Nature (IUCN) Red Book, India provides habitat to 2226 Royal Bengal Tigers also
known as Panthera Tigris Tigris. It is one of the subspecies of the Felidae family, the
largest of the ‘big cats’ in the genus Panthera. Records of occurrence for Panthera
Tigris Tigris were collected from Global Biodiversity Information Facility which
provides combination from various sources such as the Wildlife Institute of India,
iNaturalist and mainly the field museum of natural history for different temporal and
spatial data. The locations depicting the physical presence of Panthera Tigris Tigris
in terms of latitude and longitude along with 233 other attributes were extracted.
Figure 1 represents the architectural design of the steps undertaken. Data from GBIF
is preprocessed and plotted on Google Maps. Different datasets representing various
ecological, climatic, and human factors are taken from a variety of sources. These
are tested for correlation and sampling bias is removed. Following this, different
machine learning algorithms are applied and the results are studied and described.

Fig. 1 Architecture to represent species distribution modeling for Panthera Tigris Tigris

[email protected]
Species Environmental Niche Distribution Modeling … 255

3.2 Preprocessing of Distribution Data

From GBIF, the Darwin core Archive was downloaded as it is directly given to the
publisher consisting of 422 global occurrences. After extracting the coordinate posi-
tions of Panthera Tigris Tigris, there is a need to preprocess this data which contains
presence only locations. It becomes necessary to avoid errors in coordinate locations,
misleading information about the species as well as records having null attributes.
The following steps were followed and after their conclusion, the preprocessed data
was displayed to depict the geographical coordinates in R. The above preprocessing
filters the number of occurrence to 130 from 442.
• Selection of important attributes—From the gbif file of the above species, we
select only attributes of importance from the 235 attributes such as species latitude,
longitude, country, name, and any other which seems necessary.
• Removing duplicate records—For the same spatial (in real time geographical)
coordinates, a species may have presence only data at multiple time stamps. Thus,
it is necessary to remove them as it may bias the output.
• Georeferencing—There may be records with no coordinates but with locality
descriptions. These were georeferenced. This could be done with the help of
geocode function present in the dismo package in R if necessary.
• Map the dataset within the required spatial coordinates—The species records
were limited to the Indian subcontinent by imposing restriction on the raster extent.
The extent was decided to be (64.84885, 97.71038, 5.758113, 35.43777) as per
the standard (x min , x max , ymin , ymax ) to represent the Indian subcontinent.

3.3 Removing Sampling Bias

Some presence only data is derived from random sampling. These may be collected
near frequent forest treks and reserve park roads and these spatial sampling biases are
not included in presence-only data. To remove this, spatial thinning was implemented
using the R package spThin. The idea is to preserve most effective information while
removing bias through removing records on the basis of a thinning distance of 3 km
for a geographical area. Finally, from the 442 occurrences, we focus down on 48
occurrences localities after performing spatial thinning.

3.4 Environmental Data

The environmental data consists of various datasets representing various environment

variables, land cover, and human influence on the environment among others so as to
model suitable habitats for Panthera Tigris Tigris. The resolution chosen was 0.05°
as this is suitable for prediction at the national level. At the equator, this comes out

[email protected]
256 S. Bajaj and D. Geraldine Bessie Amali

to be 5.6 km2 . The Coordinate reference system (CRS) chosen is the EPSG: 4326,
WGS 84.
• Land cover—Many recent researches make use of The University of Maryland’s
(UMD) global land cover classification at a one kilometer resolution [9]. This
classification was done using the 1992–1993 satellite images. These datasets have
become outdated with time and for rapidly developing country like India these may
not represent the actual spatial land cover data in the current scenario. To mini-
mize such differences, one of the MODIS products, the Land Cover Type Climate
Modeling Grid product MCD12C1 consisting of seventeen different land cover
classes was taken to represent land cover for the Indian subcontinent. MCD12C1
is at 0.05° resolution for which the pixel size is around 5.6 km2 . The Royal Bengal
tiger usually thrives in a wide variety of habitats such as tropical and deciduous
rainforests, swamps, shrublands, grasslands, and high altitude. In the Indian region,
they are found along the Western Ghats, in deltas of Bengal, the Terai highlands,
Northeast region and parts of central India among others. Thus, the classes depict-
ing the above landscape were taken to represent their habitat for more accurate
prediction.
• Bioclimatic—Worlclim contains a set of nineteen climatic factors. The raster
dataset was then averaged for all the months which depicted the average over the
above time period. These were resampled to 0.05° using bilinear interpolation.
• Elevation—For elevation, the GTOP030 dataset was utilized. It is a part of the
elevation schema of the United States Geological Survey (USGS) updated and
released in 2000. The dataset which was available in 30 arc second resolution
was resampled to using bilinear method to 0.05°. The dataset was then cropped to
respective extent so as to represent the Indian subcontinent.
• Global Human Footprint and Population density—GHF is a dataset under the
Last of the Wild, v2 (2004) from SEDAC depicting human impact in terms of
population pressure, human access in terms of distance to roads, coastline, rail
lines, and navigable waterways and infrastructure development in terms of night
lights and built up area. The human population density raster dataset was taken
from GPW, v4 estimated for the year 2020.

3.5 Correlation Test

The test for correlation is done with the help of Pearson Correlation and environmen-
tal factors coefficient with correlation coefficient greater than 0.7 were excluded. The
environmental variables finally selected are shown in Fig. 2 based on the coefficient
values in Fig. 4. For any two pixels x and y between two corresponding raster images
representing the same spatial domain of environmental variables in the current sce-
nario, the Pearson correlation coefficient is calculated as per the below equation in
Fig. 3. It is performed over the 390,258 pixels each representing 5.6 km2 in the Indian
subcontinent.

[email protected]
Species Environmental Niche Distribution Modeling … 257

Fig. 2 Environmental variables chosen after preprocessing and performing a correlation test

Fig. 3 Formula for calculation of Pearson correlation coefficient where x and y represent any two
pixels

Fig. 4 Correlation between environmental variables after removing those with a correlation coef-
ficient greater than 0.70

4 Models and Result

Modeling was done using machine learning algorithms mainly random forest, artifi-
cial neural networks, and support vector machines along with Maxent based on the
literature survey done. Since there may yet exist some spatial sampling bias hence
spatial partition was implemented using Block scheme where k  4 whose results
are shown in Fig. 7. The maximum number of iterations was set to be 500. This was

[email protected]
258 S. Bajaj and D. Geraldine Bessie Amali

implemented using ENMeval. Based on longitudes and latitudes, the spatial extent is
divided into four. The presence points were also divided randomly into 75% training
(36) and 25% testing (12) in another instance. 10,000 background points were ran-
domly chosen. Performance is measured through AUC and Kappa statistic which is
calculated as shown in Figs. 5 and 6 ROC is calculated as the ratio between sensitivity
and 1—specificity (False positive rate).
Maxent—Maxent model performs better in accuracy and prediction than other
approaches which already exist as conveyed by Elith et al. [10]; Phillips et al. [2]. The
maximum entropy model calculates the training gain for each of the variable for the
48 presence points which are split into training and testing subsamples. The gain for
each variable here represents the maximum information extracted from each factor
keeping others constant. This is done by jackknife as shown in Fig. 8. For Maxent, the
AUC indicated a high degree of accuracy and for training, it was measured to be 0.922
while test AUC is 0.963. This is based on sensitivity which represents whether the
position is favorable for Panthera Tigris Tigris to occur when environmental factors
are present (true positive rate) and similarly the specificity which here represents the
true negative rate. Figure 7 indicates the predicted geographical range with red being
the highest probability of occurrence and blue with no occurrence. The contribution
of each variable as per the gain is also calculated and shown in Fig. 8.
Random Forest—As per Petitpierre et al. [11], RF has one of the highest accu-
racy among the machine learning models available for the majority of the datasets
regardless of its size. It can be implemented for both presence and pseudo-absence

Fig. 5 Contingency table and expected frequency for each cell

Fig. 6 Calculations for various performance measures and area under the receiver operating char-
acteristic curve for 48 presence points (36 training points and 12 testing points) for Maxent

[email protected]
Species Environmental Niche Distribution Modeling … 259

Fig. 7 Spatial partition using block (k  4) and predicted potential geographical locations for
Panthera Tigris Tigris

Fig. 8 Contribution of environmental variables and training gain for each variable (jackknife) for
Maxent

data and corrects many of the flaws in classification and regression tree including
overfitting. It also mentions the low reliability for a dataset consisting of categorical
factors with varying levels. RF being a regression tree model which does classifica-
tion, it comes to the result using various tree predictors in a forest. The AUC was
found to be 0.93 along with the Kappa statistic at 0.86.
Artificial Neural Networks—These are biologically inspired models comprising
of neurons and their corresponding weights. The AUC was found to be 0.906 along
with the Kappa statistic at 0.809. Although ANN are easy to model for species with
more presence locations yet it can be influenced by a large number of classes.
Support Vector Machines—These are multidimensional classifiers used for
regression and classification through supervised learning. The AUC is 0.898 with
a Kappa statistic of 0.795. Since they have a good generalization ability along being

[email protected]
260 S. Bajaj and D. Geraldine Bessie Amali

Fig. 9 Predicted geographic distribution using RF and ANN, respectively

Fig. 10 Binary distribution for RF and ANN modeled on a threshold calculated using TSS

able to handle high dimensional data such as raster layers, SVM was chosen for
modeling.
The potential geographical distribution for various models is shown in Figs. 9, 10,
11, 12. An ensemble model consisting of all the three models was also formed. The
contribution of each of the variables was calculated using Pearson. To generate the
binary distribution, sum of specificity and sensitivity as represented in Figs. 5 and 6
also known as TSS (True skill statistic) was taken to derive the threshold value as it
represents the point of optimization under the receiver operating characteristic curve
and is also found to minimize the mean error rate [12]. These represent locations
representing potential predicted habitats in binary terms (Figs. 10, 12).

[email protected]
Species Environmental Niche Distribution Modeling … 261

Fig. 11 Predicted geographic distribution using SVM and Ensemble modeling, respectively

Fig. 12 Binary distribution for SVM and ensemble modeling modeled on a threshold calculated
using TSS

5 Conclusion

The high AUC values of 0.963, 0.931, and 0.906 for Maxent, random forest, and ANN
indicate a greater chance for ranking a positive instance which is chosen randomly by
these classifiers as compared to a negative instance with the threshold being TSS. The
Kappa statistic comes into play whenever we need to measure inter-rater agreement
and reliability. From the above models, it can be concluded that the data collected
is a true representation of the variables measured as it is above 0.80 in all these. It
can be seen that population density plays a major role which shifts the prediction to
unhabituated areas. Woody savannas which in India is represented by the Terai region
of Uttarakhand, Uttar Pradesh, Bihar, and parts of Assam is also an important variable
along with Mixed Forest as they provide suitable ecological habitat. Altitude plays a
major role and lowers down the prediction range geographically. Annual precipitation
and temperature together contribute around 14% on average. Human footprint has

[email protected]
262 S. Bajaj and D. Geraldine Bessie Amali

Fig. 13 Tiger conservation landscape (TCL) in the Indian subcontinent and variable contribution
along with performance measures for RF, ANN, and SVM

little impact on prediction as compared to population density. Finally, shrublands,

evergreen, and deciduous forests provide suitable conditions for prey capture and
hence are important.
Also, it can be seen that the predicted areas lie in and around the tiger conservation
landscape (TCL) zones as declared by WWF (Fig. 13). These are found basically
in Madhya Pradesh, Uttarakhand and parts of Rajasthan, and a large area of pre-
dicted zones by all the models lie under these. Maxent provided a high level of AUC
at 0.964 and was chosen since it is independent of the number of occurrences and
performs better than various other models. Care was taken to proceed with correct
practices for MAXENT as mentioned by Yackulic et al. [13]. RF has one of the high-
est accuracy while SVM performs generalization well although accuracy is lower.
The above results can be used to study regions around the present tiger reserves in
India as they will provide locations suitable for reintroduction as per environmental
factors free from human influence. Each pixel of the raster images represents an area
of around 5.6 km2 . These geographical locations can then be used either as a part of
ecological corridors to prevent genetic isolation or as independent reserves. Although
several methods exist for modeling the distribution of various species yet there exist
many limitations. Resampling usually involves predicting cell values from the exist-
ing through various techniques such as bilinear, averaging, etc., which may lead to
decrease in accuracy since each cell value is depicting the predicted value rather than
the actual value. The factors incorporating and describing the habitat and needs of
the particular species must be included after proper study of the requirement of the
species. Future work requires testing on an extensive dataset, mapping the predicted
geographical locations with an actual land survey and comparing with other present
models so as to strengthen and make use of the results of this research.

[email protected]
Species Environmental Niche Distribution Modeling … 263

References

1. Grinnell, J. (1917). The niche-relationships of the California thrasher. Auk, 34, 427–433.
2. Phillips, S. J., Anderson, R. P., & Schapire, R. E. (2006). Maximum entropy modeling of
species geographic distributions. Ecological Modelling, 190, 231–259.
3. Thatte, P., Joshi, A., Vaidyanathan, S., Landguth, E., & Ramakrishnan, U. (2018). Maintaining
tiger connectivity and minimizing extinction into the next century: Insights from landscape
genetics and spatially-explicit simulations. Biological Conservation, 218, 181–191.
4. Kywe, T. Z. (2012). Habitat suitability modeling for tiger(Panthera tigris) in the Hukaung
Valley Tiger Reserve, Northern Myanmar, Niedersächsische Staats (157 pp.). Germany: Nund
Universitätsbibliothek Göttingen.
5. Hernandez, P. A., Graham, C. H., Master, L. L., & Albert, D. L. (2006). The effect of sample size
and species characteristics on performance of different species distribution modeling methods.
Ecography, 5, 773–785.
6. Kulloli, R. N., & Kumar, S. (2014). Comparison of Bioclimatic, NDVI and elevation variables
in assessing extent of Commiphora wightii (Arnt.) Bhand. ISPRS—International Archives of
the Photogrammetry, Remote Sensing and Spatial Information Sciences, XL-8, 589–595.
7. Duan, R.-Y., Kong, X.-Q., Huang, M.-Y., Fan, W.-Y., & Wang, Z.-G. (2014). The predictive
performance and stability of six species distribution models. PLoS ONE, 9(11), e112764.
https://doi.org/10.1371/journal.pone.0112764.
8. Qing, Z., Zhang, Y., Sun, G., Duo, H., Wen, L., & Lei, G. (2015). Using species distribution
model to estimate the wintering population size of the endangered scaly-sided merganser in
China. PLoS ONE, 10, e0117307. https://doi.org/10.1371/journal.pone.0117307.
9. Hansen, J., Sato, M., Ruedy, R., Lacis, A., & Oinas, V. (2000). Global warming in the twenty-
first century: An alternative scenario. Proceedings of the National Academy of Sciences, 97,
9875–9880. https://doi.org/10.1073/pnas.170278997.
10. Elith, J., Graham, C. H., Anderson, R. P., et al. (2006). Novel methods improve prediction of
species’ distributions from occurrence data. Ecography, 29, 129–151.
11. Petitpierre, B., Kueffer, C., Broennimann, O., Randin, C., Daehler, C., et al. (2012). Climatic
niche shifts are rare among terrestrial plant invaders. Science, 335, 1344–1348.
12. Freeman, E. A., & Moisen, G. G. (2008). A comparison of the performance of threshold criteria
for binary classification in terms of predicted prevalence and kappa. Ecological Modelling, 217,
48–58.
13. Yackulic, C. B., et al. (2013). Presence-only modelling using MAXENT: When can we trust
the inferences? Methods in Ecology and Evolution, 4, 236–243.
14. Adhikari, D., Barik, S. K., & Upadhaya, K. (2012). Habitat distribution modelling for rein-
troduction of Ilex khasiana Purk., a critically endangered tree species of northeastern India.
Ecological Engineering, 40, 37–43.
15. Elith, J. H., Graham, C. P., Anderson, R., Dudík, M., Ferrier, S., et al. (2006). Novel methods
improve prediction of species’ distributions from occurrence data. Ecography, 29, 129–151.

[email protected]
Organizational Digital Footprint
for Traceability, Provenance Approach

Sheetal Arya, Kumar Abhishek and Akshay Deepak

Abstract In this age of instant noodles, instant messaging, and tweeting, we have
all we want at the tips of our finger whether it be the delivery of a food product,
buying a property, or going for a date. All but a click away. As quoted, “The con-
trol of information is something successful person does.” If one could control the
information, he would be able to make worthy decisions. Now the question arises,
how does one know which information is relevant and which is fake, for making
the foundation of decisions? Here comes the role of provenance. In this paper, the
characteristics of a data such as why, when, where, what, how, and who are deter-
mined using PROV family of documents, giving full structural format to data and
determining all its characteristics and its existential source. Now all this provenance
information describing all aspects of data when connected will generate a footprint.
Apache NiFi has been used to show the flow of data and its manipulation. The digital
footprint as it is all in digital format can be traceable to determine data trustwor-
thiness, i.e., its provenance, hence determining the reliability of an organization for
merger, acquisition, data analysis, stock market, etc.

Keywords Big data · Provenance · Digital footprint · Knowledge representation

1 Introduction

As noticed by E. O. Wilson, “We are drowning in information while starving for

wisdom.” It could be seen there is too much information available but irrelevant.
This big data of real-time information, i.e., raw data, could be used to seek queries,
suggestion, and solutions, i.e., meaningful data, knowledge. The more relevant the

S. Arya (B) · K. Abhishek · A. Deepak

Department of Computer Science and Engineering, NIT Patna, Patna 800005, Bihar, India
e-mail: [email protected]
K. Abhishek
e-mail: [email protected]
A. Deepak
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2019 265
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_23

[email protected]
266 S. Arya et al.

trustworthy knowledge one has, the more reliable the decisions one would be able
to make; hence come provenance, semantics, Apache NiFi, and digital footprint in
the picture.
In Sect. 2, the related works have been described which lead to the invention
of the proposed system model. In Sect. 3, how provenance has been used in this
model has been described. In Sect. 4, the proposed workflow of the system has been
described. In Sect. 5, the methodology has described which consists of dataset used,
Sect. 5.1 describes the implementation of data flow and PROV-N, and finally, Sect. 5.2
describes digital footprint formation, i.e., result. In Sect. 6, the major achievements
and limitations are described. Section 7, finally, consists of conclusion and future
scope.

2 Related Works

The 14 papers used in this paper for provenance data implementation are summarized
in Table 1 having references [1–13]. Table 1 consists of five columns, where column 1
represents serial number. Column 2 represents the year in which paper was published.
Column 3 represents the purpose of the published paper. Column 4 represents the
scope of data, which can be used in the implementation; sometimes, format is written
as data scope because paper simply describes the syntax/format for implementation
purpose. Column 5 represents the concept described in the paper and the algorithm
used in the implementation of the framework.

3 Provenance

According to the dictionary, provenance is “The place of origin or earliest known

history of something” [14]. The structure of provenance data model is represented in
Fig. 1. Arrows are read in the backward direction, such as an entity was used by an
Activity, etc. In this diagram, symbols are used to represent attributes, and properties
are represented by arrowheads. An Entity is used to represent an object, which is
a source of information. An Activity is processed, which are involved with data
manipulation, i.e., converting data from one form to another. An Agent represents the
living or nonliving object responsible for initiating processes for data manipulation.
There are various other core and extended, attributes and properties of the provenance
data model [1].
The granularity with which provenance of a data is to be maintained is deter-
mined according to the requirement of provenance information for implementation,
because it could be as specifically defined as you want it to be. Yet again if that much
granularity is not required, then it is useless.

[email protected]
Organizational Digital Footprint for Traceability … 267

Table 1 Literature survey

S. No. Year of Purpose Data scope Concepts/
publication algorithms used
1 2013 PROV Format Provenance
introduction introduction,
summary is
described
2 2013 Ontology for Web, format Provenance
provenance ontology syntax
for provenance
data model is
described
3 2013 Data model for Web, format Structural format
provenance for attributes and
properties is
described
4 2013 Notations for Web, format Notational
provenance representation of
provenance data
model is
described
5 2013 Constraints for Web, format Implementation
provenance of constraints is
described
6 2013 Accessing and Web, format Accessing and
querying of querying
provenance mechanism for
provenance is
described
7 2013 XML schema for Web, format XML
provenance implementation
of provenance is
described
8 2013 Dictionary Web, format Collective
representation of structural
provenance representation of
provenance is
described
9 2013 Dublin core terms Web, format Mapping to DC
representation of terms vocabulary
provenance is described
10 2013 Semantics for Web, format Semantics
provenance implementation
in first-order
logic for
provenance is
described
(continued)

[email protected]
268 S. Arya et al.

Table 1 (continued)
S. No. Year of Purpose Data scope Concepts/
publication algorithms used
11 2013 Link connectivity Web, format Connectivity of
for provenance different bundles
of provenance
data using PRO
LINKS is
described
12 2010 Capturing and Web Framework,
usage of models for
provenance provenance
implementation
13 2016 A blockchain Supply chain, Blockchain,
ontology for blockchain, web solidity,
supply chain pages language,
provenance ethereum, TOVE
traceability
ontology

Fig. 1 Provenance data

model

4 Proposed System Workflow

The steps to be followed for fetching digital footprint are determined here. First,
data is searched from websites, repositories, and Wikipedia about organizations in
CSV format. That data is collected in Microsoft Excel files. This classified data is
first imported in Apache NiFi [15]. In NiFi, processors are used, connections are
established, and flow files are derived for final data flow derivation. NiFi will give us
data flow of information from different sectors of organizations. The data collected
will be written in provenance notation PROV-N deriving knowledge. Thus, finally
Fig. 2 shows, the workflow of digital footprint created using visualization tools.

[email protected]
Organizational Digital Footprint for Traceability … 269

Fig. 2 Workflow diagram

5 Methodology

In methodology, the technical implementation part of the paper is described. Here,

the datasets used, software tools used, implementation criteria used, and the final
result will be described. The dataset used for this work has been taken from the
following website:
• Wikipedia [16],
• Quandl [17], and
• Google [18].

5.1 Implementation

In this paper, finding provenance of petroleum companies’ framework has been

divided into two phases as follows:

Phase I: Dataflow using Apache NiFi

[email protected]
270 S. Arya et al.

Fig. 3 Data flow of all collected information from distributed organizations to forming flow of data
of petroleum companies in England

Information (such as location, revenue, acquisition, etc.) of petroleum companies in

England is collected from sources such as Wikipedia, Quandl, etc. The information
obtained is treated as flow files, and a data flow of these files is generated in Apache
NiFi as shown below. Through this data flow, information is transferred through
different levels. As shown, processors British Petroleum, GE Oil & Gas, and Cove
Energy Plc. represent information collected by respective petroleum companies,
headquartered in England, whereas sub-processors such as processor Revenue fetch
the data of revenue generated by the company, processor Subsidiaries collects infor-
mation relating to subsidiaries of a company, processor Operating income describes
dataset of income on which the company operates, processor Profit describes the
data of a company profits, processor Predecessor describes data of company that was
replaced and now is known as mentioned company, processor Parent describes the
parent company of the present mentioned company, processor Acquisition describes
datasets of the companies that were acquired by the mentioned company, and pro-
cessor Successor describes the company which has acquired the present mentioned
company.
All the information collected in these workspaces (GE Oil & Gas, British
Petroleum, Cove Energy plc) is transferred to “Eng_Petroleum” workspace, which
is represented by processor Petroleum Companies in England. So, all information
of these three petroleum companies is stored in Petroleum Companies in England
processor as shown in Fig. 3. Also, in the workflow, information details at each level
can be found.
British Petroleum gets information from processor Revenue, Subsidiaries, Oper-
ating income, and Profit. It has four subsidiaries located in Pakistan, Serbia, New

[email protected]
Organizational Digital Footprint for Traceability … 271

Fig. 4 Data flow of British petroleum from its attributes

Fig. 5 Data flow of British petroleum further from its acquisition

Zealand, and Mangolia. So, processor Subsidiaries gets its information from pro-
cessors named after its subsidiaries as shown in Fig. 4.
British Petroleum has also acquired four companies named Sohio, Amoco, ARCO,
and Burmah Castrol. Information of all these acquisitions described by a respective
processor is collected by the Acquisition processor as shown in Fig. 5.

[email protected]
272 S. Arya et al.

Similarly, information related to GE Oil & Gas is transferred to processor GE

Oil & Gas, and information related to company Cove Energy Plc. is collected by
processor Cove Energy plc.
Phase II: Knowledge for digital footprint formation using provenance
Metadata formation using provenance notation PROV-N: PROV-N is a W3C recom-
mended notation for provenance model. It is invented to represent information about
the data of the PROV data model as per these implementation principles:
• Readability for humans: Provenance notations determine syntax which is human-
readable and machine understandable so it can be used for illustration of examples.
• Independence of technology: Provenance notations provides format so simple that
it can easily be mapped to several technologies.
• Formal representation: Provenance notations are defined in a formal grammar
representation which is acceptable to be used with parser generators.
In phase II, we identify entities, activities, agents, and relationships among them
through data flow generated in phase II, and write code using PROV-N syntax to
obtain the provenance metadata, describing knowledge in ontologies, hence showing
digital footprint in visualization generated.

5.2 Result

The following figures in this section describe visualization generated using prove-
nance notation PROV-N, describing digital footprint generated by European
Petroleum Organizations. All the petroleum industries in England data are collected
in a directory, which is represented by entity Petroleum Companies in England as
shown in Fig. 6.
The agents are organizations attributed to entities BP, Cove, GE which are repre-
sented by data directories as shown above. The activities SendInfo1, SendInfo2, etc.,
are associated with agent ApacheNiFi as it is basically the backbone to collect and
transfer all data information. The activity SendInfo8 used information from entities
BP, Cove, GE to generate entity PetroleumCompaniesInEngland.

Fig. 6 Digital footprint generated by Europe petroleum industries

[email protected]
Organizational Digital Footprint for Traceability … 273

Fig. 7 Digital footprint generated by British petroleum describing its attributes

Fig. 8 Digital footprint generated by British petroleum further describing its subsidiaries

The entity BP used information from entities BP_Predecessor, BP_Profit,

BP_Revenue, BP_Operating_Income, BP_Acquisition, and BP_Subsidiaries. Enti-
ties ARCO, BurmahCastrol, Sohio, Amoco represent the information of companies
that have been acquired by the company British Petroleum. So, information from
these entities has been used by activity SendInfo2 to generate entity BP_Acquisition
as shown in Fig. 7.
Similarly, British Petroleum industry has subsidiaries in Pakistan, Magno-
lia, Serbia, and New Zealand and their collected data are represented by enti-
ties BP_Pakistan, BP_Mangolia, BP_Serbia, and BP_NewZealand. These enti-
ties are used by activity SendInfo1 to generate entity BP_Subsidiaries. Hence,
BP_Subsidiaries represents information relating to subsidiaries of British Petroleum
as shown in Fig. 8.
The Cove_Profit entity was used by activity SendInfo7 to generate entity Cove,
which is finally attributed to similarly for Cove_Profit entity which is attributed
to agent CoveEnergyPlc an organization and for entity GE that is attributed to
agent GE_Oil_Gas also an organization. Hence as could be seen, this is all a digital
footprint representation of companies catching all milestones of their journey.

[email protected]
274 S. Arya et al.

6 Major Achievements and Limitation

The limitation discovered from recent papers is that the data about organizations is
not available publicly; hence, further granularity is not possible easily.
The major achievements are that Trustworthiness of Organization are fetched,
which has many benefits such as lower cost of transactions as valid decisions could be
made for reducing expenditure. Apart from reliability, digital footprint could also be
used to validate data quality, i.e., increasing automatic data processing and analyzing
using scientific workflows could be influenced for automated collection, processing,
and analysis of provenance information. The provenance metadata generating digital
footprint will be used to track how market information is modified and used across
the financial system.

7 Conclusion and Future Scope

This paper presented a model for data flow and digital footprint using provenance
from W3C Provenance family of documents and related papers. The 14 published
papers were summarized using tables; they were used in contributing to model for
digital footprint formation. After going through these papers, it is clear that the
enhancements of provenance are an open field for research in a financial institution,
as where else authorization, validation, and trustworthiness are more required other
than where money is involved.
Paper have explained, provenance approach for deriving dataflow and addi-
tionally creating a digital footprint for tracing of required information for
merger, acquisition, subdivision, and other financial matters. This knowledge
will help in determining the source of default or identifying whether a sub-
sidiary is working efficiently or not, when to merge two departments, dur-
ing acquisition how perfect decisions could be made by derived knowledge,
etc.
In the next phase of this data flow, using open-source Java API for OWL Ontologies
and other API implementation is proposed. So that automatic implementation of the
model could be done. Hence, automatic retrieval of data from sources could be done
and their footprint will be derived, without manual labor for any domain.

References

1. PROV-PRIMER. (2013). World wide web consortium W3C. Retrieved from: https://www.w3.
org/TR/2013/NOTE-prov-primer-20130430/.
2. PROV-O: The PROV Ontology. (2013). World Wide web consortium W3C. Retrieved from
https://www.w3.org/TR/2013/REC-prov-o-20130430/.

[email protected]
Organizational Digital Footprint for Traceability … 275

3. PROV-DM: The PROV Data Model. (2013). World Wide web consortium W3C. Retrieved from
https://www.w3.org/TR/2013/REC-prov-dm-20130430/.
4. PROV-N: The PROV Notation. (2013). World Wide web consortium W3C. Retrieved from
https://www.w3.org/TR/2013/REC-prov-n-20130430/.
5. PROV-CONSTRAINTS: The PROV Constraints. (2013). World Wide web consortium W3C.
Retrieved from https://www.w3.org/TR/2013/REC-prov-constraints-20130430/.
6. PROV-AQ: The PROV Access and Query. (2013). World Wide web consortium W3C. Retrieved
from https://www.w3.org/TR/2013/NOTE-prov-aq-20130430/.
7. PROV-XML: The PROV XML Schema. (2013). World Wide web consortium W3C. Retrieved
from https://www.w3.org/TR/2013/NOTE-prov-xml-20130430/.
8. PROV-DICTIONARY: The PROV Dictionary. (2013). World Wide web consortium W3C.
Retrieved from https://www.w3.org/TR/2013/NOTE-prov-dictionary-20130430/.
9. PROV-SEM: The PROV Semantics. (2013). World Wide web consortium W3C. Retrieved from
https://www.w3.org/TR/2013/NOTE-prov-sem-20130430/.
10. PROV-LINKS: The PROV Link Across Bundles. (2013). World Wide web consortium W3C.
Retrieved from https://www.w3.org/TR/2013/NOTE-prov-links-20130430/.
11. Allen, A., & David, M. et al. (2010). Provenance capture and use: A practical guide. The
MITRE Corporation.
12. Kim, H. M., & Laskowski, M. (2016). Towards an ontology-driven blockchain design for
supply chain provenance.
13. Sheth, A., & Kapanipathi, P. (2016). Semantic filtering for social data. IEEE Internet
Computing, 20(4), 74–78. Retrieved from http://ieeexplore.ieee.org/abstract/document/
7529038/.
14. Dictionary Meaning of provenance. Retrieved from https://en.oxforddictionaries.com/
definition/provenance.
15. Apache NiFi overview. Retrieved from https://nifi.apache.org/docs.html.
16. Wikipedia. Retrieved from https://en.wikipedia.org/wiki/GE_Oil_and_Gas.
17. Quandl. Financial data repository. Retrieved from https://www.quandl.com.
18. Google browser. Retrieved from https://www.google.co.in/.

[email protected]
Bidirectional Long Short-Term Memory
for Automatic English to Kannada
Back-Transliteration

B. S. Sowmya Lakshmi and B. R. Shambhavi

Abstract Transliteration is the key component in various Natural Language Pro-

cessing (NLP) tasks. Transliteration is the process of converting one orthographic
system to another. This paper demonstrates transliteration of Romanized Kannada
words to Kannada script. Our system utilizes a bilingual corpus of around one lakh
words, which comprise pairs of Romanized Kannada word with its corresponding
word in Kannada script and employs orthographic and phonetic information. Recur-
rent Neural Networks (RNNs) are widely used Neural Networking model for text
and speech processing as they better predict the next word based on past informa-
tion. Long Short-Term Memory (LSTM) Networks are exceptional kind of RNNs
which handles long-term dependencies. A Character level Bidirectional Long Short-
Term Memory (BLSTM) paradigm which drives down the perplexity with respect to
word-level paradigm has been employed. Knowledge of Characters uncovers struc-
tural (dis)similarities among words, thus refining the modeling of uncommon and
unknown words. Test data of 3000 Romanized Kannada words is used for model
evaluation and we obtained an accuracy of 83.32%.

Keywords Transliteration · Bilingual corpus · RNN, LSTM

1 Introduction

Transliteration is the task of mapping graphemes or phonemes of one language into

phoneme approximations of another language. It has got various applications in
the domain of NLP like Machine Translation (MT), Cross Language Information
Retrieval (CLIR), and information extraction. Even though the task appears trivial,
prediction of pronunciation of the original word is a crucial factor in translitera-
tion process. Transliteration process among two languages is minimum if they share

B. S. Sowmya Lakshmi (B) · B. R. Shambhavi

Department of ISE, BMS College of Engineering, Bangalore 560019, India
e-mail: [email protected]
B. R. Shambhavi
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2019 277
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_24

[email protected]
278 B. S. Sowmya Lakshmi and B. R. Shambhavi

identical alphabet set. However, for languages which practice nonidentical set of
alphabets, words have to be transliterated or portrayed in the native language alpha-
bets.
Majority of multilingual web users have a tendency to represent their native lan-
guages in Roman script on social media platforms. In spite of many recognized
transliteration standards, there is an intense inclination to use unofficial transliter-
ation standards in many websites, social media, and blog sites. There are ample of
issues such as spelling variation, diphthongs, doubled letters, and reoccurring con-
structions which are to be taken care while transcribing.
Neural Networks is a rapidly advancing approach to machine learning [1, 2] and
has shown promising performance when applied to a variety of tasks like image
recognition, speech processing, natural language processing, cognitive modeling,
and so on. It involves using neural networks for training a model for a specific task.
This paper demonstrates the application of neural network for machine transliteration
of English–Kannada, two linguistically distant and widely spoken languages.
The rest of this paper is arranged as follows. Section 2 describes prior work in this
area. An introduction to LSTM and BLSTM is described in Sect. 3. The methodology
adopted to build corpus is presented in Sect. 3. Proposed transliteration network
is portrayed in Sect. 4. Section 5 provides details of results obtained. Section 6
communicates conclusion and future work of the proposed method.

2 Previous Research

Research on Indic languages within the perspective of social media is quite ample,
with numerous studies concentrating on code-switching has become a quite familiar
phenomenon. There are a few substantial works being done on Transliteration or,
more precisely, back-transliteration of Indic languages [3, 4]. A shared task which
included, back-transliteration of Romanized Indic language words to its native scripts
was run in 2014 [5, 6]. In many areas, including machine transliteration, end-to-end
deep learning models have become a good alternative to more traditional statistical
approaches. A Deep Belief Network (DBN) was developed to transliterate from
English to Tamil with restricted corpus [7] and obtained an accuracy of 79.46%. A
character level attention-based encoder in deep learning was proposed to develop a
transliteration model for English–Persian [8]. The model presented a good accuracy,
with BLEU score of 76.4.
In [9], authors proposed transliteration of English to Malayalam using phonemes.
English–Malayalam pronunciation dictionary was used to map English graphemes
to Malayalam phonemes. Performance of the model was fairly good for phonemes
in pronunciation dictionary. However, it suffered from out-of-vocabulary issue when
a word is not in pronunciation dictionary.
The most essential requisite of transliterator is to retain the phonetic structure
of source language after transliterating in target language. Different transliteration
techniques for Indian languages were proposed. In [10], input text was split into

[email protected]
Bidirectional Long Short-Term Memory … 279

phonemes and was classified using Support Vector Machine (SVM) algorithm. Most
of the methods adopted features like n-grams [11], Unicode mapping [12], or a
combination-based approach by combining phoneme extraction and n-grams [13,
14].
Antony et al. [15–17] have proposed Named Entities (NE) transliteration tech-
niques from English to Kannada. In [15, 16], authors adopted a statistical approach
using widely available tools such as Mosses and Giza++ which yielded an accuracy
of about 89.27% for English names. System was also evaluated by comparing with
Google transliterator. A training corpus of 40,000 Named Entities was used to train
SVM algorithm [17] and obtained an accuracy of 87% for 1000 test dataset.

3 LSTM and BLSTM Network

RNNs have been employed to produce promising results on a variety of tasks includ-
ing language model [18] and speech recognition. A RNN foresees present output
based on the preserved memories of its past information. RNNs are designed for
capturing information from sequences or time series data.
RNN networks [19] comprises an input layer, hidden layer, and an output layer
where each cell preserves a memory of previous time. Figure 1 demonstrates a simple
RNN model where X 0 , X 1 , X 2 are inputs at timestamps t 0 , t 1 , t 2 and hidden layer
units are h0 , h1 , h2 :
The new state (ht ) of RNN at time t is a function of its previous state at time t −
1 (ht −1 ) and the input at time t (x t ). Output (yt ) of the hidden layer units at time t are

Fig. 1 A simple RNN model

[email protected]
280 B. S. Sowmya Lakshmi and B. R. Shambhavi

calculated using new state calculated and the weight matrix. The maths behind RNN
to calculate output from hidden and output layers are as follows:
 
h (t)  gh Wi X (t) + W R h (t−1) + bh (1)
 
Y (t)  g y WY h (t) + b y (2)

where W Y , W R, and Wi are weights which are to be calculated during training phase,
gh and gy are activation functions computed using Eqs. (3) and (4) respectively and
bh and by are bias. RNN uses backpropagation algorithm, but it is applied for every
time stamp. It is commonly known as Backpropagation Through Timestamp (BTT).
The dimensionality of output layer is same as labels and also characterizes likelihood
distribution of labels at time t.

1
gh  (3)
1 + e−z
ezm
g y   zk (4)
ke

where gh is a sigmoid function and gy is softmax activation function which maps

input to the output nonlinearly.
LSTM Networks are special kind of RNNs and these RNNs are accomplished to
learn long-term dependencies. Consider a language model of text prediction, which
predicts a next word based on the previous word. In order to predict German as the
last word in the sentence “I grew up in Germany. I speak fluent German” recent
information suggests that next word might probably be the language but in order to
narrow down the language context of German is needed which is quite long back
in the sentence. This is known as long-term dependencies. LSTMs are capable of
handling this type of dependencies where RNNs fail. Figure 2 shows a LSTM cell.

Fig. 2 A LSTM cell

[email protected]
Bidirectional Long Short-Term Memory … 281

where
σ  logistic sigmoid function
i  input gate
f = forget gate
o  output gate
c  cell vectors
h  hidden vector
W  weight matrix
LSTM is implemented as the following:

• Primary step in the LSTM is to decide the data to be neglected from the cell state
which is made by forget gate layer. This decision is made by a sigmoid layer
called the “forget gate layer”. It is a function of ht −1 and xt as shown in Eq. (5),
and outputs a number between 0 and 1 for each number in the cell state ct −1 . A
1 represents “completely keep this” while a 0 represents “completely get rid of
this.”

 
f t  σ (W f h t−1 , X t + b f ) (5)

• The second step is to decide the new data to be stored. This has two steps, input
obtained from the previous timestamp and the new input are passed through a
sigmoid function called “input gate layer” to get it as shown in Eq. (6). Next, input
obtained from the previous timestamp and the new input are passed through a tanh
function. Both the steps are combined with ft passed from the previous step as in
Eq. (7).

 
i t  σ (Wi h t−1 , X t + bi ) (6)
 
ct  f t ct−1 + i t tanh (Wc h t−1 , X t + bc ) (7)

• Last step is to obtain output using Eqs. (8) and (9), which is based on cell state.
First, a sigmoid layer decides what parts of the cell state are to be outputted. Then,
the tanh function pushes cell state values between −1 to 1, which is multiplied by
the output of the sigmoid gate, so that only decided parts happen to be the output.

   
ot  σ Wo h t−1 , X t + bo (8)
   
ot  σ Wo h t−1 , X t + bo (9)

[email protected]
282 B. S. Sowmya Lakshmi and B. R. Shambhavi

3.1 BLSTM Network

Sequence learning task requires previous and forthcoming input features at time t.
Hence, BLSTM network is used to utilize previous features and upcoming features
at a given time t. BLSTM [20] network hidden layer contains a sequence of forward
and backward recurrent neural components connected to the identical output layer.
Figure 3 shows a simple BLSTM network of four input units X 0 to X 4 . Network
hidden layer has four recurrent components h0 to h3 in the forward direction and
four recurrent components ho to h3 in the backward direction to help predict output
Y 0 to Y 3 by forming an acyclic graph. Most of the text processing task BLSTM
would provide reasonable results in the prediction of sequence of data.

4 Dataset Collection

To develop a transliteration model using neural networks necessitates a significant

amount of bilingual parallel corpus. For resource-poor languages like Kannada, it
is hard to obtain or build a large corpus for NLP applications. Therefore, we built
a training corpus of around 100,000 English–Kannada bilingual words. Bilingual
words were collected from following various sources.
• Various websites were scraped to collect most familiar Kannada words and their
Romanized words. Special characters, punctuation marks, and numerals were
removed by preprocessing. This data contributed around 20% of the training data.

Fig. 3 A BLSTM network

[email protected]
Bidirectional Long Short-Term Memory … 283

• Majority of the bilingual words were collected from music lyrics websites which
consist of song lyrics in Kannada script and its corresponding song lyrics in Roman-
ized Kannada. Non-Kannada words, punctuations, and vocalize words in song
lyrics were removed. Obtained list comprehends viable syllable patterns in Kan-
nada and contributed around 70% of the training data.
• The subsequent share of the corpus was manually transliterated NEs.

5 Experiments

5.1 Setup

The proposed approach is implemented on python platform and packages used are
numpy and neural network toolkit keras with Tensorflow in the backend. The network
parameters are set up as in Table 1.

5.2 Training Procedure

The proposed model is implemented using simple BLSTM network as shown in

Fig. 4. Paradigm is trained with the collected dataset which contained bilingual
corpus Romanized Kannada words and its corresponding word in Kannada script.
During Training, 20% of the training data is set as the validation data. Algorithm 1
describes network training procedure. In each epoch, entire training data is divided
into batches and one batch is processed at given time t. Batch size determines number
of words to be included in a batch. Characters in each input word are embedded and
provided as input to forward and backward state of LSTM. Later, we backpropagate
the errors from the output to the input to update the network parameters.

Table 1 Model parameters Parameter Value

No. of epochs 30
Batch size 128
Hidden units 128
Embedding dimension 64
Validation split 0.2
Output activation function Softmax
Learning rate 0.001
Training model Bidirectional LSTM

[email protected]
284 B. S. Sowmya Lakshmi and B. R. Shambhavi

Fig. 4 BLSTM network

Algorithm 1 BLSTM Model Training Procedure

for each epoch do
for each batch do
1) bidirectional LSTM model forward pass:
forward pass for forward state LSTM
forward pass for backward state LSTM
2) bidirectional LSTM model backward pass:
backward pass for forward state LSTM
backward pass for backward state LSTM
3) update parameters
end for
end for

6 Results

Model was tested for the dataset of around 3 K words collected from random websites.
Test dataset contains Romanized words and its transliterated words in Kannada script
which is kept as reference to compare with the result.
Snapshot of results obtained from the model is shown in Table 2. The correctness
of the transliteration is measured by Accuracy (ACC) or Word Error Rate (WAR)
yielded by a transliteration model. For completeness, other transliteration results
obtained by RNN and LSTM networks which are trained for the same datasets are
reported in Table 3.

[email protected]
Bidirectional Long Short-Term Memory … 285

Table 2 Snapshot of results

Romanized Kannada Gold standard Resultant word Transliteration result
word transliterated
word
Tirugu Correct

Setuve Correct

Aggalikeya Correct

Sadbhava Incorrect

Anaupacharika Incorrect

Table 3 Evaluation results Model Accuracy obtained (%)

RNN 74.33
LSTM 79.76
BLSTM 83.32

7 Conclusion and Future Work

Transliteration is the task of mapping graphemes or phonemes of one language into

phoneme approximations of another language. It is the elementary step for most of
the NLP applications like MT, CLIR, and text mining. English and Kannada language
trail dissimilar scripts and also vary in their phonetics. Furthermore, Romanization
of Kannada words does not go along a standard pattern as far as their pronunci-
ation is concerned. Thus, a particular set of rules do not guarantee an effective
back-transliteration. In this paper, we have presented a Transliteration model for
English–Kannada language pair.
A character level BLSTM model was investigated, which utilizes character embed-
ding for words. Along with BLSTM, model was also tested for LSTM and RNN
for English and Kannada. The correctness of the transliteration was measured by
Accuracy for a test data of 3000 Romanized Kannada words. As BLSTM has two
networks, one access information in the forward direction and another access in the
reverse direction, the output generated is from both the past and future context. Accu-
racy obtained by BLSTM was more when compared to LSTM and RNN for this test
data.
There are several possible courses for future improvement. First, model would
be further progressed by combining other algorithms with BLSTM. For example,
combination of CNN with BLSTM or CRF with BLSTM would yield better results.
Another improvement is to expand the training data by collecting data from other
domains such as social media (Twitter and Weibo) which would include all pos-

[email protected]
286 B. S. Sowmya Lakshmi and B. R. Shambhavi

sible orthographic variations. Since model is not restricted to specific domain or

knowledge, social media text would also provide a fair share in training data.

References

1. Kalchbrenner, N., & Blunsom, P. (2013). Recurrent continuous translation models. In Proceed-
ings of the 2013 Conference on Empirical Methods in Natural Language Processing.
2. Sutskever, I., Vinyals, O., & Le, Q. V. (2014). Sequence to sequence learning with neural
networks. In Advances in neural information processing systems.
3. Sharma, A., & Rattan, D. (2017). Machine transliteration for indian languages: A review.
International Journal, 8(8).
4. Dhore, M. L., Dhore, R. M., Rathod, P. H. (2015). Survey on machine transliteration and
machine learning models. International Journal on Natural Language Computing (IJNLC),
4(2).
5. Sequiera, R. D., Rao, S. S., Shambavi, B. R. (2014). Word-level language identification and
back transliteration of romanized text: A shared task report by BMSCE. In Shared Task System
Description in MSRI FIRE Working Notes.
6. Choudhury, M. et al. (2014). Overview of fire 2014 track on transliterated search. Proceedings
of FIRE, 68–89.
7. Sanjanaashree, P. (2014). Joint layer based deep learning framework for bilingual machine
transliteration. In 2014 International Conference on Advances in Computing, Communications
and Informatics (ICACCI). IEEE.
8. Mahsuli, M. M., & Safabakhsh, R. (2017). English to Persian transliteration using attention-
based approach in deep learning. In 2017 Iranian Conference on Electrical Engineering (ICEE).
IEEE.
9. Sunitha, C., & Jaya, A. (2015). A phoneme based model for english to malayalam translitera-
tion. In 2015 International Conference on Innovation Information in Computing Technologies
(ICIICT). IEEE.
10. Rathod, P. H., Dhore, M. L., & Dhore, R. M. (2013). Hindi and Marathi to English machine
transliteration using SVM. International Journal on Natural Language Computing, 2(4),
55–71.
11. Jindal, S. (2015, May). N-gram machine translation system for English to Punjabi translit-
eration. International Journal of Advances in Electronics and Computer Science, 2(5). ISSN
2393-2835.
12. AL-Farjat, A. H. (2012). Automatic transliteration among indic scripts using code mapping
formula. European Scientific Journal (ESJ), 8(11).
13. Dasgupta, T., Sinha, M., & Basu, A. (2013). A joint source channel model for the English to
Bengali back transliteration. In Mining intelligence and knowledge exploration (pp. 751-760).
Cham: Springer.
14. Dhindsa, B. K., & Sharma, D. V. (2017). English to Hindi transliteration system using
combination-based approach. International Journal, 8(8).
15. Antony, P. J., Ajith, V. P., & Soman. K. P. (2010). Statistical method for English to Kannada
transliteration. In Information Processing and Management (pp. 356–362). Berlin, Heidelberg:
Springer.
16. Reddy, M. V., & Hanumanthappa, M. (2011). English to Kannada/Telugu name transliteration
in CLIR: A statistical Approach. International Journal of Machine Intelligence, 3(4).
17. Antony, P. J., Ajith, V. P., & Soman, K. P. (2010). Kernel method for English to Kannada
transliteration. In 2010 International Conference on Recent Trends in Information, Telecom-
munication and Computing (ITC). IEEE.

[email protected]
Bidirectional Long Short-Term Memory … 287

18. Mikolov, T. et al. (2011). Extensions of recurrent neural network language model. In 2011
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE.
19. Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14(2), 179–211.
20. Schuster, M., & Paliwal, K. K. (1997). Bidirectional recurrent neural networks. IEEE Trans-
actions on Signal Processing, 45(11), 2673–2681.

[email protected]
A Dominant Point-Based Algorithm
for Finding Multiple Longest Common
Subsequences in Comparative Genomics

Manish M. Motghare and Preeti S. Voditel

Abstract Finding the longest common subsequence is a classic and well-studied

problem in the field of computer science and considered as an NP-hard problem.
There are many application of LCS in the field of bioinformatics, computational
genomics, image processing, file comparison, etc. There are many algorithms are
present to find the similarity between the given strings and its special cases. As there is
a tremendous increase in the biological data and it requires an efficient mechanism to
deal with them, many efforts have been taken to reduce the time and space complexity
of the given problem. In this paper, we presented a novel algorithm for the general
case of multiple LCS problems, i.e., finding a longest common subsequence in the
given two strings. Our algorithm works on dominant point approach to compute
the LCS of the given string. When applied to multiple strings of length each 1000,
2000, 3000, 4000, and 5000, characters, it is found that our algorithm works two or
three magnitude faster than existing algorithm and it requires less space compared
to existing algorithms.

Keywords Longest common subsequence (LCS) · Dynamic programming ·

Np-hard problem · Dominant point · Problem complexity · Comparative genomics

1 Introduction

In biological computation, we need to compare two or more strings, in order to find a

similarity between them. This is usually done in finding similarity between two organ-
isms by comparing their DNA. A DNA comprises several molecules called bases.
The DNA bases are adenine, cytosine, guanine, and thymine. All of these bases
are represented as {A, C, G, T}, respectively. For example, we are having DNA
of two organisms, let ‘X  ACGGTGTCGTGCTATGCTGATGCTGACTTATAT-
GCTA’ denote DNA of one organism and ‘Y  CGTTCGGCTATCGTACGTTC-

M. M. Motghare (B) · P. S. Voditel

Department of Computer Application, Shri Ramdeobaba College of Engineering and
Management, Nagpur, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 289

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_25

[email protected]
290 M. M. Motghare and P. S. Voditel

TATTCTATGATTTCTAA’ denote DNA of other. We can say two organisms are

similar to each other if one’s DNA is a substring of another, but in this case, neither
strings are a substring of each other. So another method to find similarity between the
two organisms can be determined by generating a third string which is formed from
the given two strings, the new string will contain common subsequence between
them, it is necessary the bases of DNA will be in same organized order, but not pos-
sibly consecutively. In this case, the common subsequence between X and Y is Z 
‘CGTTCGGCTATGCTTCTACTTATTCTA’. The longer the common subsequence,
the more the similarity between the given DNA strands. But up to what extent they are
similar can be determined based on longest common subsequence. A subsequence
is a part of a huge sequence which can be obtained from another by removing one
or more elements without changing the order. For example, the sequence {C, B, C,
E, D, B, G} is obtained from the {A, C, B, D, E, G, C, E, D, B, G} [1].
Table 1 shows the list of common structures in computational biology and their
sizes [2]. The LCS algorithms are widely used in computational biology as well as
in many traditional applications like file comparison, data compression, and FPGA
synthesis.
The longest common subsequence (LCS) is a standard and well-studied problem
of computer science. It is an NP-Hard problem with many applications in image
processing, computational genomics, bioinformatics, etc [11–13]. Over many years,
significant efforts have been made to find efficient algorithm to solve the problem
of finding the longest common subsequence in the multiple strings, but the present
algorithms having limitation as they work on only few cases of MLCS [14–17], or
the problem’s special case of two or three strings [18, 19]. Several methods have been
proposed over general strings [19–22], and they could greatly be benefitted if there
is improvement in computational time of the algorithms. As mentioned this method
of MLCS [23–26] can be applied in bioinformatics and computational genomics
which deals with biological data. Due to tremendous increase in biological data and
widespread use of sequence analysis tools, we expect the usage of MLCS method in
Computational genomics and their applications.
In this paper, we present an algorithm which is fast and space efficient for the
multiple longest common subsequences (MLCS) problem. The algorithm works on
larger strings and returns a longest common subsequence faster than existing algo-

Table 1 A details of  
Biological data Alphabet ( ) | | Typical
biological data and its sequence
component length
DNA {A C G T} 4 104 –1011 [3–6]
RNA {A C G U} 4 10–104 [7]
Genome {gene1 , 104 10–104 [8, 2]
gene2 …genek }
Protein {A, C…..,W} 20 102 –104 [9,
10]

[email protected]
A Dominant Point-Based Algorithm … 291

rithms. Our algorithm works on the dominant point approach [18, 19, 25]. Procedure
which uses a dominant point method is more efficient, which results in reducing the
size of search space compared to other existing algorithm [19]. A dominant point
represents matching character in the row, or simply k-dominant or dominant at level
k and it is represented by circle. The key idea behind dominant point method is to
recognize matching values of all levels in a matrix M, instead of all values in matrix
[14, 26, 27]. In our algorithm, we are filling the values in the matrix only if match is
found, i.e., “1”, reducing the space required to store other values. The place where
the value “1” is printed is called dominant point. In this, there is no requirement to
store the value of pointers to trace the longest common subsequence, result in saving
of time. Whereas LCS based on dynamic programming require trace-back approach,
which follows the arrow backwards, starting from the last entry of the matrix [28].
The paper is organized as follows: In the next section, we discussed some of the
existing policies and related work. Later, we will discuss our new algorithm and its
complexity and last we will compare our algorithm with other existing algorithm.

2 Problem Formulation and Existing Methods

In this paragraph, we will discuss existing techniques and number of ways to find
MLCS of a given string.

2.1 Problem Definition

Definition 1 The longest common subsequence is defined as follows. Consider the

string X which is of length ‘L’ and another string Y of length ‘M’. The main goal
is to find the longest common subsequence present in the given two strings. The
subsequence will be calculated from left to right order, which can be derived from
another sequence by deleting some elements without changing the order of remaining
elements. The longer the subsequence the more the similarity in between the strings.

Let us consider the following strings:

X  TGAC
Y  AGCT

In this case, longest common subsequence has length 2, i.e., subsequence ‘AC’.
An alternative approach to look at this is that we are finding 1–1 matching between
the character of X and Y strings.

Definition 2 Let A  {a1 , a2 ….an } be a set of alphabets over , then B  {b1 ,
b2 ….bn }, are the subsequence generated from ‘A’, then multiple longest common

[email protected]
292 M. M. Motghare and P. S. Voditel

subsequence can be considered only if, b is generated from set ‘A’ without changing
the order of other elements.

The subsequence ‘b’ is the longest sequence of all the available sequences present
in ‘A’, satisfying point mentioned above.
For the given strings, there can be one or more subsequence will be present.
Consider the given strings X  {A, B, C, B, D, A, B} and Y  {B, D, C, A, B,
A}, then the common subsequence of both the strings could be Z  {BDAB, BDA,
BD, B}. So, the longest common subsequence present in the multiple string is Z 
BDAB. For multiple longest common subsequences, finding LCS is declared as its
special case.

2.2 Dynamic Programming Methods

The traditional method for multiple longest common subsequence is based on

dynamic programming approach [17]. In this approach, two sequences are given n1
and n2 each of length L 1 and L 2 , respectively, the algorithm maintains two matrixes,
one for storing the length of the string, i.e., ‘b’ and second matrix, i.e., c to store few
pointers in a parallel array, c[0…m, 0…n] value for particular ith row and jth column.
Let us consider a string of protein sequences as characters. Given two sequences L 1
 {AATCCGCTAG} and L 2  {AAACCCTTAG}, the motivation is to compute the
LCS of the two whole strings. In this, we will compute the LCS for every pair of
prefixes and will store the value in C[i][j]. In the dynamic programming approach,
we have come up with three observations.
First Approach: b[i, 0]  b[j, 0]  0. If out of the given two sequences, any
string found to be empty then LCS in between the string will be empty.
Second Approach: Assume X i  Y j . If the last characters of the given two
sequences are same, then the LCS must end with same character. Hence X i  Y j then
b[i, j]  b[i − 1, j − 1] +1.
Third Approach: Assume X i  Y j . In this approach, both X i and Y j will not
appear in the longest common subsequence. From this, we can conclude either X i is
not a part of the LCS or Y j . Hence, X i  Y j . Then b[i, j]  max (b[i − 1, j], b[i, j −
1]).
By applying the dynamic programming approach, we get the matrix ‘b’. The
length of the longest common subsequence is calculated from the end point of the
matrix b[i, j] (Fig. 1).

3 Related Work

Since last decade, numerous algorithms have been proposed to find the longest com-
mon subsequence of the multiple strings, and most of them are based on dynamic

[email protected]
A Dominant Point-Based Algorithm … 293

Fig. 1 The LCS is calculated using all the three approaches of dynamic programming mentioned
above. The MLCS present in the above string is “BDAB”. Arrows represent pointers which are
stored in matrix ‘c’, which helps in trace-back of matching sequence from the end point, i.e.,
b[i][j] to b[0][0]. The dynamic programming approach having time complexity of O(m + n), as
at least one of the ‘i’, or ‘j’, decrements at each stage. The resultant algorithm is having space
complexity of O(nl), for ‘l’, sequence of length ‘n’. Currently, there are many algorithms [14–16]
present for finding the longest common sequence, but they are having large complexity and require
more memory than our proposed algorithm

programming. But due to tremendous increase in the dataset of biological data, the
algorithm which is based on dynamic programming scale badly and found to be less
space and time efficient when dealing with larger strings to compare. Due to this, the
focus has been shifted to heuristic techniques which are capable to deal with large
data sets around gigabytes of data which are produced during genome sequencing
[29]. The algorithm developed by Hirschberg [30] or Irving and Fraser [31], having
complexity of O(l n ), which shows they are exponential in nature. Many improve-
ments have been done to reduce this complexity to O(ln−1 ), like reducing the space
by calculating dominant points [32] which are used in practice. Parallelization of this
algorithm has been done by Chen et al. [33], i.e., Fast_LCS and Wang et al. [34], by
running algorithm over multiple cores of processor and distributing the work among
them. The implementation of this algorithm work over 10 strings (n  10), with
more strings their runtime becomes impractical. Various approximation algorithms
like long run proposed by Jiang and Li [35] which create a longest string containing
only a single character that is valid subsequence in all strings there which finds the
optimal solution to NP-hard problems with reduced time complexity [36] as compare
to heuristics approach, they have better solution and runtime. Heuristics algorithm
does not provide good solution and runtime but they provide solution quickly which

[email protected]
294 M. M. Motghare and P. S. Voditel

is “good enough”. Jones et al. [37, 38] and Bryant [39, 40] proposed a GA algorithm
for finding the longest common subsequence of an arbitrary string. They had done
comparison of genetic algorithm with Fraser and Irving which are based on dynamic
programming [31] and reported a shorter runtime for genetic algorithm.

4 New Fast_LCS Algorithm

Our algorithm works on the principal of point dominant approach. In our proposed
algorithm LCS is calculated in row major order. Figure 2, shows a pseudocode for
our proposed algorithm. The algorithm works parallel, i.e., it starts comparing first
row and first column value of the said matrix, in row major order and also prints the
value in the output array. Following observations are associated with the algorithm
and combining all the observations jointly we have the following rule:
Rule 1: C[i, j]  “1”, if x i  yj , stop scanning for particular row and column and
increments the value C[x i + 1, y + 1].
Rule 2: C[i, j]  “0”, if x i  yj , increment the value in row major order till the end
of the row.

Fig. 2 Pseudocode of dominant MLCS algorithm

[email protected]
A Dominant Point-Based Algorithm … 295

4.1 Quick Version of Algorithm to Find MLCS

In this section, we present a new memory-efficient dominant point-based algorithm

for finding multiple longest common subsequence (Fig. 3).
First Step: Compare the first letter of ‘Y’ axis with all character of ‘X’ axis in
row major order.
a. Compare the first letter ‘B’ of Y -axis with first letter ‘A’ of X-axis, as there is no
match, put value “0”.
b. Compare the first letter ‘B’ of Y -axis with second letter ‘B’ of X-axis, as there
is a match found, put value “1”.
c. Compare the first letter ‘B’ of Y -axis with third letter ‘C’ of X-axis, as there is
no match, put value “0”.
d. Compare the first letter ‘B’ of Y -axis with fourth letter ‘B’ of X-axis, as there is
a match found, put value “1”.
e. Compare the first letter ‘B’ of Y -axis with fifth letter ‘D’ of X-axis, as there is
no match, put value “0”.
f. Compare the first letter ‘B’ of Y -axis with sixth letter ‘A’ of X-axis, as there is
no match, put value “0”.
g. Compare the first letter ‘B’ of Y -axis with seventh letter ‘B’ of X-axis, as there
is a match found, put value “1”.
h. The process will be same for the next letter ‘D’, ‘C’, ‘A’, ‘B’, and ‘A’.

Fig. 3 The matrix for two

sequences X  ABCBDAB A B C B D A B
and Y  BDCABA, the LCS X/Y
found in the given
 string is
“BDAB”, over  |4|
B 0 1 0 1 0 0 1

D 0 0 0 0 1 0 0

C 0 0 1 0 0 0 0

A 1 0 0 0 0 1 0

B 0 1 0 1 0 0 1

A 1 0 0 0 0 1 0

[email protected]
296 M. M. Motghare and P. S. Voditel

4.2 Computing Longest Common Subsequence

Step 1: The algorithm will search for match in the first row, if match is found, it will
print the value of the matching character, i.e., “1”. And then it will stop searching
for next matching character in that row. Now LCS is “B”.
Step 2: Again the algorithm will search for matching pair in the second row, but this
time it will not consider that row and column from where first match pair was found
and also it will not consider the left column values for the first matching pair while
searching for next computation. In the second row, we are getting another matching
pair so it will print “1”. Now LCS at this moment is “BD”.
Step 3: Discard Value of Row and Column after 2nd step.
Step 4: Search for matching pair at row “3”, here match between “C-C”, is found,
i.e., value ‘1’, but this value falls in the left column of second matching pair of row,
so this matching value will not be considered. Current LCS value is “BD”.
Step 5: Search for matching pair at fourth row, here match between “A-A” is found,
so it will print value ‘1’, and LCS will be “BDA”. This time it will not consider upper
row values and left column values.
Step 6: Search fifth row for matching pair, here two matches are found, but first
matching pair will not be considered because it comes before the left column of the
first selection, so it will be discarded. Then, next match is found, i.e., last “B-B”,
whose value is ‘1’, so it will be considered. Now algorithm will print LCS “BDAB”.
Step 7: This will be the final Multiple LCS for the given sequence given in Table 3.

5 Results

In our experiment, the algorithm is executed in a system having Intel Core i5-2330 M
Processor 2.20 GHz Windows 10 Home Basic (64-bit), Memory 4 GB/Hard Disk
Drive 720 GB. The execution programming environment is GNU C++. The algorithm
is tested on strings of length 1000, 2000, 3000, 4000, and 5000, each consists of four
characters{A, C, G, T}, i.e., a nucleic acid sequence.

5.1 Dominant Point Multiple Longest Common Subsequence

Algorithm

The developed algorithm is compared with Quick-DP [20] and Hakata et al. A and
C algorithms [18]. The C algorithm proposed by Hakata et al. is designed for any
number of strings. The A algorithm works for three strings and so far it is the fastest
algorithm for three sequences. The QUICK-DP is a generalized algorithm which
is developed to work with any number of sequences [20]. We implemented both
algorithms as mentioned in their paper and the same environment was created for

[email protected]
A Dominant Point-Based Algorithm … 297

Fig. 4 Longest common

subsequence “BDAB” is
found between the sequences

execution of algorithm. As multiple longest common subsequence algorithms is

having many application from finding of similarity among the biological sequences,
computational genomics to other than biological domain, for every application there
is a different representation of dataset. In our experiment, we considered a database
which is generated from the four characters of nucleotide sequence, i.e., {A, C, G,
T} and each dataset consists of sequence of length 100, 200, 300, 400, 500, 600, 700,
800, 900, 1000, 2000, 3000, 4000, and 5000. In Table 1 and Fig. 4, the comparison is
shown among Quick-Dp, Hakata and Imai algorithm, and our D_MLCS algorithm
(Tables 2 and 3).

6 Conclusion

In this paper, we have proposed a new algorithm for finding the longest common
subsequence, which is applied to both biological sequences which are randomly
generated and real biological sequence and it is found that it is working efficiently
compared to the existing algorithm with less time and space complexity. Recently,
there is a tremendous change found in the study of protein and genomic sequences
[20, 27], so our algorithm development and improvement will be on the basis of
requirement by the bioinformatics community. In the current implementation of this
algorithm, we focused on finding the longest common subsequence on genomic and
protein sequences, each consists of combination of  4 and  20 characters,
respectively. So in our future implementations of this algorithm, we will try to handle
larger protein families. Our contribution in this paper is designing of new, efficient
algorithm and its comparison with the existing algorithm.

[email protected]
298 M. M. Motghare and P. S. Voditel

Table 2 Comparison is done among the latest algorithm and our implementation mentioned above
in terms of avg. running time of algorithm in seconds by giving inputs of 15 available random
sequences of different size where input character is  |4|
Sequence length Quick-DP [28] Hakata and Hakata and Our
Imai A Imai C implementation
algorithm [19] algorithm [19] of algorithm
100 0.0 0.6 1.7 0.03
200 0.1 4.2 6.6 0.04
300 0.3 13.0 26.0 0.14
400 0.8 62.4 71.5 0.16
500 1.3 174.2 203.3 0.17
600 NA NA 560.3 0.18
700 NA NA NA 0.20
800 NA NA NA 0.22
900 NA NA NA 0.30
1000 NA NA NA 0.35
2000 NA NA NA 1.13
3000 NA NA NA 2.37
4000 NA NA NA 4.45
5000 NA NA NA 6.83

Table 3 Implementation results of dominant LCS algorithm

S. No. X array Y array Time Processor RAM RAM
sequence sequence (s) type
Execution Name Speed Core
time
1 100 100 0.03 Intel 2.2 GHz 2 4 GB DDR3
2 200 200 0.04 core i5-
2330 M
3 300 300 0.14
4 400 400 0.16
5 500 500 0.17
6 600 600 0.18
7 700 700 0.20
8 800 800 0.22
9 900 900 0.30
10 1000 1000 0.35
11 2000 2000 1.13
12 3000 3000 2.37
13 4000 4000 4.45
14 5000 5000 6.83

[email protected]
A Dominant Point-Based Algorithm … 299

References

1. Cormen, T. H., Leiserson, C. E., Rivest, R. L., Stein, C. (2009). Introduction to algorithm (3rd
ed.). The MIT Press.
2. Nekrutenko, A., & Li, W.-H. (2001). Transposable elements are found in alarge number of
human protein-coding genes. Trends in Genetics, 17(11), 619–621.
3. Gregory, T. R. (2005). Animal genome size database. Retrieved from http://www.Genomesize.
com.
4. Lodish, H. F. (2003). Molecular cell biology. WH Freeman.
5. Paterson, M., Dančík, V. (1994). Longest common subsequence’s. In Proceedings of the 19th
International Symposium on Mathematical Foundations of Computer Science (pp. 127–142).
Springer.
6. Fortnow, L. (2009). The status of the P versus NP problem. Communications of the ACM, 52(9),
78–86.
7. Elbashir, S. M., Harborth, J., Lendeckel, W., Yalcin, A., Weber, K., & Tuschl, T. (2001).
Duplexes of 21-nucleotide RNAs mediate RNA interference in cultured Mammalian cells.
Nature, 411(6836), 494–498.
8. Blanchette, M., Kunisawa, T., & Sankoff, D. (1999). Gene order breakpoint evidence in animal
mitochondrial phylogeny. Journal of Molecular Evolution, 49(2), 193–203.
9. Brocchieri, L., & Karlin, S. (2005). Protein length in eukaryotic and prokaryotic proteomes.
Nucleic Acids Research, 33(10), 3390–3400.
10. Zastrow, M. S., Flaherty, D. B., Benian, G. M., & Wilson, K. L. (2006). Nuclear titin interacts
with A-and B-type lamins in vitro and in vivo. Journal of Cell Science, 119(2), 239–249.
11. Luce, G., & Myoupo, J. F. (1998). Systolic-based parallel architecture for the longest common
subsequences problem. VLSI Journal Integration, 25, 53–70.
12. Sankoff, D., & Blanchette, M. (1999). Phylogenetic invariants for genome rearrangements.
Journal of Compuational Biology, 6, 431–445.
13. Sheridan, R. P., & Venkataraghavan, R. (1992). A systematic search for protein signature
sequences. Proteins, 14(1), 16–18.
14. Hirschberg, D. S. (1977). Algorithms for the longest common subsequence problem. Journal
of the ACM, 24, 664–675.
15. Masek, W. J., & Paterson, M. S. (1980). A faster algorithm computing string edit distances.
Journal of Computer and System Sciences, 20, 18–31.
16. Rick, C. (1994, October). New algorithms for the longest common subsequence problem (Tech-
nical Report No. 85123-CS). Computer Science Department, University of Bonn.
17. Smith, T. F., & Waterman, M. S. (1981). Identification of common molecular subsequences.
Journal of Molecular Biology, 147, 195–197.
18. Hakata, K., & Imai, H. (1992). Algorithms for the longest common subsequence problem. In
Proceedings of Genome Informatics Workshop III (pp. 53–56).
19. Hakata, K., & Imai, H. (1998). Algorithms for the longest common subsequence problem for
multiple strings based on geometric maxima. Optimization Methods and Software, 10, 233–260.
20. Chen, Y., Wan, A., & Liu, W. (2006). A fast parallel algorithm for finding the longest common
sequence of multiple biosequences. BMC Bioinformatics, 7, S4.
21. Korkin, D. (2001). A new dominant point-based parallel algorithm for multiple longest common
subsequence problem (Technical Report TR01-148). University of New Brunswick.
22. Xu, X., Chen, L., Pan, Y., He, P. (2005). Fast parallel algorithms for the longest common
subsequence problem using an optical bus. In Lecture Notes in Computer Science (pp. 338–348).
Springer.
23. Bork, P., & Koonin, E. V. (1996). Protein sequence motifs. Current Opinion in Structural
Biology, 6, 366–376.
24. Korkin, D., & Goldfarb, L. (2002). Multiple genome rearrangement: A general approach via
the evolutionary genome graph. Bioinformatics, 18, S303–S311.

[email protected]
300 M. M. Motghare and P. S. Voditel

25. Korkin, D., Wang, Q., & Shang, Y. (2008). An efficient parallel algorithm for the multiple
longest common subsequence (MLCS) problem. In Proceedings of the 37th International
Conference on Parallel Processing (ICPP’08) (pp. 354–363).
26. Bergroth, L., Hakonen, H., & Raita, T. (2000). A survey of longest common subsequence algo-
rithms. In Proceedings of International Symposium. String Processing Information Retrieval
(SPIRE’00) (pp. 39–48).
27. Chin, F. Y., & Poon, C. K. (1990). A fast algorithm for computing longest common subsequences
of small alphabet size. Journal of Information Processing, 13(4), 463–469.
28. Wang, Q., Korkin, D., & Shang, Y. (2011, March). A fast multiple longest common subsequence
(MLCS) algorithm. IEEE Transactions on Knowledge and Data Engineering, 23(3).
29. Yang, J., Yun, X., Sun, G., & Shang, Y. (2013). A new progressive algorithm for a multiple
longest common subsequences problem and its efficient parallelization. IEEE Transactions on
Parallel and Distributed Systems, 24(5), 862–870.
30. Hirschberg, D. S. (1975, June). A linear space algorithm for computing maximal common
subsequences. Communications of the ACM, 18, 341–343.
31. Irving, R. W., & Fraser, C. (1992). Two algorithms for the longest common subsequence of
three (or more) strings. In Proceedings of the Third Annual Symposium on Combinatorial
Pattern Matching (pp. 214–229). London, UK: Springer.
32. Wang, Q., Korkin, D., & Shang, Y. (2009). Efficient dominant point algorithms for the multi-
ple longest common subsequence (MLCS) problem. In Proceedings of the 21st International
Joint Conference on Artifical Intelligence (pp. 1494–1499). San Francisco, CA, USA: Morgan
Kaufmann Publishers Inc.
33. Chen, Y., Wan, A., & Liu, W. (2006). A fast parallel algorithm for finding the longest common
sequence of multiple biosequence. BMC Bioinformatics, 7, 4.
34. Wang, Q., Korkin, D., & Shang, Y. (2011). A fast multiple longest common subsequence
(MLCS) algorithm. IEEE Transactions on Knowledge and Data Engineering, 23(3), 321–334.
35. Jiang, T., & Li, M. (1994). On the approximation of shortest common supersequences and
longest common subsequences. In Proceedings of the 21st International Colloquium on
Automata, Languages and Programming (pp. 191–202). London, UK: Springer.
36. Cormen, T. H., Leiserson, C. E., Rivest, R. L., & Stein, C. (2009). Introduction to algorithms
(3rd ed.). Cambridge, MA, USA: MIT Press.
37. Jones, E., Oliphant, T., & Peterson, P. (2013). SciPy: open source scientific tools for python.
Retrieved fromhttp://www.scipy.org/. Accessed April 19, 2013.
38. Maier, D. (1978). The complexity of some problems on subsequences and supersequences.
Journal of the ACM, 25, 322–336.
39. Julstrom, B. A., & Hinkemeyer, B. (2006). Starting from scratch: Growing longest common
subsequences with evolution. In Proceedings of the 9th International Conference on Parallel
Problem Solving from Nature (pp. 930–938). Berlin, Heidelberg: Springer.
40. Bergroth, L., Hakonen, H., & Raita, T. (2000). A survey of longest common subsequence algo-
rithms. In Proceedings Seventh International Symposium on String Processing and Information
Retrieval SPIRE 2000 (pp. 39–48).
41. Attwood, T. K., & Findlay, J. B. C. (1994). Fingerprinting G protein coupled receptors. Protein
Engineering, 7(2), 195–203.
42. Bourque, G., & Pevzner, P. A. (2002). Genome-scale evolution: Reconstructing gene orders in
the ancestral species. Genome Research, 12, 26–36.

[email protected]
Fast and Accurate Fingerprint
Recognition in Principal
Component Subspace

S. P. Ragendhu and Tony Thomas

Abstract In the case of fingerprint-based person recognition, the most widely used
discriminating features are minutiae (end points and bifurcations of ridges). Majority
of fingerprint matching algorithms are dealing with comparing the parameters di-
rectly derived from or relative to minutiae points extracted from the templates. Hence
eventually fingerprint matching based on minutiae can be reduced to a 2D point set
matching problem. Various security pitfalls like impersonation using one’s minutiae
coordinates and performance issues related to enhancement as well as spurious minu-
tiae removal are obvious in such a system. Certain non-minutiae based schemes are
able to give acceptable performance at the cost of increased complexity which results
in increased execution time. In order to overcome these issues, we propose a simple
yet efficient and faster fingerprint alignment and matching scheme based on statisti-
cal features which will not reveal the unique local features of the template. Proposed
matching technique is based on the weighted similarity score obtained by comparing
the principal component subspaces of fingerprint templates. Proposed method also
utilizes an alignment scheme based on principal components calculated for the 2D
coordinates of fingerprint region with minimal overhead without any helper data.

1 Introduction

Fingerprint is the pattern formed by ridges and valleys at the palmar side of the
finger tip, which is found to be unique for each individual. Fingerprint biometrics
deal with extracting the distinguishing features of this pattern for each individual and
identification and/or authentication by comparing the extracted features [28].

S. P. Ragendhu (B) · T. Thomas

Indian Institue of Information Technology and Management-Kerala,
Research Centre of Cochin University of Science and Technology, Kazhakkuttom,
Thiruvananthapuram, Kerala, India
e-mail: [email protected]
T. Thomas
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 301

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_26

[email protected]
302 S. P. Ragendhu and T. Thomas

A recent classification of fingerprint features into a three-level hierarchy has been

proposed by experts [6, 29]. Even though there is difference of opinion among
researchers regarding this classification, as a whole we can conclude that Level 1
features include global features like orientation field, ridge frequency field or singular
regions (cores and deltas). Level 2 features include less global features like minutiae
or ridge skeleton. More quantitative features like sweat pores, ridge contours etc.
can be considered as Level 3 features. Level 1 feature can only be used for initial
classification of fingerprints into different groups such as whorls, loops, or arches;
whereas Level 2 features like minutiae can be used for more refined classification and
person authentication as they represent local discriminating features even within the
fingerprints with similar global structure. Different security issues associated with
unauthorized reconstruction of Level 2 features from Level 1 features and vice versa
while employing minutiae-based schemes are available in the literature [16, 22].
Nevertheless, non-minutiae-based features are also not exempted from the threat of
unauthorized reconstruction of original templates [23].
In both minutiae-based and non-minutiae-based schemes, securing the features
against the reconstruction of templates is inevitable. However, when existing template
protection mechanisms such as cancellable biometrics and biometric cryptosystems
are suffering from degradation of matching performance along with increased com-
plexity when combined with state of the art fingerprint recognition systems [19].
Increased complexity slows down the recognition system. A simple, efficient, and
secure feature representation for fingerprints, which can be secured easily is need of
the hour.
Person authentication (one to one) and person identification (one to many) are
the two major categories of biometric applications. As the query template will be
compared with only one registered template in the case of one to one authentica-
tion, most of the algorithms perform efficiently in spite of complex calculations.
However, for person identification, query template needs to be compared with all the
templates present in selected template database. If complex calculations are involved,
the execution time becomes the key factor in deciding the performance of matching
technique. Hence person identification systems require comparatively simpler and
faster method to match fingerprint templates.
In this paper, we propose principal components based alignment and matching for
fingerprint templates, which reduces the execution time significantly without com-
promising the performance. Since we use only global statistical features (Level 1)
for authentication, there is no need to store the intrinsic details (Level 2 and Level
3) of the biometric data in the system and it will ensure the security and privacy of
the overall system. In the proposed alignment scheme, orientation estimation is done
by calculating the principal components of the 2D coordinate points of fingerprint
region and rotation by estimated angle is performed. Proposed matching technique
compares fingerprints by analyzing the similarity of principal component subspaces
of corresponding templates. Similarity of fingerprint images are calculated by esti-
mating weighted cosine similarity of corresponding principal component subspaces.
The calculations are quite simple and straight forward; which makes the method
simpler and faster.

[email protected]
Fast and Accurate Fingerprint Recognition in Principal Component Subspace 303

Rest of this paper is organized as follows. Section 2 discusses about some re-
lated literature in the area of fingerprint recognition. Introduction to the concept of
principal components is given in Sect. 3. Principal components based orientation
estimation and alignment of fingerprint is detailed in Sect. 4.1. Fingerprint matching
scheme based on similarity and weighted similarity calculation of fingerprint tem-
plates using principal components are explained in Sect. 4.2. Section 5 describes
the experimental results and performance of proposed matching scheme and Sect. 6
briefs the conclusion drawn and future directions for research.

2 Related Works

Vast varieties of approaches have been proposed in the domain of fingerprint recogni-
tion, in both minutiae-based [28] and non-minutiae-based [26] categories. Overhead
associated with preprocessing techniques (spurious minutiae removal, minutiae ver-
ification, etc.) itself is really high in the case of minutiae-based algorithms; whereas
complexity of calculations associated with non-minutiae-based schemes with ac-
ceptable performance are higher compared to minutiae-based systems.
Minutiae are coordinate dependent and will get affected largely by distortion and
rotation. Most of the works in the literature are dealing with either relative distance
or relative coordinates of minutiae points and the match is determined by the overlap
between selected features of registered and query templates [24]. The most widely
accepted and recent minutiae-based recognition system is based on Minutiae Cylin-
der Code (MCC) representation of fingerprint templates [4]. In MCC representation,
a cylinder structure is used which encodes spatial and directional relationships be-
tween the minutiae and its neighborhood. In spite of high matching accuracy, higher
execution time and higher memory requirement are the major disadvantages of MCC-
based approach.
One of the popular non-minutiae-based approaches used in fingerprint recognition
is Gabor features based fingercode (filterbank-based) representation of fingerprint
[10]. Matching accuracy of this technique is found to be high, though at the cost
of increased complexity. We have compared the performance of proposed work
with filterbank-based approach and MCC-based approach, where the former is non-
minutiae-based approach and latter is minutiae-based approach.
There are vast varieties of applications for principal components in different do-
mains [13]. The most important application of principal components is principal
component analysis (PCA); which is a statistical technique widely used mainly for
linear dimensionality reduction. When we consider the domain of biometric recogni-
tion, conventional PCA is commonly used in face recognition. However, conventional
PCA (similar to eigen fingerprints and weight calculation) is not able to give accept-
able performance in the case of fingerprint recognition [27]. When compared to face
images, the richness and variance of features are less in fingerprint templates as we
only have some structures (singular points, endpoints and bifurcations) formed by the
flow of ridges and valleys in fingerprints. The sensitiveness of principal components

[email protected]
304 S. P. Ragendhu and T. Thomas

to the outliers might be another reason for poor performance of PCA in fingerprints.
There are certain works which use principal components for structural classification
of fingerprints [2]. Principal components have been used for minutiae verification
and spurious minutiae removal in some works [9].
When we use conventional PCA in biometric recognition [3], if there are m regis-
tered images (N × N ), the first step is to find mean image (N × N ) of the registered
images and it will be subtracted from each of the registered image. After this, a
matrix A N 2 ×m will be formed; where each column vector of A corresponds to each
mean subtracted image. Each column of A.A T represents principal component of
A. Each of the registered templates will be reconstructed using the obtained prin-
cipal components. While reconstruction, weights associated with different principal
components for each image will be calculated. This weight vectors are considered as
the features. When query images are coming, weights of those images will also be
calculated corresponding to the obtained principal components. Distance between
weight vectors of query image and each of the registered images will be calculated.
If there is a registered image with distance less than particular threshold, that will be
considered as the matching template.
In the proposed scheme, instead of using conventional PCA technique, principal
components are used directly for estimating similarity of the fingerprints. We are not
performing dimensionality reduction or weight calculation for registered images.
Instead, principal component subspace of each of the registered template will be
compared with the principal component subspace of the query image. This technique
has been used commonly in pattern recognition of historical data as well as in process
control [5, 11, 25].
Matching techniques which deal with similarity measure based on less complex
calculations can provide optimal tradeoff between performance and complexity. Al-
most all the similarity measures proposed in literature for fingerprint templates are
based on minutiae points [7]. Proposed work discusses a different application of prin-
cipal components, where the similarity between the biometric images is calculated
based on the cosine similarity of principal components.

3 Principal Components

Principal Component Analysis (PCA) deals with finding out new variables (Principal
components) that are linear functions of those in the original dataset, which succes-
sively maximize variance and that are uncorrelated with each other [13]. Principal
components can be interpreted as the new uncorrelated variables extracted from a
high dimensional dataset. The first principal component gives the axis of maximum
variance and the second principal component represents the axis of second most
variance. The problem of extracting the orthonormal vectors (principal components)
can be reduced to an eigenvalue/eigenvector problem.

[email protected]
Fast and Accurate Fingerprint Recognition in Principal Component Subspace 305

The basic idea behind principal components can be given by the Eq. 3:


m
x≈ wi ai + c, (3)
i=1

where, ai , i = 1, 2, . . . , m represent orthonormal vectors (principal components),

wi represent weights associated with each principal component and c is the minimal
squared reconstruction error while reconstructing the data x. If the dimensionality
of the data x is n, Eq. 3 shows how principal components can be used to reconstruct
x in a dimensionally reduced space (m < n). Hence the most important application
of Principal Component Analysis (PCA) is linear dimensionality reduction.
Different variants of applications have been proposed for principal component
analysis in different disciplines. One of the applications of principal components
in image processing is the estimation of orientation of objects in the images [18].
We have utilized this property of principal components for aligning the fingerprint
images and detailed steps are given in Sect. 4.
Another application of principal components which has got recent attention from
researchers is pattern recognition in two multivariate datasets based on the similarity
of principal components of those two datasets. This technique is derived from the
conventional factor analysis method [12] and it was first proposed by Krzanowski
[14]. Multivariate data analysis is a well-researched area and many statistical tech-
niques have been proposed for pattern recognition in multivariate datasets. Different
fields such as finance, process control, multimedia and image processing are dealing
with multivariate data. In the case of biometric systems, the biometric data collected
for authentication in almost all the commonly used modalities (fingerprint, face,
iris, finger vein, palmprint,palm vein, etc.) are in the form of images, which is a
multivariate data. Hence we propose principal components based similarity measure
commonly used for the pattern recognition in multivariate datasets for comparing
the biometric images. When data is represented in high dimensions, it is difficult
to analyze the similarity between them. Principal components basically represent
optimized coordinate system to represent the given data. Hence it becomes com-
paratively easier to compare and analyze the data which is represented in terms of
principal components.

3.1 Calculation of Principal Components

Principal components of a data matrix (biometric image) can be calculated either

by singular value decomposition of a column centered data matrix or by the eigen
value decomposition of a covariance matrix. We have used the latter approach for
calculating the principal components. Different steps involved in the calculation of
principal components are given in Algorithm 1.

[email protected]
306 S. P. Ragendhu and T. Thomas

Algorithm 1 CalcPC
Step 1: Normalize the input image I(r ×c) , by subtracting the mean image of size (r × c) and
obtain I˜(r ×c) .
Step 2: Calculate the covariance matrix of I˜(r ×c) using either Equation 1 or Equation 2
(depends on the need of dimensionality reduction) as given below:

Cov I m (r ×r ) = ( I˜)(r ×c) × ( I˜)(c×r

T
) (1)

Cov I m (c×c) = ( I˜)(c×r

T ˜
) × ( I )(r ×c) (2)

Step 3: Find out the eigen vectors (ei , i = 1, 2....c or r ) corresponding to the eigen values
(λi , i = 1, 2....c or r ) (sorted in descending order) of covariance matrix. Eigen vec-
tors (ei ) represent the principal components and the amount of variance is given by
corresponding eigen values (λi ).

4 Proposed System

Different steps involved in proposed matching technique are shown in Fig. 1. In-
put grayscale fingerprint image will be aligned according to the alignment method
explained in Sect. 4. Angular and orientation alignment will be achieved for the tem-
plates after this step. The enhancement technique proposed by Jain et al. [8] is applied
on the aligned images and morphological operations are applied for smoothing the
image. Noise removal and thinning are performed and region of interest (R O I ) is
extracted out based on the core point [1]. The results of the different preprocessing
steps (enhancement, thinning and ROI extraction) are shown in Fig. 2. Principal
components of the extracted ROI image are calculated by eigen value decomposition
method as explained in Sect. 3.1.

4.1 Principal Components Based Fingerprint Alignment

Fingerprint alignment is the first step to be performed in the proposed system. There
are various complicated procedures proposed in the literature for aligning the query
template with the registered template [21]. Some of them are based on the core
point which require executing different algorithms for core point detection [17].
Some others are based on the helper data stored along with the registered templates
[20], which can become a security threat to the system. In both of these approaches,
significant amount of processing time is required in the alignment process. We use
a comparatively fast, secure and simple method based on the properties of principal
components for aligning fingerprint templates.
In the proposed approach, orientation of fingerprint is estimated by finding out
the axis of maximum variance of the fingerprint region. Fingerprint region is iden-
tified from the image and coordinates of fingerprint region are extracted. Hence the

[email protected]
Fast and Accurate Fingerprint Recognition in Principal Component Subspace 307

Fig. 1 Steps in proposed Grayscale Fingerprint Query Image

matching technique
Alignment Using Proposed PC
Based Technique

Image Enhancement, Noise

Removal and Thinning

ROI ExtracƟon

Eigen Value DecomposiƟon of

ROI Image

Principal Components of ROI

Principal
Responsible for 95% Variance
Components of
Registered ROI
Responsible for Calculate Weighted Similarity
95% Variance Score

Match Yes score >

threshold?
Match

No
Not Match

entire fingerprint region, I is reduced to a set of 2D coordinate points, which can

be represented as I = {(xi , yi ) : I (xi , yi ) < 230, i = 1, 2, . . . , n} (where the value
of n depends upon the fingerprint template). Principal components of this 2D data
are calculated; wherein the first principal component(PC1) will give the axis of
maximum variance and the second principal component(PC2) will give the axis of
second most variance.
The orientation of fingerprint image w.r.t. current axis is calculated by measuring
the angle between PC1 and vertical axis. Different steps in the alignment process is
explained in Algorithm 2. Results of the alignment process are shown in Fig. 3. The
red lines in the input images shown in Fig. 3a, b represent horizontal and vertical
axes. The reference point calculated in Step 4 of the Algorithm 2 is used for visual-
ization of principal components. The blue lines in the figures represent the principal
components. The longest blue line represents PC1 which gives the axis of maximum
variance. The second blue line orthogonal to PC1 is PC2, which represents the axis
of second most variance. The angle estimated in Step 5 of the Algorithm 2 is the
angle between PC1 and the vertical axis. We will perform a clockwise rotation if
x-coordinate of PC1 is positive and we will perform counter clockwise rotation if

[email protected]
308 S. P. Ragendhu and T. Thomas

Fig. 2 Result of
preprocessing steps in intra
class images. a and b are the
aligned images after
enhancement. c and d are the
corresponding thinned
images. e and f are the
images after ROI extraction

x-coordinate of PC1 is negative. The alignment is only based on the angle of rotation
and no reference point is used for alignment. Results show that if the images have
acceptable amount of information (if it is not latent), proposed alignment process
can improve the performance of the system.

4.2 Fingerprint Matching

Proposed matching technique considers a fingerprint image as a multivariate data

and hence quantifies the similarity of fingerprint templates based on the similarity
of principal component subspaces. When data is represented in high dimensions,
it is difficult to analyze the similarity between them. Principal components basi-
cally represent optimized coordinate system to represent the given data. Hence it
becomes comparatively easier to compare and analyze the data which is represented

[email protected]
Fast and Accurate Fingerprint Recognition in Principal Component Subspace 309

Algorithm 2 FPAlign
Step 1: Input I = {(xi , yi ) : I (xi , yi ) < 230, i = 1, 2, . . . , n}
⎡ ⎤
x1 y1
⎢ . . ⎥
⎢ ⎥
Step 2: Form n × 2 matrix X Y , X Y = ⎢ . . ⎥
⎣. .⎦
xn yn
Step 3: Find the covariance matrix of X Y . Calculate the principal components corresponding
to the two eigen vectors of the covariance matrix using eigen value decomposition
α α
method as discussed in Section 3.1. Let PC1 = x1 and PC2 = x2 be the
α y1 α y2
principal components corresponding to the eigen values λ1 and λ2 , λ1 > λ2 .
n x n y
Step 4: Calculate the mean values x̃ = i=1 n , ỹ =
i i=1 i
n and set the reference point as
(x̃, ỹ).
Step 5: Calculate the angle(θ) between unit vector along Y axis (y) and PC1.
θ = cos−1 |y|.|PC1|
y.PC1

Step 6: In order to rotate the points X Y , we need to first translate it to origin by subtracting
the means. x̃i = xi − x̃, ỹi = yi − ỹ.
⎡ ⎤
x̃1 ỹ1
⎢. .⎥
⎢ ⎥
Step 7: Form the matrix of translated points X̃ Ỹ , X̃ Ỹ = ⎢ . . ⎥
⎣. .⎦
x̃n ỹn
⎡ ⎤
x̃1 ỹ1
⎢. .⎥
  ⎢ ⎥ cosθ sinθ
Step 8: Rotate the matrix X̃ Ỹ by the estimated angle θ, X Y = ⎢ . . ⎥ ×
⎣. .⎦ −sinθ cosθ
x̃n ỹn
 
Step 9: Translate the points back to original position, xi" = xi + x̃, yi" = yi + ỹ.

in terms of principal components. Principal components obtained corresponding to

each dimension will be compared with principal components of the query image. Two
similarity measures are experimented for estimating the similarity between principal
components of fingerprint templates.

4.2.1 Similarity Measure Based on Principal Components

If X 1 and X 2 represent two fingerprint ROI images of size p × q, PCA can model
X 1 and X 2 using k principal components each(k  q). If L and M represent corre-
sponding subspaces of size ( p × k) of each image, the similarity of these subspaces
can be used as a measure of similarity between the images. The subspaces L and
M also represent the eigen vector matrices corresponding to the first k eigen values,
which can be obtained by the eigen value decomposition of covariance matrices of
X 1 and X 2 . Hence PCA similarity factor SPCA , compares these subspaces and is
defined by the Eq. 4 [14]:

[email protected]
310 S. P. Ragendhu and T. Thomas

Fig. 3 Intra class fingerprint images before and after alignment. a and b are the input images. c
and d are the corresponding aligned images

trace(L T M M T L)
SPCA = , (4)
k
Basically we are checking only the similarity of orientation of principal compo-
nents in the case of PCA similarity factor (SPCA ). Principal components responsible
for providing at least 95% variance in each image are selected and similarity calcula-
tion is done in these subspaces formed by k principal components. Equation 4 can be
represented as the squared cosine values of the angle between principal components
of templates and is given in Eq. 5 [14]:

1 k
SPCA =   k cos2 θi j , (5)
k i=1 j=1
where k is the smallest number of principal components responsible for at least 95%
variance in each input biometric images and θi j is the angle between ith principal

[email protected]
Fast and Accurate Fingerprint Recognition in Principal Component Subspace 311

component of first image and jth principal component of second image. The value
of S PC A will range between 0 and 1, where 1 corresponds to maximum similarity.
Performance of SPCA was not desirable for fingerprint templates because of the
high FAR value. The results of SPCA based experiments were analyzed and it was
found that templates having similar structure with comparatively similar orientations
are the reason for increase in FAR. Hence we concluded that SPCA is more suitable for
structural classification of fingerprint images than person authentication. The reason
for poor performance of SPCA is that only orientation correspondences were involved
in the calculation of match score.

4.2.2 Weighted Similarity Measure Based on Principal Components

λ
Modified weighted PCA similarity(SPCA ) measure considers the contribution or
variance associated with each principal component along with orientation corre-
spondence while calculating the similarity score [15]. Calculation of the modified
similarity score is given by the Eq. 6:

λ
i=1
k
 kj=1 λi1 λ2j cos2 θi j
SPCA = , (6)
i=1
k
 kj=1 λi1 λ2j

where k is the number of principal components responsible for at least 95% variance
in each input biometric images, θi j is the angle between ith principal component of
first image and jth principal component of second image and λi 1 represents the eigen
value corresponding to ith principal component of first image and λ2j represents the
eigen value corresponding to jth principal component of second image. In Eq. 6 we
have represented the similarity w.r.t. all possible combinations of eigen values of two
images. Significant reduction of FAR to 6% is obtained against FRR of 11%.

5 Results and Discussions

Experiments are conducted using both FVC2000 and FVC2004 databases, wherein
8 images per person were present. We started experiments with fingerprint templates
of 40 different people available in Db2a of FVC2000 database (40 × 8 = 320).
After analyzing the results obtained from FVC2000 database, we extended our ex-
periments to FVC2004 also. Fingerprint images of 60 different people available in
FVC2004 fingerprint database (40 from D B2 A and 20 from D B2B) are also used
in experiments (60 × 8 = 480). Four images per person are used as training images
and remaining four images are used for testing. Cross validation is performed using
modified weighted similarity scores of all images. Figure 4 shows the values of FAR
against GAR at different thresholds obtained for FVC2000 and FVC2004. Accuracy

[email protected]
312 S. P. Ragendhu and T. Thomas

Fig. 4 Calculation of area

under curve (AUC)

Table 1 Accuracy calculated Database Accuracy (AUC) (%)

from Fig. 4
FVC2000 96.6
FVC2004 96.33

Table 2 Comparison of Method Accuracy at 2%FAR (%)

results
Jain et al. [10] 96.78
Cappelli et al. [4] 97.25
Proposed 97.31

is computed through Area under curve(AUC) method from Fig. 4. It is observed that
overall accuracy of proposed method is almost equal for the two databases, which is
given in Table 1.
Performance of the proposed system is compared with two major approaches in
fingerprint recognition. First one is filterbank-based system proposed by Jain et al.
[10] and second one is Minutiae Cylider Code(MCC) method proposed by Cappelli et
al. [4]. Both the approaches are tested against the same set of FVC2004 images. The
accuracy of the methods are calculated at different thresholds. Different FAR values
are obtained corresponding to different thresholds. Number of genuine acceptance
(True Positives) and genuine rejection (True Negatives) are calculated at each of the
selected FAR levels. Accuracy is calculated using Eq. 7. Comparison of accuracy of
methods at 2% FAR is shown in Table 2. We can observe that the proposed method is
able to obtain better performance compared to filterbank-based system and slightly
higher performance compared to MCC.

True Positives + True negatives

Accuracy = (7)
Total number Templates Tested

[email protected]
Fast and Accurate Fingerprint Recognition in Principal Component Subspace 313

Table 3 Average matching Method Average matching time

time (ms) in one to one
authentication Jain et al. [10] 18
Cappelli et al. [4] 15
Proposed 2

Table 4 Execution time (s) Method Execution time

in one to many identification
Jain et al. [10] 6
Cappelli et al. [4] 4.5
Proposed 1.2

We also compared the time taken for the proposed method with the other two
approaches. The two parameters we took for comparison are matching time and
execution time. Matching time means the time required for calculating the similarity
score and it does not include the time required for preprocessing techniques. It is
evident from Table 3 that the proposed method is far better in terms of matching time.
Execution time includes the time taken for entire steps in the approach, including the
time required for preprocessing steps. We calculated the execution time of proposed
method in case of one to many identification using a template database with 400
templates (4 images per person). Result is compared with the methods proposed by
Jain et al. and Cappelli et al. and the result is given in Table 4.
The most appreciable feature of the proposed method is that the execution time
for one to many identification is found to be very less compared to the other two
methods in terms of execution time. One of the notable applications of biometric
person identification is criminal identification by law enforcement agencies. In most
of the cases, the template database will be large and hence the existing schemes will
take pretty good amount of time to complete the searching. However, the performance

Fig. 5 Calculation of EER

[email protected]
314 S. P. Ragendhu and T. Thomas

of the proposed system show that it can reduce the search space, and converge to a
small set of matching templates in almost no time.
Average value of FAR and FRR at different thresholds are calculated and Fig. 5
shows the average FAR and FRR values plotted against the normalized threshold
range 0–0.25, and it is observed that EER of 10% is obtained at 0.10625.

6 Conclusion and Future Works

Fingerprint alignment and matching based on statistical properties of the templates

is a promising direction of thought. Vulnerabilities of minutiae-based approach raise
serious security and privacy concerns in fingerprint systems. Another serious draw-
back is the increased execution time due to highly complicated procedures. Results
show that acceptable matching performance can be ensured even from statistical
features based matching schemes without large computations. Analyzing the appli-
cability and performance of the proposed technique in multi-biometric systems and
designing suitable template protection mechanism for proposed system with minimal
complexity are the major future directions for research.

Acknowledgements This research is supported by Kerala State Planning Board project CEPIA
(2017–18).

References

1. Bahgat, G., Khalil, A., Kader, N. A., & Mashali, S. (2013). Fast and accurate algorithm for
core point detection in fingerprint images. Egyptian Informatics Journal, 14(1), 15–25.
2. Ballan, M., & Gurgen, F. (1999). On the principal component based fingerprint classification
using directional images. Mathematical and Computational Applications, 4(2), 91–97.
3. Bhattacharyya, S., & Chakraborty, S. (2014). Reconstruction of human faces from its eigen-
faces. International Journal, 4(1).
4. Cappelli, R., Ferrara, M., & Maltoni, D. (2010). Minutia cylinder-code: A new representation
and matching technique for fingerprint recognition. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 32(12), 2128–2141.
5. Damarla, S. K., & Kundu, P. (2011). Classification of unknown thermocouple types using
similarity factor measurement. Sensors & Transducers, 124(1), 11.
6. Feng, J., & Jain, A. K. (2011). Fingerprint reconstruction: From minutiae to phase. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 33(2), 209–223.
7. Ghany, K. K. A., Hassanien, A. E., & Schaefer, G. (2014). Similarity measures for fingerprint
matching. In Proceedings of the International Conference on Image Processing, Computer
Vision, and Pattern Recognition (IPCV) (p. 1). The Steering Committee of The World Congress
in Computer Science, Computer Engineering and Applied Computing (WorldComp).
8. Hong, L., Wan, Y., & Jain, A. (1998). Fingerprint image enhancement: Algorithm and perfor-
mance evaluation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(8),
777–789.

[email protected]
Fast and Accurate Fingerprint Recognition in Principal Component Subspace 315

9. Hsieh, C.-T., & Shyu, S.-R. (2007). Principal component analysis for minutiae verification on
fingerprint image. In Proceedings of the 7th WSEAS International Conference on Multimedia
Systems & Signal Processing, Hangzhou, China.
10. Jain, A. K., Prabhakar, S., Hong, L., & Pankanti, S. (2000). Filterbank-based fingerprint match-
ing. IEEE Transactions on Image Processing, 9(5), 846–859.
11. Johannesmeyer, M. C., Singhal, A., & Seborg, D. E. (2002). Pattern matching in historical data.
AIChE Journal, 48(9), 2022–2038.
12. Jolliffe, I. T. (1986). Principal component analysis and factor analysis. In Principal component
analysis (pp. 115–128). Springer.
13. Jolliffe, I. T., & Cadima, J. (2016). Principal component analysis: a review and recent devel-
opments. Philosophical Transactions of the Royal Society A, 374(2065), 20150202.
14. Krzanowski, W. (1979). Between-groups comparison of principal components. Journal of the
American Statistical Association, 74(367), 703–707.
15. Liao, T. W. (2005). Clustering of time series dataa survey. Pattern Recognition, 38(11), 1857–
1874.
16. Liu, E., & Cao, K. (2016). Minutiae extraction from level 1 features of fingerprint. IEEE
Transactions on Information Forensics and Security, 11(9), 1893–1902.
17. Msiza, I. S., Malumedzha, T. C., & Leke-Betechuoh, B. (2011). A novel fingerprint re-alignment
solution that uses the tfcp as a reference. International Journal of Machine Learning and
Computing, 1(3), 297.
18. Mudrova, M., Prochazka, A. (2005). Principal component analysis in image processing. In
Proceedings of the MATLAB Technical Computing Conference, Prague.
19. Nandakumar, K., & Jain, A. K. (2015). Biometric template protection: Bridging the perfor-
mance gap between theory and practice. IEEE Signal Processing Magazine, 32(5), 88–100.
20. Nandakumar, K., Jain, A. K., & Pankanti, S. (2007). Fingerprint-based fuzzy vault: Imple-
mentation and performance. IEEE Transactions on Information Forensics and Security, 2(4),
744–757.
21. Ramoser, H., Wachmann, B., & Bischof, H. (2002). Efficient alignment of fingerprint images.
In Proceedings. 16th International Conference on Pattern Recognition, 2002 (Vol. 3, pp. 748–
751). IEEE.
22. Ross, A., Shah, J., & Jain, A. K. (2007). From template to image: Reconstructing fingerprints
from minutiae points. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(4),
544–560.
23. Rozsa, A., Glock, A. E., & Boult, T. E. (2015). Genetic algorithm attack on minutiae-based
fingerprint authentication and protected template fingerprint systems. In 2015 IEEE Conference
on Computer Vision and Pattern Recognition Workshops (CVPRW) (pp. 100–108). IEEE.
24. Shen, W., & Eshera, M. (2004). Feature extraction in fingerprint images. In Automatic Finger-
print Recognition Systems (pp. 145–181). Springer.
25. Wang, J., Zhu, Y., Li, S., Wan, D., & Zhang, P. (2014). Multivariate time series similarity
searching. The Scientific World Journal.
26. Yang, J. (2011). Non-minutiae based fingerprint descriptor. InTech: In In biometrics.
27. Yongxu, W., Xinyu, A., Yuanfeng, D., & Yongping, L. (2006). A fingerprint recognition al-
gorithm based on principal component analysis. In TENCON 2006. 2006 IEEE Region 10
Conference (pp. 1–4). IEEE.
28. Zaeri, N. (2011). Minutiae-based fingerprint extraction and recognition. InTech: In biometrics.
29. Zhang, D., Liu, F., Zhao, Q., Lu, G., & Luo, N. (2011). Selecting a reference high resolution
for fingerprint recognition using minutiae and pores. IEEE Transactions on Instrumentation
and Measurement, 60(3), 863–871.

[email protected]
Smart Meter Analysis Using Big Data
Techniques

Neha Pandey, Nivedita Das, Sandeep Agarwal, Kashyap Barua,

Manjusha Pandey and Siddharth Swarup Rautray

Abstract In the present-day, many government firms and global companies pay
emphasis on energy conservation and efficient use of energy. The smart meter data
have mapped a way to use energy efficiently. The need to use energy in an efficient
way is very much required for developing countries like India. The emergence of
smart meter gave us access to huge amounts of energy consumption data. It is an
electronic component that records utilization of electric energy at regular intervals
of time, be it hours, minutes, or seconds. This paper proposes a different method
for grouping electricity consumption. Through smart meter, we get a huge amount
of energy consumption data. These data are analyzed by various energy distribution
companies which further leads to prediction of demand and consumption of user. Our
paper uses a business intelligence tool such as map reduction to handle these data
sets. Taking the advantage of this tool, energy distribution companies can reduce the
investment by making the use of community hardware. Using distributed computing
tools we can reduce the processing time appreciably to enable real-time monitoring
and decision-making. Further, R is integrated to it to perform analysis. Various data
sets are used to check the potential of the proposed models and approaches.

Keywords Smart meter data · Meter data analysis · Big data analytics · Hadoop ·
Map reduction · HDFS

N. Pandey (B) · N. Das · S. Agarwal · K. Barua · M. Pandey · S. S. Rautray

School of Computer Engineering, KIIT University, Bhubaneswar, Orissa, India
e-mail: [email protected]
N. Das
e-mail: [email protected]
S. Agarwal
e-mail: [email protected]
K. Barua
e-mail: [email protected]
M. Pandey
e-mail: [email protected]
S. S. Rautray
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 317

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_27

[email protected]
318 N. Pandey et al.

1 Introduction

The basic energy demand is the consumption of electricity. The analysis is being made
on energy consumption data to gain perception about the customer usage pattern.
It is the job of smart meter to keep the record of data that has been generated for
every minute. These data can be mined by the energy Utilities Company and new
insights thus produced can lead to business benefit. The data from smart meter are
raw stream and is also associated with high volume. For this, there is a requirement
of framework which can accumulate and handle massive raw data stream generated
by smart meter and to apply various logical and statistical techniques to correlate
events with different conditions and finally predict the result. Well, the resolution to
handle the above problem is Apache Hadoop. It is a tool which deals with distributed
computing, which has large storage property along with computing capability. The
framework of Apache Hadoop deals with the processing of large sets of data across
a clutch of computers [1]. Parallel processing of data set is done by Hadoop and
MapReduce. Unstructured and not originally intended for machine processing
• Data: A set of values of qualitative or quantitative variables is known as data.
The idea of data is usually associated with scientific research. It is collected over
by a massive range of institutions and corporation including governmental, non-
governmental, and business firms. Data can be analyzed either by measuring and
collecting the information or by visualizing the graphs and images.
• RDBMS: A database management system (DBMS) that is built upon the relational
model is called a relational database management system (RDBMS). It has been
one of the most popular choices for the storehouse of statistics. It is used for
commercial and economical documentation, logical information, private data, and
other applications for a long period of time [2]. Relational databases have often
replaced other database techniques, because they are familiar in understanding and
using. RDBMSs are used to store various kind of information.
• BIG DATA: Big data consists of data sets that are so massive and tangled that
normal data processing applications are not sufficient. Managing data, storing,
analyzing, finding, sharing, shifting, conceptualizing, querying, updating, infor-
mation privacy, and data source [3]. Although big data is not specific to equate
any amount of data, the term can be used for attending various bytes of data. Data
might contain large variation in file types such as documents, sensors, images.
• HADOOP: The data generated by smart meter is sent by energy consumption
data to the server at a regular interval of time leading to the generation of massive
amount of data and existing tools were not functioning well in handling such
massive data. It is an open architecture for the development of scattered application
that can pass over very large amount of data [4]. It is a base which provides us
with various cache and computing techniques. The advantages of Hadoop are:
(1) It is an extraordinarily ascendible framework, as it accumulates and distributes
extremely massive data sets through several of retailers that work in alignment. (2)
Hadoop is a method of low-cost effective cache solution for businesses’ exploding

[email protected]
Smart Meter Analysis Using Big Data Techniques 319

data. (3) Hadoop allows the businesses to easily reach to the fresh data sources
and wander into various types of data to create attributes from that data (4).
• Map Reduce: It is a kind of computing architecture which is also a similar exe-
cution for handling and promoting massive data values along with collateral, dis-
tributed algorithm on a clutch [5, 6]. This framework consists of a Map() method
to performs sieving and categorizing and a Reduce() method to performs a wrap-
ping up operation. The “Map Reduce System” (also called “infrastructure” or
“framework”) organizes the processing through collecting the distributed retain-
ers, coordinating number of tasks in parallel, handling different communique and
data transmission between the various parts of the framework, and providing for
redundancy and fault tolerance [7].
• HDFS: The Hadoop Distributed File System (HDFS) is a scattered file system
modeled to run on commodity hardware [8, 9]. It is similar to extant distributed
file systems. Still, the variation from other is consequential. It is very much fault-
tolerant and is modeled to be redistributing on less-expensive hardware. It provides
fast rate access to data and is relevant to large data sets applications [10]. It was
initially designed as groundwork for the Apache engine project.
• ENERGY: Energy basically deals with power derivation. It deals with derivation
of power from chemical and physical resources, mainly to provide light and heat
and to work on machines [11, 12]. It is a kind of conserved quantity. The basic fact
about energy is that, “Energy can be converted in form, but it cannot be created or
destroyed”.

1.1 Smart Meter Data Analysis

Electricity generation is dependent on unrenewable resources [13, 14]. Though India

is wealthy in these resources, they are extinct at a very speed which means they will be
almost gone soon whereas the renewable resources are not used to their best ability.
Keeping these points in consideration the government has taken many measures
towards the improvement of the electricity system [15]. The concept presented in
this paper is an optimal procedure which if carried away properly will lead to an
efficient and well-mannered electricity system in the future (Fig. 1).

1.1.1 Application of Big Data

Big data has a very vast application scenario be it in manufacturing, health care,
education, media, government or it industries. It has increased the demand for man-
agement of the data sets [14]. In the government sector, the usage of big data provides
efficiency in terms of cost, productivity, and usage [16]. Whereas in health care id
helps in keeping record of various rage of data related to disease, patient, range of
infection. Basically in every aspect of today’s world big data has been applied. One

[email protected]
320 N. Pandey et al.

Fig. 1 Application of big

data
Bank and Healthcare
Insurance and research

Big Data
Application

Energy and Telecom

Utilities Industry

Fig. 2 Characteristics of big

data Volume

Velocity Variety

Value Veracity

of its major applications is in energy consumption on which this paper has been
worked upon (Fig. 2).

1.1.2 Components of Big Data

The basic five components of big data are: Volume; It basically deals with the amount
of data generated. Variety; What kind of data are we dealing with and what is its
nature comes under variety. Velocity; The speed of the generation of data basically
comes under this category. Variability; [17–19] This deals with the consistency of
data. Varacity; This deals with the quality of data (Fig. 3).

2 Related Work

See Table 1.

[email protected]
Smart Meter Analysis Using Big Data Techniques 321

Table 1 Previous work done in smart meter analysis

Sl. No. Year Author Paper title Paper Tools and techniques
description
1 2015 Balaji K Analysis For efficient 1. Smart meter
Bodkhe of smart energy 2. Apache hadoop
et al. meter data consumption,
using apache hadoop
hadoop is used, which
uses the data
generated by
smart meter for
predicting
energy
consumption
2 2014 Javier Analysis In this paper 1. Media analysis
Conejero of hadoop investigation 2. Instrumentation
et al. power and and monitoring
consump- measurement is
tion done on energy
consumption
[20]
3 2017 Dr. Analyzing This paper 1. SAX
Mohammed the proposes a novel 2. Markov model
Abdul behavior approach for 3. Distributed
Waheed of clustering of algorithm for
et al. electricity electricity large data sets
consump- consumption
tion using behavior
hadoop dynamics
4 2017 Yimin A hierar- In this paper an 1. Energy
Zhou et al. chical energy consumption
system of consumption modeling
energy monitoring and 2. Information
consump- Information management
tion management
monitor- system based on
ing and energy
informa- conservation
tion model is
manage- developed [18]
ment
system

[email protected]
322 N. Pandey et al.

Fig. 3 Proposed framework

Smart meter

3 Proposed Methodology

The system that is implemented on smart meter data is providing a result for an
electricity provider. Smart meter produces huge amount of data which is handled
by Hadoop. The data has been processed through MapReduce and HDFS is used to
store these data which includes various components like date, time, active, reactive,
meter number, voltage, etc. MapReduce programming is done on these data sets that
produces key/value pair. R language is used for this work. In this module, we have
to create data set for electricity consumption which we usually get from smart meter
data. It consists of various attributes such as customer details, billing details, payment
details for the last few years. These data set has been loaded in HDFS from the local
file system. We use HDFS to store massive amount of data which is later exported
to R for load profile analysis. MapReduce is a preparing strategy and a program for
appropriated processing. The MapReduce calculation contains two vital activities,
in particular, Map and Reduce. In this module likewise utilized for dissecting the
informal index using MapReduce (Fig. 4).

Fig. 4 Methodology

[email protected]
Smart Meter Analysis Using Big Data Techniques 323

4 Implementation

4.1 Understanding the Problem and Extracting Data

This phase is quite essential in determining the success of the forecast. The main
focus is made on extracting data and exploring those data from various industries
and sub-industries. After analyzing the data simple forecast models are predicted.
The data set is obtained from the site—https://open-enernoc-data.s3.amazonaws.
com/anon/index.html. This data set consists of time series of various consumers and
their corresponding metadata. Exploring and cleaning of data is done with help of a
package named data. table. For visualization of relations ggplot package is used and
for manipulation of date and time lubricate is used. Initially a frequency table for
industry and sub-industry was made. We use the package gmap to map the location
of our consumer on the map of use. After plotting the map we calculate the sq_m
(square meter) of buildings for all the consumers. Further mean load for the sub-
industry and median load for the industry was calculated (Figs. 5, 6 and 7; Table 2).

Fig. 5 Location of
consumer

Fig. 6 Mean of sub-industry

[email protected]
324 N. Pandey et al.

Fig. 7 Median of industry

Table 2 Plot for industry and Industry Sub-industry No.

sub-industry
1 Commercial property Shopping center, 14
shopping mal
2 Commercial property Corporate office 2
3 Commercial property Business services 3
4 Commercial property Commercial real 4
estate
5 Commercial property Bank financial 2
services
6 Education Primary secondary 25
school
7 Food sales and storage Grocer market 25
8 Light industrial Food processing 19
9 Light industrial Manufacturing 5
10 Light industrial Other light industrial 1

4.2 Implementation and Testing

In this phase data set is prepared to forecast and explore time series of load. Initially,
we need to transform all the attributes of dates to classic date and time. The unwanted
columns are being removed and we observe that the current structure of the present
data set is full of time series. Then ids are being extracted (Figs. 8 and 9).
Now we need to fetch all the dates with all features taking the dimension, 288/days,
now we will randomly check for any one id and its corresponding time series (Figs 10
and 11).
We observe that there is a strong dependency on time. So it would be better if
the time series is aggregated at lower dimensions, around 48/day and once it is done
then the plotting is done for four sub-industries providing them id’s respectively.

[email protected]
Smart Meter Analysis Using Big Data Techniques 325

Fig. 8 Screenshot of extraction of id

Fig. 9 Observation of consumption versus time series for 1 day

4.3 Evaluation of Result

During this phase, we consider all the data set from the previous implementation
made and try to calculate the aggregate consumption by the user. For the forecast of
electricity consumption, this aggregate value is used. Aggregate for all the customer
is 45 and it has been plotted.

[email protected]
326 N. Pandey et al.

Fig. 10 Load versus date for 4 industries

Fig. 11 Plot of aggregate consumption of load versus time series

[email protected]
Smart Meter Analysis Using Big Data Techniques 327

5 Conclusion

In conclusion, we see that the project concentrates on the forecast model. This model
helps us to make analysis about the daily consumption of various consumers of
different industries and sub-industries with respect to time series. While we construct
the histogram for plotting various consumers on map we observe that majority of the
buildings are under 20,000 m2 . Through density plot we analyze that commercial
buildings have viable size whereas food sales and storage buildings have smaller size.
Through plotting the mean we get to know that the largest consumer of electricity
is manufactures, shopping center, and business services whereas schools have the
lowest consumption. The medial plot tells about the viable relation between the load
and sq_m area. The forecast models help us to find the aggregate consumption to be
45%. The forecast model also helps to find out daily profile of consumer or word
area.

6 Future Scope

In this paper, we have not made a forecast model for weekly or monthly consumption
which can also be done further. These models can be further used to make compar-
isons between models of two different attributes. ARIMA can be used to predict more
accurate forecast. By making the observation on consumption we can mark those
areas where consumption is high and try to reduce it for conserving energy. In future
methods like regression or neural network approach can be used for forecasting.

References

1. Herodotou, H., et al. (2011). Starfish: A self-tuning system for big data analytics. In Cidr (Vol.
11. No. 2011).
2. Quilumba, F. L., et al. (2015). Using smart meter data to improve the accuracy of intraday load
forecasting considering customer behavior similarities. IEEE Transactions on Smart Grid, 6(2),
911–918.
3. LaValle, S., et al. (2011). Big data, analytics and the path from insights to value. MIT Sloan
Management Review, 52(2), 21.
4. Roy, C., Pandey, M., & Rautaray, S. S. (2018). A proposal for optimization of horizontal
scaling in big data environment. In Advances in data and information sciences (pp. 223–230).
Singapore: Springer.
5. Thusoo, A., et al. (2009). Hive: A warehousing solution over a map-reduce framework. Pro-
ceedings of the VLDB Endowment, 2(2), 1626–1629.
6. Chu, C.-T., et al. (2007). Map-reduce for machine learning on multicore. In Advances in neural
information processing systems.
7. Patel, A. B., Birla, M., & Nair, U. (2012). Addressing big data problem using hadoop and
map reduce. In 2012 Nirma University International Conference on Engineering (NUiCONE).
IEEE.

[email protected]
328 N. Pandey et al.

8. Deshmukh, A. P., & Pamu, K. S. (2012). Introduction to hadoop distributed file system. IJEIR,
1(2), 230–236.
9. Ananthanarayanan, G., et al. (2010). Reining in the outliers in map-reduce clusters using Mantri.
OSDI, 10(1).
10. Sitto, K., & Presser, M. (2015). Field guide to hadoop: an introduction to hadoop, its ecosystem,
and aligned technologies. O’Reilly Media, Inc.
11. Feeney, L. M., & Nilsson, M. (2001). Investigating the energy consumption of a wireless
network interface in an ad hoc networking environment. In Proceedings of the Twentieth Annual
Joint Conference of the IEEE Computer and Communications Societies INFOCOM 2001 (Vol.
3). IEEE.
12. Efthymiou, C., & Kalogridis, G. (2010). Smart grid privacy via anonymization of smart meter-
ing data. In 2010 First IEEE International Conference on Smart Grid Communications (Smart-
GridComm). IEEE.
13. Kavousian, A., Rajagopal, R., & Fischer, M. (2013). Determinants of residential electricity
consumption: Using smart meter data to examine the effect of climate, building characteristics,
appliance stock, and occupants’ behavior. Energy, 55, 184–194.
14. Borthakur, D., et al. (2011). Apache hadoop goes realtime at facebook. In Proceedings of the
2011 ACM SIGMOD International Conference on Management of Data. ACM.
15. Michael, Katina, & Miller, Keith W. (2013). Big data: New opportunities and new challenges
[guest editors’ introduction]. Computer, 46(6), 22–24.
16. Das, N., et al. (2018). Big data analytics for medical applications.
17. Jiang, L., Li, B., & Song, M. (2010). The optimization of HDFS based on small files. In
2010 3rd IEEE International Conference on Broadband Network and Multimedia Technology
(IC-BNMT). IEEE.
18. Szargut, J., Morris, D. R., Steward, F. R. (1987). Exergy analysis of thermal, chemical, and
metallurgical processes.
19. Borthakur, D. (2008). The hadoop distributed file system: Architecture and design.
20. Liu, X., et al. (2009). Implementing WebGIS on hadoop: A case study of improving small
file I/O performance on HDFS. In IEEE International Conference on Cluster Computing and
Workshops, 2009. CLUSTER’09. IEEE.
21. Leverich, J., & Kozyrakis, C. (2010). On the energy (in) efficiency of hadoop clusters. ACM
SIGOPS Operating Systems Review, 44(1), 61–65.

[email protected]
Movie Recommendation System

S. Rajarajeswari, Sharat Naik, Shagun Srikant, M. K. Sai Prakash

and Prarthana Uday

Abstract Recommender systems are called information filtering tools, which use
big data to recommend likes of the user according to their preference and interest.
Moreover, they also help in matching users with similar tastes and interests. Due
to this, a central part of websites and e-commerce applications is taken up by the
recommender systems. Systems using recommendation algorithms like Collabora-
tive Filtering, Content-Based Filtering, etc., are called Recommendation systems.
Recommendation systems are transforming the way quiescent websites corresponds
with their users. Rather than providing an orthodox experience in which users search
for the products which they want and potentially buy products, recommender sys-
tems increase communication to provide a better experience. The work would be
to implement both collaborative as well as content-based recommenders available
and try to extend the knowledge obtained to a more efficient hybrid model. Then
benchmark these hybrid algorithms for accuracy and computation time as well.

Keywords Recommendation systems · Collaborative · Cosine similarity · SVD ·

Machine learning · Python · Surprise library

1 Introduction

It is expected that in future with the increased use of computers, the use of advanced
and latest technologies by users and professionals and decision-makers will increase
a lot. The basic idea of recommendation systems is that, if some users share the same
interests, or maybe in the past, e.g., they liked the same book, they might also have
similar tastes in the future. Most recommender systems take three basic approaches:
simple recommender, collaborative filtering, and content-based filtering. A hybrid
approach can combine these approaches to come up with a more optimal solution.
A recommendation system has become an indispensable component in various e-
commerce applications. Recommender systems collect information about the user’s

S. Rajarajeswari (B) · S. Naik · S. Srikant · M. K. Sai Prakash · P. Uday

Department of Computer Science Engineering, RIT, Bangalore 560054, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 329

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_28

[email protected]
330 S. Rajarajeswari et al.

preferences of different items (e.g., movies, tourism, TV, shopping,) by two ways,
either implicitly or explicitly.
A movie recommendation system ideally analyzes factors like review, cast, plot,
crew, genre, popularity and comes up with the best possible result. It takes into
account a user preference and hence provides results based on the user’s personal taste
and choice. By increasing efficiency and accuracy of recommender systems, users
expect lightning fast results and different types of recommender systems available
today provide the same. There are four different kinds of recommendation systems,
improving the accuracy and user personalization with each.
The four systems are the following:
Simple Recommendation System: A simple recommender is the most basic rec-
ommendation system available to us. It usually compares a metric/weight for all the
available entities and provides the recommendation on the basis of this comparison.
Content-Based Recommendation System: This recommendation system
involves the use of more attributes from the dataset to provide a well-filtered rec-
ommendation to the user. In this project, at first, the recommender just provides a
recommendation using the movie overview and taglines. But there is a possibility
that a user watches a certain movie for the director or the cast rather than the ratings.
For this reason, the recommender takes into consideration the director, cast, genre,
and keywords to provide a better search result. This algorithm makes use of k-means
algorithm through cosine similarity to find how similar two movies are and hence
can be recommended together.
Collaborative Recommendation System: Both the recommendation systems
mentioned above are efficient but they are not personalized. They would show the
same results to all the users. But that should not be the real case. Movies should be
recommended based on the user’s personal taste. That is where collaborative filtering
comes into the picture.
Here, the movie ratings provided by a particular user and using the SVD sur-
prise library functions, evaluate the RMSE and MAE values. Since these values are
comparable, a predict function is used that provides an estimate for a movie for that
particular user using his ratings. This way, the movie recommendations provided to
each user would be different depending on their tastes.
Hybrid Recommendation System: Having demonstrated all the three available
recommendation models, a hybrid system is developed which uses the concepts of
both content base as well as collaborative recommendation systems. Here, the user’s
ID and movie title are taken as the inputs. Using the movie title, 30 similar movies
are collected in a database based on the title, cast, crew, overview, etc., using the
cosine similarity and linear kernel concept. And then using the user’s ID, his ratings
are procured which is then passed as an argument to the svd.predict(). Through the
list of those 30 movies, the ones most suited to the user’s taste are recommended.

[email protected]
Movie Recommendation System 331

2 Related Work

In a Movie Recommender System called “MOVREC” by Manoj Kumar, D. K. Yadav,

Vijay Kr Gupta.
It uses collaborative and content-based filtering using the k-means algorithm.
This asks a user to select his own choices from a given set of attributes and then
recommends him a list of based on the cumulative weight of different attributes
using k-means algorithm [1].
The disadvantage is that it does not use large datasets, hence there will not be
meaningful results.
In a paper by Rahul Kataria and Om Prakash named An effective movie recom-
mender system which uses cuckoo search, a bio-inspired algorithm such as cuckoo
search has exclusive background sensing abilities and employs a special method to
facilitate the evolution of continuing resolutions into novel and quality recommen-
dations by generating clusters with reduced time [2].
It has a limitation where if the initial partition does not turn out well then the
efficiency may decrease.
In 2007 Weng, Lin, and Chen performed an evaluation study which says using
multidimensional analysis and additional customer’s profile increases the recom-
mendation quality. Weng used MD recommendation model (multidimensional rec-
ommendation model) for this purpose. Multidimensional recommendation model
was proposed by Tuzhilin and Adomavicius [3].

2.1 Sajal Halder: Movie Recommendation Using Movie

Swarm

Movies swarm mining that mines a set of movies, which are suitable for producers and
directors for planning new movie and for new item recommendation. Popular movie
mining can be used to solve new user problems. It has the capability of handling both
new users and new items. It is better than content-based and collaborative approaches
when it comes to new users. This method, however, has a drawback of finding a group
of users depending on the genre [4].
Yueshen Xu: Collaborative recommendation with User-Generated Content
(UGC).
It proposes a UGC-based collaborative recommendation which integrates proba-
bilistic matrix factorization and collaborative regression models. It reaches a higher
prediction accuracy than most of the available models. It is efficient when it comes
to cold start problem. The cost per iteration is more as compared to traditional CTR
models [5] (Table 1).

[email protected]
332 S. Rajarajeswari et al.

Table 1 Comparative study related works and algorithms

S. No. Technique Algorithm Drawback
1 Collaborative and content K-means It is not suited for large
based datasets
2 Cuckoo search Clustering Efficiency decreases if the
initial partition is not proper
3 Movie swarm Mining Drawback of finding a group
of users based on the genre
4 User-generated content Regression Cost per iteration is more
(UGC) analysis when compared to
traditional models

3 Design

See Table 2.
General Architecture Diagrams
See Fig. 1.
Simple Recommender System
In Fig.2 Simple recommender gives a prediction based on the weighted rating (WR)
calculated. It takes in vote_ count and vote_average from Movies_metadata.csv.
The TMDB Ratings is used to come up with our Top Movies Chart. IMDB’s
weighted rating formula is used
Mathematically, it is represented as follows:

Weighted Rating (WR)  (v/v + m.R) + (m/v + m.C)

Table 2 Description of Dataset name (.csv) Description

attributes and dataset used in
the project Movies_metadata Contains information about 45,000
movies featured in TMDB
Keywords Contains movie plot keywords
available in the form of a stringified
JSON object
Credits Contains cast and crew information,
also saved as a stringified JSON object
Links Contains TMDB and IMDB Ids of all
the movies
Links_small Contains a subset of links database
with 9000 movie entries
Ratings_small Contains a subset of 100,000 ratings
from 700 users on 9000 movies

[email protected]
Movie Recommendation System 333

Fig. 1 General architecture of the recommendation system

Movies metadata.csv

vote_count
vote average

Weighted Rating

OUTPUT

Fig. 2 Simple recommender

[email protected]
334 S. Rajarajeswari et al.

Movies_metadata.csv
Links_small.csv

Keywords, cast, crew,

taglines

Cosine Similarity

OUTPUT

Fig. 3 Content-based recommender

Content Based Recommender System

In Fig.3, Content-based recommender gives prediction based on cosine similarity.

It takes in keywords, taglines from Movies_metadata.csv & links_small.csv, it also
takes director and cast as attributes.
To personalize our recommendations more, an engine is built that computes the
similarity between movies based on certain metrics and suggests movies that are
most similar to a particular movie that a user liked previously. Since movie metadata
(or content) is used to build the engine, this is also known as Content-Based Filtering.
Cosine similarity is used to calculate a numeric value, which denotes the similarity
between two movies.
Collaborative Recommender System
In Fig.4, Collaborative recommender gives prediction using singular value decom-
position (svd). It takes in userID and movieID from ratings_small.csv. This is more
personalized as it gives different suggestions to different users based on their taste.
A technique called Collaborative Filtering is used to make recommendations to
People. Collaborative Filtering is based on the idea that users similar to me can be
used to predict how much I will like a particular product or service those users have
used/experienced but I have not.
Surprise library is being used that uses extremely powerful algorithms like Sin-
gular Value Decomposition (SVD) to minimize RMSE (Root Mean Square Error)
and gives great recommendations to users.
Mean value is obtained for Root Mean Square Error, then we train our dataset and
arrive at predictions.

[email protected]
Movie Recommendation System 335

Hybrid Recommender System

In Fig.5, Hybrid recommender gives a prediction based on the svd and cosine simi-
larity. It uses Movies_metadata.csv as well as ratings.csv.
In this, content and collaborative filtering techniques are brought together.
Input given will be User ID and Title of a movie the user likes.
Output will be similar movies sorted on the basis of expected ratings by that
particular user which is more personalized.
Pseudocode:

a. Simple recommendation system

build_chart(genre)
Input: genre to be searched for.
Output: Top 250 recommendation

Fig. 4 Collaborative
recommender Ratings_small.csv

UserID, MovieID
Ratings

Singular Value
Decomposition

OUTPUT

Fig. 5 Hybrid recommender

Ratings_small.csv
Movies metadata.csv

SVD
Cosine Similarity

OUTPUT

[email protected]
336 S. Rajarajeswari et al.

df movies of ‘genre’

vote_counts number of votes for the genre
vote_average average votes for the genre
C=vote_averages.mean()
m=vote_counts.quantile(percentile)
v=x[‘vote_count’]
R=x[‘vote_average’]
qualified[‘wr’] (v/(v+m) * R) + (m/(m+v) * C)
return qualified

b. Content-Based recommendation system

get_recommendations(title)
Input: Name of a movie
Output: Top 30 movie recommendation
Idx  id of the title
keywords get keywords. Strip them of spaces.
Director get_director(crew){if (job==’director’) return name}
Cast first three names in the cast
Count-_matrix=Get the count vectorizer matrix to calculate the cosine similarity
using keywords, director, cast.
Cosine_similarity=Get the cosine similarity.
Return result.

c. Collaborative Recommendation System

recommendations(userId, title)
Input: User’s ID and title of the movie
Output: Movies which cater to user’s preference

Data  Use user id and movie id to get this

svd  Calling the SVD function.
Evaluate RMSE and MAE
Train data set to arrive at function
svd.train(trainset)
Predict using SVD and return the movies.

Proposed System

The proposed hybrid system takes features from both collaborative as well as content-
based recommendation to come up with a system, which is more personalized for
users.

[email protected]
Movie Recommendation System 337

Technique
Cosine similarity technique is used to come up with movies similar to the one given
in the search engine. The selected 30 movies are then compared using the user’s
ratings and the SVD. Hence, the search advances to predict the movies most suitable
for the user.
Algorithm:

The algorithm used for the hybrid engine is as follows:

def hybrid(userId, title):
Obtain the movieID (Idx) based on the title.
Obtain the sim_scores using cosine_similarity on the procured Idx
Use cosine similarity to obtain 30 movies similar to Idx using ‘vote_count’, ‘year’,
‘vote_average’ etc.
Using svd.predict(), calculate the RMSE between the suggested movies as well as
the movie passed as an argument.
Sort the list obtained.
Return the top 10 movies of the list.

4 Results

Content and Collaborative filtering techniques are brought together to build an engine
that gives movie suggestions to a particular user based on the estimated ratings that
it had internally calculated for that user. It makes use of cosine similarity concept to
find similar movies and the SVD prediction to estimate the movie recommendation
as per the user’s taste.
Simple Recommendation System:
See Fig. 6.
Content-Based Recommendation System:
See Fig. 7.
Collaborative Filtering:
In Fig. 8, the RMSE and MAE values are evaluated by using SVD Algorithm
available in Surprise library to minimize the RMSE.
Hybrid Recommender System:
See Fig. 9.
For our hybrid recommender, we get different recommendations for different users
although the movie is the same. Hence, our recommendations are more personalised
and tailored toward particular users.

[email protected]
338 S. Rajarajeswari et al.

Fig. 6 Top 250 movies based on the popularity/user ratings

Fig. 7 Similar 30 movies based on overview, taglines, director, cast

[email protected]
Movie Recommendation System 339

Fig. 8 The RMSE and MAE Values

Fig. 9 Movie prediction for User ID 500 with movie given as ‘Avatar’

5 Conclusion

Upon surveying through different algorithms and models present for movie recom-
mendation systems, we try to come up with a hybrid recommendation algorithm in
order to provide the most accurate recommendation to a user based on his ratings
and preference.
Recommendation systems of the future will work in e-commerce to offer a more
visceral, immersive, and well-rounded experience for every step of a customer’s
journey. In addition, once a successful hybrid engine comes into the picture, the same
algorithm can be extended to other types of recommendation systems. This would

[email protected]
340 S. Rajarajeswari et al.

result in an economic boom for the various e-commerce websites and applications
with better user-specific experience.

References

1. Kumar, M., Yadav, D. K., Singh, A., & Gupta, V. K. A movie recommender system “MOVREC”.
https://pdfs.semanticscholar.org/e13e/b41de769f124b3c91771167fb7b01bc85559.pdf.
2. Katarya, R., Verma, O. P. An effective collaborative movie recommender system with
cuckoo search https://ac.els-cdn.com/S1110866516300470/1-s2.0-S1110866516300470-
main.pdf?_tid=7a06b6fe-c95d-11e7-b816-00000aacb362&acdnat=1510679077_
68b65f60dcc55e2aeeacb3588dab49e4.
3. Hande, R., Gutti, A., & Shah, K. Moviemender—A movie recommender system. https://pdfs.
semanticscholar.org/e13e/b41de769f124b3c91771167fb7b01bc85559.pdf.
4. Halder, S. Movie recommendation system based on movie swarm. http://www.academia.edu/
2396000/Movie_Recommendation_System_Based_on_Movie_Swarm.
5. Xu, Y. Collaborative recommendation with User Generated Content (UGC).
https://www.researchgate.net/profile/Yueshen_Xu/publication/280733284_Collaborative_
Recommendation_with_User_Generated_Content/links/55c3cf3408aeb97567401ddb/
Collaborative-Recommendation-with-User-Generated-Content.pdf.

[email protected]
A Semiautomated Question Paper
Builder Using Long Short-Term Memory
Neural Networks
Rajarajeswari Subramanian, Akhilesh P. Patil, Karthik Ganesan
and T. S. Akarsh

Abstract Through this research, we propose a model to generate different sets of

question papers automatically using deep learning methodologies. We first develop
LSTM classification models to classify questions into Bloom’s level of taxonomy,
chapter name and we have an LSTM prediction model to predict the marks to be
allocated. These Deep Learning techniques can help in reducing human effort in
deciding the marks, the section, and Bloom’s level to be allocated to a question.
Given a pool of hundreds of questions, our deep learning models develop a knowl-
edge base consisting of questions and predicted attributes marks, Bloom’s level, and
chapter name. We then we have a randomization algorithm to pick the questions for
different units of the question paper, keeping the standards to be maintained and even
distribution of questions across all topics.

Keywords Bi-directional long short-term memory networks · Randomization

technique · Classification

1 Introduction

In the current era of artificial intelligence, most of the processes are being automated.
This has eliminated the human effort required and improved the accuracy with which
the process can be carried out. Automation has greatly helped the industries and is
yet to find its application in the educational background. Considering this, we have
proposed a model which may help educational institutes automatize the process of
generating a question paper. We can train the model in such a way that the marks,
bloom’s level, and chapter classification of a particular question can be predicted
from the question as an input.

R. Subramanian · A. P. Patil (B) · K. Ganesan · T. S. Akarsh

Department of Computer Science and Engineering, Ramaiah Institute of Technology, Bengaluru,
India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 341

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_29

[email protected]
342 R. Subramanian et al.

1.1 Obtaining the Dataset

The data set consists of the set of questions from a textbook. The Bloom’s level,
chapter name, and the marks distribution is assigned according to each question and
appended to the database. The entry to each of the categories is done in the form
of a CSV file. This CSV file is read in the form of data-frames with the help of
Pandas Library in Python to make the manipulation of data easier. The snapshot of
the dataset is as shown in the below Figure.

1.2 Pattern of the Question Paper

The question paper to be set is of 200 marks, out of which the student is to attempt
questions for 100 marks. The question paper consists of 5 individual units of 40
marks each. The student has to attempt either one of the questions from each unit
which is 20 marks each. Each question also has a Bloom’s Level associated with it.
Our aim in building the question paper is to ensure that there are two questions in
each unit of twenty marks each and the Bloom’s Level is evenly distributed.

2 Related Work

In the paper cited in [1], a randomization technique was proposed to automatically

generate a question based on the questions fed to the system according to a particular
syllabus. The technique proposed here, which is a role-based model also takes care
that the questions are not repeated. The paper cited in [2] also uses a randomization
technique to generate sets of question papers, which are unique from each other.
This method ensures that the questions generated are not the ones as in a particular
question bank provided to students. A research cited in [3], used J2EE tools to
design an automated question paper management system. Separate management
modules were developed for the user, subject, and the classification of questions.
The algorithm used was efficient enough to identify the subject, question type, and
also the difficulty level. Some researches as cited in [4] also used fuzzy logic to
generate non-repetitive question paper that distributed questions evenly throughout
the chapters. Similarly, papers cited in [5] select questions based on the Blooms
Levels of Taxonomy. This model was developed based on genetic algorithms. In the
paper cited in [6], an adaptive technique was proposed for the generation of question
paper, but the drawback of this model was that it assumes the entry of questions made
in the database are error free. The question paper generator proposed in our research
is different from all the other researches in the sense that separate classification model
was built using the bidirectional Long Short-Term Memory Networks for each of

[email protected]
A Semiautomated Question Paper Builder Using Long … 343

the attributes. The attributes taken into consideration for classification were chapter
name, marks categorization and Bloom’s Level of Taxonomy.

3 Techniques Used

The deep learning methodologies like Long Short-Term Memory Networks, Bidi-
rectional Long Short-Term Memory Networks have been used to classify the text
in the questions into slots such as the mark allotment to each question, unit-wise
markings and Bloom’s level of taxonomy. The following section gives a brief about
the working of each of these models to serve our cause.

3.1 Bidirectional Long Short-Term Memory Networks

Bidirectional LSTMs can extract the most out of an input sequence by running over
the input sequence both in the forward direction as well as in the backward direction.
The architecture is developed by creating another copy of the first recurrent layer
in the neural network so that there are two layers side-by-side. The next step is to
provide the input sequence as it is to the first layer and feeding an inverted copy to
the second duplicated layer. This approach is effectively used along side the Long
Short-Term Memory Networks. The bidirectional LSTM is as shown in Fig. 1.

3.2 Long Short-Term Memory Networks

Humans do not start to think from scratch every time they read some new article. As
a person reads an article, he or she reads every word based on the understandings of
the previous words. Our thoughts can retain the things of the past. A simple neural
network will not have the ability to do so. Recurrent neural networks can overcome
this shortcoming, due to the presence of a looping methodology in the hidden layers.
Long Short-Term Memory networks are a special kind of Recurrent Neural Net-
works, which have the capability of learning from previously computed output
results. In our problem where we predict the sequence of words that have to be
a part of the question, the whole idea here is that the sentence in our data frame is
a sequence of words which we may consider as a vector. The RNN that obtains the
vectors as input and considers the order of vectors to generate predictions. From the
embedding layer, the new representations will be passed to LSTM cells. These will
add recurrent connections to the network, so we can include information about the
sequence of words collected. Finally, the cells of the LSTM will go to the sigmoid
output layer. We use a sigmoid because we are trying to predict the final understand-
ing of the sentence. LSTMS have chain-like structure, the repeating module has a

[email protected]
344 R. Subramanian et al.

INPUT

FORWARD BACKWARD
LSTM LSTM

DENSE

OUTPUT

Fig. 1 The functioning of the bidirectional long short-term memory network

Fig. 2 First step in the LSTM

different structure. Instead of having a single neural network layer, there are four,
interacting in a very special way. The basic structure of the LSTM architecture is as
shown in Fig. 2.
In the first step, the LSTM model decides by a sigmoid layer, looks at a previously
computed hidden layer- h(t − 1) and the present output- x(t) and outputs a number
between 0 and 1 for each number in state C(t − 1). The value of f t can be found
using the equation below (Fig. 3).
  
f t  σ W f h t−1 , xt + b f (1)

[email protected]
A Semiautomated Question Paper Builder Using Long … 345

Fig. 3 Second step in LSTM

Fig. 4 Third step in LSTM

Second, we need to determine as to what is going to be stored in the successive cell.

First, the sigmoid layer decides as to which values to update. Finally, the tanh layer
vectorizes the values to form the Ct1 values. The next step involves the combination
of the two layers. i t and Ct1 can be found using the below equations (Figs. 4 and 5).
   
i t  σ Wi . h t−1 , xt + bi (2)
   
Ct1  tanh Wc . h t−1 , xt + bc (3)

In the next step, we update the old cell state C(t − 1) to new state C(t). In the
case of our model, this is where we drop the information of the old state and add new
information, as decided earlier. Here, Ct−1 is the previous old cell state.

[email protected]
346 R. Subramanian et al.

Fig. 5 Fourth step in LSTM

Ct1  f t ∗ Ct−1 + i t ∗ Ct (4)

Finally, we come up with the output that is the final information of a sentence. For
the model, it might want to output information relevant the outcome of the result.
The output ot can be found using the below equations.
   
ot  σ Wo . h t−1 , xt + bo (5)

h t  ot ∗ tanh (6)

3.3 Randomization Technique

3.3.1 Word Embeddings

Word Embeddings are used to represent two or more similar words with the same
meaning. For the purpose of training, a pretrained genism model on Wikipedia data
was used for converting words into vectors. Word Embeddings maps word to vectors
in a high-dimensional space. If learnt the right way, these word embeddings can learn
the semantic and the syntactic information of the words—i.e., similar words are close
together and dissimilar words are further apart. This can be learnt from large size
text files such as Wikipedia data. As an illustration, we have shown here the nearest
neighbors from the word embeddings space for some words. This embedding space
was learnt by the model that we propose through the following research. The example
case of the word embeddings used is as shown in Table 1.

[email protected]
A Semiautomated Question Paper Builder Using Long … 347

Table 1 Training example of word embeddings

Sunday Training Robot Artificial intelligence Time
Wednesday Education Machine Knowledge engineering Schedule
Monday Building Automaton Neural net Times
Sunday Tuition Artificial Intelligence Machine learning Schedules
Monday Tune up Automation NLP Dinnertime
Friday Practice Gadget Information retrieval Ord
Saturday Schooling Android Tokenization f28

3.3.2 Building the Classification Model

A many to one Long Short-Term memory network is used to classify a large number
of questions into three different categories as chapter number, marks allocated, and
Bloom’s Level of Taxonomy. However, the LSTM classifier can also be built for
attributes such as difficulty level. In this section, we go through how the classifier is
built for each of the attributes.
For the chapter classifier a many to one Long Short-Term Memory Network was
built such a way that the input layer is the sequence of words in question and the
target variable is the attribute we are trying to predict such as chapter number to which
the question may belong to. The model is trained for a large corpus of questions by
allocating the chapter numbers, respectively. A feedback to the input layer is provided
when there is an error in the classification and the weights are updated accordingly.
Similarly, the classifier was built for the marks and Bloom’s Level attributes
considering the range of marks for each question and the different levels of Bloom’s
taxonomy as the target variables.

3.3.3 Randomization Algorithm

The classified questions were appended to the data set in the form of a CSV file. The
CSV file was converted to a data frame for further processing to be carried out so
that each of the classified questions can be written to a word file.
The pseudo-code for the randomization technique is as follows:
Step 1: Initializations:

• A dictionary with two sequences of marks that add up to a total of 20.

• A list with all the chapter numbers.
• A dictionary for matching Bloom’s Level of Taxonomy.
• A visited array to keep track of all the chapters taken into consideration for each
unit.
• Two lists for the two-alternate question sets in a unit.

Step 2: Chose a random sequence number from the dictionary.

[email protected]
348 R. Subramanian et al.

Step 3: For every value in the sequence, find the marks and Bloom’s Level in the
sequence. Also, select a random chapter number from the union set of chapter list
and the visited array.
Step 4: Match the chapter selected and the marks selected with that in the data frame
consisting of the classified chapters and the marks using a search technique.
Step 5: Append the chapter selected to the visited array and append the chapter name,
marks and the Bloom’s Level to the list initialized for the first alternate part of the
unit.
Step 6: Initialize another list without the current selection of the random value chosen.
Repeat steps 1 through 5 for choosing the second alternate question in the unit.
Step 7: In a similar way, repeat the procedure for all the five units for which the
questions that must be set. Lastly, we append the lists to the word document to create
the final copy of the question paper.

4 Results

Using the Bidirectional Long Short-term Memory Networks, classifiers were built
for each of the attributes. The root mean square error method was used to find the
accuracy of the marks prediction model and the misclassification error for Bloom’s
Level and chapter number classification models.
The root mean square error is found by taking the root of the variance of the
predicted value of the classifier with that of the actual value of the classifier and the
misclassification error is calculated from the confusion matrix. The accuracy for the
models was found to be 82.4, 84.8, and 81.7%, respectively (Fig. 6).
The graph showing the comparison of the accuracies and errors of each model is
shown in the figure below. The inverse relation graph is also shown in Fig. 7.

Fig. 6 Comparison of Accuracy vs Model

models
100
Accuracy Error
80
Accuracy

60

40

20

0
Model 1 Model 2 Model 3
Model Name

[email protected]
A Semiautomated Question Paper Builder Using Long … 349

Fig. 7 Graph of accuracy

versus probability

References

1. Bhirangi, R., & Bhoir, S. Automated question paper generation system. Computer Engineering
Department, Ramrao Adik Institute of Technology, Navi Mumbai, Maharashtra, India.
2. Gangar, F. K., Gopal Gori, H., & Dalvi, A. Automatic question paper generator system. Depart-
ment of Information Technology, K. J. Somaiya College of Engineering Vidhyavihar, Mumbai-
77.
3. Cena, G., Dong, Y., Gao, W., Yu, L., See, S., Wang, Q., Yang, Y., & Jiang, H. A implementation
of an automatic examination paper generation system.
4. Mohandas, M., Chavan, A., Manjarekar, R., Karekar, D., Qing, L., & Byeong Man, K. (2003).
Automated question paper generator system. In IEEE/WIC International Conference on Clus-
tering Approach for Hybrid Recommender System, in Web Intelligence, 2003. WI 2003. Pro-
ceedings, (pp. 33–38).
5. Leekha, A., Barot, T., & Salunke, P. Bharati Vidyapeeth College of Engineering Sector-7, C.B.D.
Belpada, Navi Mumbai-400614, India.
6. Naik, K., Sule, S., Jadhav, S., & Pandey, S. Automatic question paper generation system using
randomization algorithm.
7. Divate, M., & Salgaonkar, A. Automatic question generation approaches and evaluation tech-
niques.

[email protected]
Time-Critical Transmission Protocols
in Wireless Sensor Networks: A Survey

Archana R. Raut, S. P. Khandait and Urmila Shrawankar

Abstract Wireless Sensor Network (WSN) is an extremely important tool for

closely monitoring, understanding and controlling application processes to the end
users. The main purpose behind installing the wireless sensor network is to make
real-time decisions based on data received from the sensor nodes. This data trans-
mission from sensor nodes to the base station is considered to be very complicated
because of the resources and communication capability constraints of various sensor
nodes and enormous amount of data is generated by WSNs. Real-time applications
in WSN like mission-critical monitoring, surveillance systems, etc., demands well-
timed and reliable delivery of data. For such applications, besides energy, Quality
of Services (QoS) routing, i.e., requirement of message delivery timeliness is also
one of the significant issues. Based on the type of application, it is essential to grant
different levels of QoS in WSNs. In this paper, QoS requirements for mission-critical
WSNs applications are highlighted and existing QoS-aware protocols to support such
applications are discussed with their boundaries in that domain.

Keywords Wireless sensor network · Quality of service · Mission-critical ·

Delay · Reliability

A. R. Raut (B)
Research Scholar, Department of Information Technology, G. H. Raisoni College of Engineering,
Nagpur 440016, India
e-mail: [email protected]
S. P. Khandait
Research Supervisor, Department of Information Technology, G. H. Raisoni College of
Engineering, Nagpur 440016, India
U. Shrawankar
Department of Information Technology, G. H. Raisoni College of Engineering, Nagpur 440016,
India

© Springer Nature Singapore Pte Ltd. 2019 351

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_30

[email protected]
352 A. R. Raut et al.

Sensor Network

Sensor Network Sensor Network

Gateway

Internet
Gateway Gateway

Client Client

Storage

Fig. 1 Wireless sensor network

1 Introduction

The most important capability of WSN is to sense the environmental parameters

and communicate it between sensors nodes deployed in the broad area of atten-
tion. Grouping and installing huge number of sensor nodes in the interested area
in an unplanned or planned manner can do this. These nodes correspond mutually,
wirelessly in self-organized way after installation and send collected data to base sta-
tion as shown in Fig. 1. WSN have numerous applications areas [1–4]. Some major
applications are in agriculture, in wildlife monitoring, in military surveillance, for
industry control applications [5, 6], environmental sensing, and personal health con-
cern as well as for home security and automation, etc. In health care applications,
wireless devices are used for patient monitoring and personal health care. In remote
monitoring, the necessary requirement is that wireless systems can be used in place
of wired systems to reduce cost of wiring and to permit various types of measure-
ments. Remote monitoring also includes Environmental monitoring of water, soil
and air monitoring, Industrial machine monitoring, Structural monitoring for vari-
ous constructions, object tracking and in military. In all these applications, reliable
and real-time monitoring is the necessary requirement.
Because of constraints like limited power, bandwidth, memory, etc., in sensor
networks, energy efficiency of nodes turns out to be an important parameter for
designing an efficient data collection schemes for sensor networks. Major part of
the energy of the sensors is required for the data transmission in sensor network. As
sensor nodes life in WSN is based on energy remained in the node so, reducing the
energy utilization enlarge the network lifetime. With the emergence of critical, mul-
timedia and real-time (RT) applications, QoS is becoming a significant parameter in

[email protected]
Time-Critical Transmission Protocols … 353

WSNs. So, in addition to conserving the overall network energy, QoS is also a key
measure to compute the overall network performance in RT delays sensitive WSN
applications [4–6]. As there is tradeoff between various application necessities such
as energy efficiency in addition to delaying performance, systems that grant real-
time services with guaranteed data delivery time, assured reliability and additional
required QoS-related parameters are essential in WSNs. Data transmission in such
critical applications demands both energy and QoS support for effective accessibility
of the collected measurements and efficient utilization of sensor resources. There-
fore, time-critical and reliable transmissions of collected data are two most important
parameters QoS provisioning in WSNs. In this paper, research is performed in the
direction of concentrating on the issues in existing data transmission schemes con-
sidering the reliability and delay in time-critical WSN applications.
The entire paper is organized as given below. Section 2 discusses different chal-
lenges in providing QoS in sensor networks. Some QoS-aware MAC and routing
protocols are reviewed in Sect. 3. A comparative revision on some QoS protocols
based on few important parameters in context of WSNs is also done. Section 4 dis-
cusses some issues that need to attend to enhance the performance of WSNs. Last, in
Sect. 5, concluding observations in the literature are presented with some research
directions in critical WSN applications.

2 Quality of Service

2.1 QoS Challenges in WSNs

In WSNs, different layers demand different QoS requirements to get better system
throughput [2–4]. Different application areas demand different kinds of QoS related
to various layers. To conserve RT applications the following are the main challenges
related to QoS in WSNs (Fig. 2). Amongst lots of WSNs applications, with a range
of QoS requirements, the most challenging application can be found in real-time
WSNs [4–6].
• Resource constraints: It mainly involves energy constraints, bandwidth constraints,
memory constraints, processing capability, and transmission power constraints.
• Node Deployment: Sensor nodes can be deployed randomly or in a pre-planned
manner.
• Topology changes: In WSN, topology rapidly changes in view of sensor node
mobility, link failures between nodes, faulty behavior of nodes or due to energy
reduction of nodes.
• Data Redundancy: Redundant data transmission in the network leads to network
congestion. To decrease the redundancy, we can make use of data aggregation
mechanisms [7].

[email protected]
354 A. R. Raut et al.

Resource
Constraints
Node
Real Time Deployment
Traffic
QoS
Challenges in
WSN
Multiple
Data Traffic
Redundancy Types
Topology
Changes

Fig. 2 QoS challenges in WSNs

Real-time traffic: In some critical applications like security inspection or natural

disaster monitoring, collected/sensed data should reach the target before its time limit.
So, such type of significant data must be handled by sufficient QoS mechanisms.
Critical applications like fire-monitoring, intruder-tracking, and medical care
are mostly used in real-time WSNs. All these applications include high needs of
bandwidth, delay guarantees and delivery time. When an application forwards data,
TCP/IP stack is used to transmit it from the source to the destination nodes. Every
layer of TCP/IP stack is associated with some QoS parameters. Figure 3 point up QoS
requirements for basic five layers of TCP/IP model such as application layer, trans-
port layer, network layer, medium access control (MAC) layer and physical layer. In
delay-protected RT applications in WSN, QoS guarantees can be considered as either
providing a deterministic time delay barrier or a probabilistic delay guarantee should
be provided. In deterministic approach, data arrived later than its deadline is of no use
or it can be considered as collapse of the system whereas in probabilistic approach,
some delay is acceptable. Hence, for supporting QoS in such applications WSNs,
either required probabilistic or deterministic time delay assurance. Also, based on the
research issues, QoS protocols are again classified as soft RT and hard RT protocols.
So, choosing the best QoS protocol to obtain better throughput based on the purpose
requirements is the most important concern mission-critical applications in WSN.

[email protected]
Time-Critical Transmission Protocols … 355

IP Stack QoS Requirements

Response Time
Detection Probability
Application Layer
Data Reliability
Data Resolution
Reliability
Bandwidth
Transport Layer Latency
Cost
Path Latency
Routing Maintenance
Network Layer Congestion Probability
Energy Efficiency

Communication Range
MAC Layer Throughput
Transmission Reliability
Energy Efficiency
Wireless unit capabilities
Physical Layer
Processor capabilities
Sensing unit capabilities

Fig. 3 QoS associated with different layers

End-To-End delay
Criticality
Event Driven Interactivity

Data End-To-End delay

Delivery Criticality
Models in Query Driven Query Specific Delay Tolerant
WSN Interactivity

Continuous Delay Tolerant

Criticality

Fig. 4 Data delivery models requirement for various application classes

2.2 Data Delivery Models in WSNs

Various data delivery models used in WSN that are utilized in special applications
are given in Fig. 4. Depending on QoS needs in different applications, three data
delivery model found in the literature are event driven, query-driven in addition to
continuous model [8, 9].
• Event-driven model: For applications like interactive, delay-intolerant, mission-
critical, event-driven model is used as these applications require detecting the

[email protected]
356 A. R. Raut et al.

events and as a result make the right decisions at the earliest. The accomplish-
ment of these applications based on the efficient event detection and delivery of
notification of that event to the end user.
• Query-driven model: This includes applications like query-specific delay lenient,
interactive and time-critical applications. Here queries can be forwarded as
required so as to minimize the energy usage. This model is alike to that of first
model. Only the difference is that in query driven the data is fetched by base station
whenever it is required whereas in the event-driven model, data is pressed to the
base station when desired event occurs. This data delivery model involves WSN
applications like environmental control or habitat monitoring.
• Continuous model: Sensor node sends the gathered data to the base station con-
tinuously at a particular time in the continuous model.
Applications requirements in all above three models are different. By making use
of these models we can include desired application-specific functionalities in WSNs.
The characteristics of the application requirements are depend on the following
factors shown in Fig. 4:
• Interactivity: Application can be either interactive or not.
• End-to-end: Application may need end-to-end management or not.
• Delay tolerance: Application may delay accepting or not.
• Criticality: Application may be time-critical or not.

3 Related Works

While expanding RT applications in WSN, resource constraints must be considered

in addition to reliability constraints to support desired performance at all time. In
recent years, many QoS-aware MAC and routing protocols were proposed for WSNs
[8–24]. Many of these protocol works for improving energy efficiency, considering
energy of sensor nodes as a significant resource. Besides, some additional challenges
also caused by time-critical applications where both energy proficient and QoS sup-
ported routing is required in order to utilize the sensors efficiently and accessing the
collected data successfully within time bounds. In QoS provisioning, MAC layer
should provide delay guarantee for getting access to channel whereas network layer
should provide the transmission time barrier. Figure 5 shows the classification of
various protocols in view of probabilistic or deterministic approaches.
In WSNs, main role of the MAC layer is to determine delay required for get-
ting access to the channel, channel utilization, and energy requirements. Data trans-
port, reliability, and delay are fundamental performance objectives in mission-critical
application scenarios [8]. Delay-aware MAC protocols like Sensor-MAC (S-MAC),
Timeout-MAC (T-MAC) Berkeley-MAC (B-MAC) [9–11], etc., only reduce delay
to offer best-effort services but RT assurance is not granted. An energy efficient and
low latency MAC protocol (D-MAC) [12] and a Delay-Bounded MAC Protocol (DB-
MAC) [13] are proposed for application-specific data gathering tree but are bounded

[email protected]
Time-Critical Transmission Protocols … 357

QoS Protocols for WSN

MAC Layer Protocols Network Layer Protocols

(Single –hop) (Multi-hop)

Probabilistic Deterministic Probabilistic Deterministic

S-MAC Dual Mode MAC SPEED Ergen

T-MAC PEDAMACS MMSPEED
B-MAC IEEE 802.15.4 Dwarf
D-MAC Gin-MAC WirelessHART
DBMAC RPAR
QoSMAC DEGRAM
BCMN/A
EA-QoS
QEMPAR

Fig. 5 Classification of QoS protocols for WSN

to use in general topology and even if latency is reduced but precise RT guarantee is
not offered by these protocols. In 2006, TDMA-based MAC protocol PEDAMACS
[14] is proposed by Ergen, Varaiya to accomplish energy and delay effectiveness. It
provides hard RT guarantee but due to limitation of requirement of powerful AP it
cannot be used in variety of applications. IEEE 802.15.4 [15] protocol is developed
with guaranteed time slot (GTS) mechanism for managing time significant data to
provide explicit QoS guarantees and is still developing to support hard RT guarantee.
Even though, MAC layer provides RT guarantee, still without incorporating trans-
mission delay bound in the network layer packet deadline is hard to achieve. There-
fore, in real-time applications, routing protocols capable of providing delay assurance
are preferential. A Stateless Protocol for Real-Time Communication (SPEED) [16]
is a RT routing protocol proposed in 2003 to achieve flexible end-to-end time limit
guarantee by supporting a preferred speed in the network. A Multi-Path and Multi-
SPEED (MMSPEED) [17] protocol is an improved SPEED, which provides service
discrimination and data transmission delay guarantee in addition to data reliability.
By providing high link reliability MMSPEED meets both timeliness and reliabil-
ity requirements over SPEED. However, both in SPEED and MMSPEED energy
expenditure of the network was not considered. Some other protocols like Quality of
Service MAC (QoSMAC) [18] ensure reliability and transmission delay from source
to destination node. Priority-MAC protocol [25] was designed for handling time-
critical data transmission for industrialized wireless sensors and actuators network
by categorizing the traffic in four different classes depending on priorities consider-
ing 1 for highest and 4 for lowest priority and traffic in each of the class utilizes its

[email protected]
358 A. R. Raut et al.

committed scheme for medium access. Furthermore, In Dwarf [19] message delay
is reduced with increased reliability, whereas in WirelessHART [20], a wireless sen-
sor networking technology based on the Highway Addressable Remote Transducer
Protocol only delay from source to destination is promised without guaranteed reli-
ability. In further literature, a novel TDMA-based MAC protocol called Gin-MAC
[21] guarantees time delay as well as data delivery reliability in addition to efficient
energy utilization.
In 2009, author proposed a routing protocol [22] for WSNs, which enhances
the quality of service in real-time data delivery. This protocol improves throughput
of the system by reducing the packet deadline miss ratio and efficiency of energy
consumption by using energy balancing approach. In [23], author proposes a gateway
control protocol (H-NAMe) which uses grouping approach to avoid interference
between overlapping clusters and to eliminate collisions of hidden nodes in a single or
several clusters WSNs by using contention-based MAC protocols. Data aggregation
cross-layered design protocol Lump is proposed in 2010, which maintain QoS in
various applications by giving priorities to packets for distinguished services and
helps aggregation decisions [24]. It reduces overheads in network processing, and
thus suitable in many applications.
In 2011, author proposed energy efficient QoS routing protocol (EEQAR) [26]
to manage both the energy efficiency and QoS the requirement. EEQAR protocol
uses cellular topology scheme for clustering with reasonable energy consumption to
gain the enhanced performance. Also, QoS and network lifetime is highly improved
using EEQAR for multimedia sensor networks. Similar kind of protocol QoS and
Energy-Aware Multi-Path Routing (QEMPAR) [27] is proposed in the same era for
providing QoS requirements in RT applications in WSNs. As the protocol utilizes
four main parameters of QoS, end-to-end delay is optimized. Here author also high-
lighted some architectural and other operational challenges for managing the QoS
data transmission in sensor networks in addition to enhance energy efficiency of the
sensors. Single-hop hybrid cluster architecture is proposed in 2013, which includes
two kinds of nodes transmit-only nodes as well as standard nodes [28]. Here author
also proposed a framework for MAC layer protocol called RARE with the purpose of
managing the one hop hybrid cluster efficiently and reliably in a self-structured fash-
ion in densely deployed area. This work brings reliable scheduling scheme using
transmit-only nodes as well as standard nodes. In direction to reduce the sensor
nodes required to perform various tasks, energy unconscious protocols may effect
in unequal expenditure of sensor energy by passing on irregular workloads to the
sensors. In this, heavily loaded sensors may create energy holes, which may result
in partitioning the network into incoherent parts. To overcome this trouble and to
enhance the network lifetime, in [29] author proposed two energy aware task manage-
ment protocols: wherein combination centralized and distributed protocols is used for
improving network lifetime using load balancing technique amongst sensor nodes.
Two stages are used in distributed approach to guess the maximum energy altogether
between them. These protocols improve the network lifetime by utilizing the sensor
nodes energy on event basis as well as controlling energy gaps between the sensor
nodes exist in identical region. Balancing the sensor nodes energy utilization with no

[email protected]
Time-Critical Transmission Protocols … 359

change in effective density of the sensor network is the key benefit of this approach.
In [30], Joint Optimization of Lifetime and Transport Delay under Reliability con-
straint (BCMN/A) protocol also improves network lifetime and reduces network
delay by using energy as well as delay efficient data aggregation process both in
intra- as well as inter-cluster communication. A TDMA-based protocol (DGRAM)
that gives significant delay guarantee with desired energy efficiency is proposed in
[31]. This protocol is completely self-configuring protocol in which assignment of
slots is done with no control messages exchanging. As it does not provide significant
delay guarantee, it is not useful in a wide range of applications.
Thus, here some existing MAC and Network layer protocols in Wireless Sensor
Networks designed with intension of providing reliable and timeliness data trans-
mission are discussed. In addition, some cross-layer solutions invented for offering
QoS in WSNs are also discussed. We elaborated on some of the existing approaches
with special focus on the techniques used to boost the network performance of delay-
bounded applications considering reliability, delay, energy efficiency, type of service
as specific QoS parameter in order to find the technical issues for providing energy
efficient, QoS guarantee in WSN. Table 1 summarizes some QoS features provided
by various existing data transmission techniques in WSNs.

4 Research Directions

Efficient collection of data from various sensors installed in the field and transmission
of the accumulated data to the respective base station is the important responsibil-
ities of the sensor nodes. The major challenge here is to conserve the sensor node
energy to maximize their existence, which leads to enhance the network lifetime. It
is a most important and considerable factor in efficient data collecting schemes in
favor of wireless sensor networks. In addition, QoS is the most important issue in
time-critical applications in WSNs. Reliability along with timeliness is the signifi-
cant parameter to be concern in providing QoS in such applications [32–34]. Data
redundancy concept can be used to accomplish Reliability. But, it will affect timeli-
ness factor as it introduces delay in transmission. So, optimum techniques should be
launched which will take care of both parameters reliability and timeliness in QoS
provisioning. Presently, only some protocols deal with the dual purpose of attaining
delay and reliability bounds metrics, but only some of the proposed protocols can
preserve time-critical applications. Hence, it is essential to develop event based reli-
able data transmission methods to support time-critical applications within sensor
network domain which gives optimal performance considering QoS guarantee in
terms of delay bound and reliability. So, while designing a QoS scheme, researchers
must focus on event-based reliable transmission techniques to guarantee the delivery
of crucial data in timely as well as reliable manner which gives optimal performance
considering delay and reliability bearing in mind the growing needs of various time-
critical application domains.

[email protected]
360 A. R. Raut et al.

Table 1 Comparison of various QoS protocols in WSN

Protocol name Type Reliability Delay Energy RT type Issues
S-MAC, CSMA/CA- Not Decrease Duty Best effort No
T-MAC, based assured node-to- cycling real-time
B-MAC MAC node data
delay delivery
guarantee
D-MAC, CSMA/CA- Not Decrease Duty Best effort No delay
DB-MAC based assured overall cycling guarantee
MAC end-to-
end
delay
QoSMAC TDMA- Yes (Node Yes (Node Duty Soft RT No end-
based level) level) cycling to-end
MAC + delay
Routing guarantee
Dual Mode TDMA- Yes (End Yes (End Not con- Hard RT Only
MAC based level) level) sidered applicable
MAC when load
is low
PEDAMACS TDMA- Not Yes (End Duty Hard RT Requires
based assured level) cycling high
MAC power AP,
low
scalability
IEEE802.15.4 Slotted Yes (End Yes (End Moderate Best Need to
CSMA/CA level) level) effort/Hard improve
RT reliability
Gin-MAC TDMA- Yes (End Yes (End Duty Hard RT Low
based level) level) cycling scalability
MAC
Priority-MAC MAC Increases Decreases Yes Soft RT No
(End (End collision
level) level) avoidance
SPEED MAC + Soft (End Soft (End No Soft RT Energy
Routing level) level) metric not
consid-
ered
MMSPEED MAC + Probabilistic Probabilistic No Soft RT Energy
Routing (End (End metric not
level) level) consid-
ered
Dwarf MAC + Increase Decrease Duty Soft RT More
Routing (End (End cycling focus on
level) level) reliability
than delay
parameter
(continued)

[email protected]
Time-Critical Transmission Protocols … 361

Table 1 (continued)
Protocol name Type Reliability Delay Energy RT type Issues
WirelessHART TDMA/ Increase Guarantees Duty Soft RT More
FDMA (End (End cycling communi-
MAC + level) level) cation
Routing overhead
SchedEx TDMA- Guarantees Decreases Duty Soft RT Guaranteeing
based (End (End cycling reliability
MAC level) level) bounds
effects
latency
bounds
RPAR Routing Increase Decreases Yes Soft RT No end-
(End (End to-end
level) level) delay &
reliability
guarantee
DEGRAM TDMA- No Guarantees Yes Soft RT No end to
based (End end
MAC + level) reliability
Routing guarantee
BCMN/A Routing Overall Overall Yes Soft RT Channel
Network Network access
increase delay delay not
decreases consid-
ered
EA-QoS Routing Moderate Guarantees Yes Soft RT Channel
increase (End access
level) delay not
consid-
ered
QEMPAR Routing No Increase Yes Soft RT Throughput
in latency affect due
to
increased
latency
Ergen Routing No Guarantees Yes Hard RT Channel
(End access
level) delay not
consid-
ered

[email protected]
362 A. R. Raut et al.

5 Conclusion

The time-critical application in WSN poses several challenges in concern with timely
and reliable delivery of data collected at sensor nodes to the base station. This paper
discusses the existing QoS provisioning data transmission techniques in WSNs with
different requirements and necessity for different approaches. This paper has dis-
cussed the requirements; challenges for sustaining QoS in WSNs. Current develop-
ments discussed in the paper indicate that there are still many issues and challenges
that need to attend. Considering the various application domains, in addition to energy
QoS is also the main concern for successful implementation of WSNs. So, for assured
and reliable event awareness in the network is the major requirement of many delay
constraint and mission-critical WSN applications. To address this problem, it is nec-
essary to discover most efficient solution in WSNs scheme, which gives reliability,
and timeliness in delay sensitive applications, in WSNs. Thus, it looked for research
overcome both end-to-end level reliability and delay limitations in real-time systems
in WSNs.

References

1. Al-Karaki, J. N., & Kamal, A. E. (2004). Routing techniques in wireless sensor networks: A
survey. IEEE Wireless Communications, 11, 6–28.
2. Li, Y., Chen, C. S., Song, Y.-Q., & Wang, Z. (2007). Real-time QoS support in wireless sen-
sor networks: A survey. In 7th IFAC International Conference on Fieldbuses & Networks in
Industrial & Embedded Systems - FeT’2007, Nov 2007, Toulouse, France.
3. Arampatzis, T., Lygeros, J., & Manesis, S. (2005). A survey of applications of wireless sensors
and wireless sensor networks. In Proceedings of the 20th IEEE International Symposium on
Intelligent Control (ISIC 05) (pp. 719–724), June 2005.
4. Chen, D., & Varshney, P. K. (2004). QoS support in wireless sensor network: A survey. In Pro-
ceedings of International Conference on Wireless Networks (ICWN2004), Las Vegas, Nevada,
USA, June 2004.
5. Balen, J., Zagar, D, & Martinovic, G. Quality of service in wireless sensor networks: a survey
and related patents.
6. Raut, A. R., & Malik, L. G. (2011, May). ZigBee based industrial automation profile for power
monitoring systems. International Journal on Computer Science and Engineering (IJCSE),
3(5), 2028–2033 ISSN: 0975-3397.
7. Raut, A. R., & Malik, L. (2011). ZigBee: The emerging technology in building automation.
International Journal on Computer Science and Engineering, 3(4), 1479–1484.
8. Suriyachai, P., Roedig, U., & Scott, A. (2012, Second Quarter). A survey of MAC protocols
for mission-critical applications in wireless sensor networks. IEEE Communications Surveys
& Tutorials, 14(2).
9. Ye, W., Heidemann, J., & Estrin, D. (2002). An energy-efficient MAC protocol for wireless
sensor networks. In Proceedings of the 21st Annual Joint Conference IEEE Computer and
Communications Societies, New York, NY, USA (Vol. 3, pp. 1567–1576).
10. Van Dam, T., & Langendoen, K. (2003). An adaptive energy-efficient MAC protocol for wire-
less sensor networks. In Proceedings of the 1st ACM Conference on Embedded Networked
Sensor Systems, Los Angeles, CA, USA, 2003 (pp. 171–180).
11. Polastre, J., Hill, J., & Culler, D. (2004). Versatile low power media access for wireless sensor
networks. In: Proceedings of ACM Sensys (pp. 95–107).

[email protected]
Time-Critical Transmission Protocols … 363

12. Lu, G., Krishnamachari, B., & Raghavendra, C. S. (2004). An adaptive energy efficient and
low-latency MAC for data gathering in wireless sensor networks. In Proceedings of the 18th
International Parallel and Distributed Processing Symposium, Santa Fe, NM, USA, 2004
(pp. 224–231).
13. Bacco, G. D., Melodia, T., & Cuomo, F. (2004). A MAC protocol for delay-bounded applica-
tions in wireless sensor networks. In Proceedings of the Med Hoc-Networks 2004 (pp. 208–220).
14. Ergen, S. C., & Varaiya, P. (2006). PEDAMACS: power efficient and delay aware medium
access protocol for sensor networks. IEEE Transactions on Mobile Computing, 5, 920–930.
15. Yoo, S., Chong, P. K., Doh, Y., Pham, M.-L., Kim, D., Choi, E., & Huh, J. (2010, November).
Guaranteeing real-time services for industrial wireless sensor networks with IEEE 802.15.4.
IEEE Transactions On Industrial Electronics, 57(11), 3868–3876.
16. Hea, T., Stankovica, J. A., Lub, C., & Abdelzahera, T. SPEED: A stateless protocol for real-time
communication in sensor networks. In: Proceedings of the ICDCS (pp. 46–55).
17. Felemban, E., Lee, C., & Ekici, E. (2006). MMSPEED: Multipath multi SPEED protocol for
QoS guarantee of reliability and timeliness in wireless sensor networks. IEEE Transactions on
Mobile Computing, 5, 738–754.
18. Suriyachai, P., Roedig, U., & Scott, A. (2009). Implementation of a MAC protocol for QoS
support in wireless sensor networks. In Proceedings of the 1st International Workshop Informa-
tion Quality and Quality of Service for Pervasive Computing, in conjunction with 7th Annual
IEEE International Conference on Pervasive Computing and Communications, Galveston, TX,
USA, 2009 (pp. 1–6).
19. Strasser, M., Meier, A., Langendoen, K., & Blum, P. (2007). Dwarf: Delay-aware robust for-
warding for energy-constrained wireless sensor networks. In Proceedings of the 3rd IEEE
International Conference on Distributed Computing in Sensor Systems, Santa Fe, NM, USA,
2007 (pp. 64–81).
20. HART Communication Foundation. WirelessHART technology. [Online]. Available: http://
www.hartcomm.org/protocol/wihart/wireless_technology.html, December 2009.
21. Suriyachai, P., Brown, J., & Roedig, U. (2010). Time-critical data delivery in wireless sensor
networks. In Proceedings of the 6th IEEE International Conference on Distributed Computing
in Sensor Systems, Santa Barbara, CA, USA, 2010 (pp. 216–229).
22. Li, Y., & Chen, C. S. (2009, May). Enhancing real-time delivery in wireless sensor networks
with two-hop information. IEEE Transactions On Industrial Informatics, 5(2), 113–122.
23. Koubâa, A., Severino, R., Alves, M., & Tovar, E. (2009, August). Improving quality-of-service
in wireless sensor networks by mitigating “hidden-node collisions”. IEEE Transactions on
Industrial Informatics, 5(3), 299–313.
24. Jeong, J., & Kim, J. (2010). A QoS-aware data aggregation in wireless sensor networks. In
12th International Conference on Advanced Communication Technology (ICACT), February
7–10, 2010.
25. Subramanian, A. K., & Paramasivam, I. (2016). PRIN: A priority-based energy efficient MAC
protocol for wireless sensor networks varying the sample inter-arrival time. https://doi.org/10.
1007/s11277-016-3581-5 (Springer Science Business Media, New York).
26. Lin, K., & Rodrigues, J. J. P. C. (2011, December). Energy efficiency QoS assurance routing
in wireless multimedia sensor networks. IEEE Systems Journal, 5(4).
27. Heikalabad, S., Rasouli, H., Nematy, F., & Rahmani, N. (2011, January). QEMPAR: QoS
and energy aware multi-path routing algorithm for real-time applications in wireless sensor
networks. IJCSI International Journal of Computer Science Issues, 8(1).
28. Zhao, J., Qiao, C., & Sudhaakar, R. S., & Yoon, S. (2013, March). Improve efficiency and
reliability in single-hop WSNs with transmit-only nodes. IEEE Transactions on Parallel and
Distributed Systems, 24(3).
29. AbdelSalam, H. S., & Olariu, S. (2011, November). Toward efficient task management in
wireless sensor networks. IEEE Transactions on Computers, 60(11).
30. Dong, M., & Ota, K. Joint optimization of lifetime and transport delay under reliability con-
straint wireless sensor networks. IEEE Transactions on Parallel and Distributed Systems,
TPDS-2013-12-1250.

[email protected]
364 A. R. Raut et al.

31. Shanti, C., & Sahoo, A. (2010, October). DGRAM: A delay guaranteed routing and AC protocol
for wireless sensor networks. IEEE Transactions on Mobile Computing, 9(10).
32. Pöttner, W.-B., & Seidel, H. (2016, January). Constructing schedules for time-critical data
delivery in wireless sensor networks. ACM Transactions on Sensor Networks, V (N), Article A.
33. Mohammad, B. M., Abd Kader, S., & Konber, H. A. Designing a channel access mechanism
for wireless sensor network. Wireless Communications and Mobile Computing, 2017, Article
ID 7493269, 31 p. https://doi.org/10.1155/2017/7493269.
34. Richert, V., & NaumanIsrar, B. Implementation of a modified wireless sensor network MAC
protocol for critical environments. Wireless Communications and Mobile Computing, 2017,
Article ID 2801204, 23 p. https://doi.org/10.1155/2017/2801204.

[email protected]
Impact of Shuffler Design Pattern
on Software Quality

G. Priyalakshmi, R. Nadarajan, Joseph W. Yoder, S. Arthi and G. Jayashree

Abstract Modern software has become intricate and versatile due to the world-
wide growth of new software technologies. In this regard, the evolution of soft-
ware quality metrics to support software maintainability is studied. The impact of
the Shuffler design pattern on software quality metrics is evaluated in this paper.
The Shuffler design pattern provides an efficient design approach for shuffling. The
pattern helps to choose generic shuffling alternatives that make the client program
loosely coupled, and thus attaining high reusability. A few software quality met-
rics, which has a higher influence on software reusability and maintainability, are
experimented on three gaming applications like Jigsaw, Poker, and Scramble. These
gaming applications are redesigned using Shuffler design pattern and a combination
of other patterns. The three software quality metrics, which show improvement on
the redesigned applications, are McCabe Cyclomatic Complexity, Lack of Cohesion
of Methods and Specialization Index. The authors also have tested the pattern with
a reusability metrics suite, which measures the reusability of the black-box compo-
nents of the aforementioned applications without any source code. The results with
high cohesive and low coupling values would help software designers in the industry
to be more confident in using the Shuffler pattern along with other design patterns.
The interdependence of the three software metrics on the software quality attributes
is finally tabulated to show their impact on software quality.

1 Introduction

Design patterns represent solutions that have emerged and matured over time. Pat-
terns solve particular design problems and make object-oriented designs more flexi-
ble, refined, and ultimately reusable. They assist designers to reuse successful designs

G. Priyalakshmi (B) · R. Nadarajan · S. Arthi · G. Jayashree

Department of Applied Mathematics and Computational Sciences, PSG College of Technology,
Coimbatore, India
e-mail: [email protected]
J. W. Yoder
The Refactory, Inc, Urbana, USA

© Springer Nature Singapore Pte Ltd. 2019 365

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_31

[email protected]
366 G. Priyalakshmi et al.

by depending upon new designs on prior experience. A software designer who is

acquainted with such patterns can put them into effect to design problems directly
without ascertaining them [1]. According to ISO 9126, software must be portable,
completely functional, reliable, usable, efficient, and maintainable. Each quality sub-
attribute (low-level quality attributes, such as complexity, cohesion, etc.), can be
assessed by a set of metrics that can be used as indicators for the score of a system
with respect to the corresponding high-level quality attribute.
Apostolos, Sofia, and Ioannis have presented an extensive survey of 120 papers
to impart an overview of GOF pattern research [2]. They have accounted that, the
most prominent research field is the impact of patterns on software quality attributes.
Finally, they have formally stated that the reported research on the effect of patterns on
software quality attributes is controversial, since some studies identify one pattern’s
effect as beneficial whereas others report the same pattern’s effect as harmful. Hence,
this work attempts to narrow these gaps by investigating the effect of applying patterns
on software quality attributes and providing an empirical study.
Shuffler design pattern creates a design technique from existing solutions that
make them available to software developers [3]. Shuffle is a method that refers to
jumbling or interchanging the positions of objects. Online games like Jigsaw puzzle,
Sudoku, Poker, Jumbled word games, etc., more commonly use the shuffling design
solution. In music applications, shuffling refers to the capability to randomize the
order of the songs in a playlist. Thus, shuffling has an immense number of well-known
uses in the software industry. The other real-time practical applications are shuffling
questions and choices in quiz apps or online tests, image shuffling in games, shuffling
wallpapers in desktops or mobiles, shuffling characters in password generators, etc.
The research work proposed by the authors is to measure quantitatively the impact
of Shuffler design pattern on open source gaming applications. An Eclipse plug-in
called Metrics calculates 22 metrics for these applications. Out of 22 metrics, three
of them showed a significant improvement, when the Shuffler design pattern was
applied. Table 3 provides a brief description of a few of the metrics. In addition, few
coupling metrics listed in Table 4, were used to study the impact of Shuffler pattern
on the software quality of the above mentioned open source systems. A comparative
analysis of the open source applications for each of the significant metrics yields
better results.
To emphasize the proposed work, the authors also consider reusability met-
rics for software components. The metrics studied were EMI (Existence of Meta-
Information), RCO (Rate of Component Observability), RCC (Rate of Component
Customizability), SCCr (Self-Completeness of Component’s Return Value), SCCp
(Self-Completeness of Component’s Parameter), for computing the existence of
meta-information, observability, customizability, and external dependency of soft-
ware components without source code [4]. These metrics were applied to the applica-
tions Poker, Jigsaw, and Scramble. The Poker gaming application after redesigning
using Shuffler design pattern had greater flexibility in the metrics EMI and SCCp.
When applied to Scramble game, it showed similar values for all the metrics except
EMI, indicating minimal impact of Shuffler design pattern on this application. When
the Jigsaw system was measured with the same metrics, except for the RCC and

[email protected]
Impact of Shuffler Design Pattern on Software Quality 367

SCCp, all other metrics showed improved values when compared to the same mod-
ule designed without Shuffler design pattern.
A decision table, showing the interdependence between software quality mea-
sures and object-oriented metrics, coupling metrics, and reusability metrics for com-
ponents, was reconstructed. This could be very effective for software practitioners to
interpret the direct impact of design patterns on software quality. Finally, the study
was extended with the impact of Shuffler design pattern coupled with Singleton [5]
and a similar analysis was done on Poker and Scramble projects. The effect of Shuf-
fler and Prototype coupled was measured on the Scramble project, and the results
were tabulated.

2 Related Work

There has been a booming involvement in the adoption of design patterns, since
their initiation in 1995 by Gamma et al. [1]. In this section, some lines of work
which performs a study of the impact of design patterns on quality attributes are
presented. Lange and Nakamura proved [6] that patterns help in inspection of a
software program thus increasing its understandability. However, this study was
limited to a single quality attribute and to a few subset of patterns. Foutse and Yann
hypothesized [7] that though object-oriented best practices help producing systems
with good quality; it is surprising that their adoption seems to decrease quality. The
negative evaluations may be just an a priori on the pattern because their respondents
considered the pattern not suitable. Also, their evaluation may not reflect the real
impact of the implementation of the pattern on quality. In addition, they inferred
that the consequences for some design patterns cannot be justified, thus calling for
further studies on the impact of these principles on quality.
Bansiya and Davis [8] evaluated high-level quality features in designs of object-
oriented software systems. The hierarchical QMOOD model, which the authors pro-
posed, relates various characteristics like cohesion, coupling, encapsulation, modu-
larity, etc. to high-level quality attributes using experimental and informal informa-
tion. The authors have examined qualities like understandability, reusability, flexibil-
ity, etc. The model also provides a unique feature to include different relationships
and weights. Hence, it delivers a practical quality assessment tool compliant to a
variety of needs. Wendorff [9] warned developers of commercial products that the
magnitude of problems in maintenance phase is proportional to the unconstrained
use of patterns. The intention of this paper is not to condemn the idea of patterns;
instead, it is motivated by the remark that the improper application of patterns can
possibly go wrong. The paper identified two categories of inappropriately applied
patterns: first, patterns that were simply misused by software developers who had not
understood the rationale behind the patterns; second, patterns that do not match the
project’s requirements. There are many more studies which include deep learning
and theoretical evaluation of the design patterns. This led to the hypothesis of vari-

[email protected]
368 G. Priyalakshmi et al.

Table 1 Survey on existing work with their limitations

References Authors Content Limitations
[1] Erich Gamma, Richard Introduction to
Helm, Ralph Johnson, object-oriented design.
John Vlissides Applications,
advantages,
disadvantages of
design patterns
[7] Foutse Khomh, Object-oriented The outcome of their
Yann-Gael Gueheneuc practices help research could not be
producing systems with justified for some
good quality design patterns
[6] D. B. Lange, Y. Demonstration of Limited to a single
Nakamura design patterns that quality attribute.
serve as a guide in Limited to certain
program exploration patterns
[8] Jagdish Bansiya, Carl Proposed a hierarchical
G. Davis QMOOD model to
assess design with
abstract quality
characteristics
[9] Peter Wendorff In many commercial
softwares, the increase
in maintenance issues
is a result of
uncontrolled use of
patterns
[14] Ronald Jabangwe, Analysis of the Can expand types of
Jürgen Börstler, Darja association between external attributes that
Šmite, Claes Wohlin design measures and were explored
quality attributes not
related to source code
was done and the
stability of the
relationship was
evaluated across other
studies
[2] Apostolos A survey of existing Investigation of the
Ampatzoglou, Sofia work in the field of impact of patterns on
Charalampidou, GoF patterns is done by design quality using
Ioannis Stamelos a mapping review of other empirical
120 basic researches methods or theoretical
methods

ous models and theories for assessing the design patterns and analyzing their impact.
Table 1 shows a consolidated list of few studies and their limitations.
Object-oriented environments are evaluated using various metrics. These metrics
are developed as user-friendly tools in various programming languages like Object

[email protected]
Impact of Shuffler Design Pattern on Software Quality 369

Pascal, Java and C++. A survey of few of these tools and the metrics evaluated by
them are provided in Table 2.

3 Object-Oriented Metrics

The object-oriented metrics are used to measure, foresee and enhance the quality
of software products. There are various metrics that are used in different phases
of project management, but this paper focuses on object-oriented design metrics.
Object-oriented design metrics are classified by Archer into coupling level, inheri-
tance level, class level, and method level. This study focuses on the object-oriented
metrics tabulated in Table 3.
McCabe Cyclomatic Complexity counts the number of flows through a piece of
code. Each time a branch occurs (if, for, while, do, case, catch, ?: ternary operator,
as well as the && and || conditional logic operators in expressions) this metric is
incremented by one. It is calculated for methods only. If a method has more than 10
different loops, break the method to reduce McCabe Cyclomatic Complexity. For a
project, McCabe Cyclomatic Complexity must be less. The complexity M is defined
as

M  E − N + 2P (1)

where
E the number of edges in the graph.
N the number of nodes in the graph.
P the number of connected components
Lack of Cohesion of methods (LCOM) is a measure for the cohesiveness of a class.
It is calculated using Henderson-Sellers method. If m(A) is the number of functions
retrieving an attribute A, compute the average of m(A) for all attributes From the
mean, subtract the number of functions, m and divide the resultant expression by
(1 − m). A small value signifies a cohesive class and a value almost nearer to 1
implies a lack of cohesion. A high LCOM value means low cohesion [10]. A class
with high cohesion might show a better design if split into a number of subclasses
[11]. Take each pair of methods in the class. If they access disjoint sets of instance
variables, increase P by one. If they share at least one variable access, increase Q by
one. LCOM is computed as

LCOM  P−Q (if P > Q) (2)

LCOM  0 (otherwise) (3)

[email protected]
370 G. Priyalakshmi et al.

Table 2 Software tools with the metrics they measure

S. No. Tools Metrics measured
1 Eclipse metrics plug-in [15] Cyclomatic complexity
Coupling
Specialization index
Depth of inheritance tree
Abstractness
Instability
Normalized distance
2 Analyst4j [16] Cyclomatic complexity
Halstead Effort
3 CCCC [17] McCabe’s Complexity
Chidamber & Kemerer metrics
Henry & Kafura metrics
4 Chidamber & Kemerer Java Afferent coupling
metrics [18] Number of public methods
Depth of inheritance tree
Response for a class
5 Dependency Finder [19] Static classes per group
Single lines of code
Inner classes per group
6 Testwell CMT Java [20] McCabe’s Cyclomatic number
Depth of control structure nesting
Maintainability index
Halstead metrics
7 Resource standard metrics Cyclomatic complexity
[21] Overall complexity
LOC/SLOC
8 CodePro AnalytiX [22] Comments ratio
Efferent coupling
Block depth
Cyclomatic complexity
9 Java Source Code Metrics Cyclomatic complexity
[23] Depth of Inheritance tree
10 JDepend [24] Coupling
Abstractness
Instability
(continued)

[email protected]
Impact of Shuffler Design Pattern on Software Quality 371

Table 2 (continued)
S. No. Tools Metrics measured
Package dependency cycle
11 JHawk [25] Average cyclomatic complexity
Maintainability index
Cumulative Halstead effort
Cumulative Halstead Bugs
12 jMetra [26] Lines of code
Method tallies
13 JMetric [27] OO metrics
14 Krakatau Metrics [28] Line counting metrics
Complexity metrics
Halstead metrics
15 Refactorit [29] Weighted method per class
Lines of code
Depth of Inheritance tree
16 SonarJ [30] Cyclic dependencies
17 OOMeter [31] Ratio of cohesive inheritance
Coupling between objects
Tight class cohesion
Loose class cohesion
18 SemmleCode [32] Number of lines of code
19 Understand [33] Average complexity
Average line code
Average line blank
20 VizzAnalyzer [34] Class complexity

LCOM  0 indicates a cohesive class. LCOM > 0 indicates that the class needs
or can be split into two or more classes since its variables belong to disjoint sets.
Classes with a high LCOM are fault-prone.
The specialization index is defined as in Eq. 4. This is a class level metric.

Specialization Index  (NORM ∗ DIT)/NOM (4)

To maintain high-quality software, developers need to aim for a low-coupled

and highly cohesive design. The coupling and cohesion metrics are computed by
dealing with a number of associations proposed by various authors doing research.
Coupling metrics are influenced by acyclic dependencies, stable dependencies, and
stable abstractions. The authors have also considered the impact of Shuffler pattern

[email protected]
372 G. Priyalakshmi et al.

Table 3 Object-oriented metrics with their interpretations

Object-oriented metrics Interpretation
Number of children Total number of direct subclasses of a class
Number of interfaces Total number of interfaces in the selected scope
Depth of Inheritance Tree (DIT) Distance from class Object in the inheritance
hierarchy
Number of Overridden Methods (NORM) Total number of methods in the selected scope
that are overridden from an ancestor class
Number of Methods (NOM) Total number of methods defined in the selected
scope
Number of fields Total number of fields defined in the selected
scope
Lines of code TLOC: Total lines of code that will count
non-blank and non-comment lines in a
compilation unit
MLOC: Method lines of code will count and
sum non-blank and non-comment lines inside
method bodies
Specialization index It is defined as NORM * DIT/NOM. This is a
class level metric
McCabe Cyclomatic Complexity Counts the number of flows through a piece of
code. Calculated for methods only
Weighted Methods per Class (WMC) Sum of the McCabe Cyclomatic Complexity for
all methods in a class
Lack of Cohesion of Methods (LCOM) A measure for the cohesiveness of a class

Table 4 Coupling metrics with their meaning

Coupling metrics Interpretation
Afferent coupling (C a ) The number of classes outside a package
that depends on classes inside the package
Efferent coupling (C e ) The number of classes inside a package that
depends on classes outside the package
Instability (I) C e /(C a + C e )
Abstractness (A) The number of abstract classes and
interfaces, divided by the total number of
types in a package
Normalized distance from main sequence (Dn ) | A + I – 1 |, this number should be small,
close to zero for good packaging design

on coupling metrics. The coupling metrics which are used in this work are elaborated
in Table 4.

[email protected]
Impact of Shuffler Design Pattern on Software Quality 373

4 Reusability Metrics for Components

Reusability is one of the key features of well-designed object-oriented software. In

order to evaluate the software reusability, Component-based Development (CBD)
is carried out. CBD is a software development method to state, execute and consti-
tute loosely coupled individualistic components into software. CBD is characterized
by two activities: development of components for reuse, and development of soft-
ware systems with reuse of components [12]. As discussed, the various reusability
metrics include EMI, RCO, RCC, SCCr, and SCCp. For each of the gaming appli-
cations redesigned using the Shuffler design pattern, the components are evaluated
for reusability based on the above-mentioned metrics as shown in Figs. 4, 5 and 6.
It can be inferred from the analysis that there is a positive impact on reusability of
the redesigned software components.
Reusability metrics such as EMI, RCO, RCC, SCCr, and SCCp [4] are applied on
the above mentioned gaming software Poker, Scramble, and Jigsaw. EMI measures
how much the users of the target component can understand its usage. RCO indicates
the component’s degree of observability and understandability. RCC indicates the
component’s degree of customizability and adaptability. Thus, RCO is the ratio of
readable fields to all the fields declared and defined within the target class in the
component. This metric signifies that a large value of readability would assist the
user to comprehend the working of a component from an external environment.
Similarly, RCC is the ratio of writable fields to all the fields declared and defined
within the target class in the component. High value of RCC metric stipulates higher
customizability of the component as per the demand of the user, thus causing inflated
adaptability. SCCr is the ratio of business functions with no return to all the business
functions defined within a component. SCCp is the ratio of business functions with
no parameters to all the business functions defined within a component.

5 Results

This section of the paper presents the results of the assessment, according to two
perspectives. The first section presents the analysis of metrics on applications with
only Shuffler pattern. The second part provides the research outcomes with applica-
tions redesigned with both Shuffler pattern and the related patterns, Singleton and
Prototype.

5.1 Analysis of Metrics—Shuffler Pattern

The games such as Poker, Jigsaw puzzle, and Scramble were redesigned and imple-
mented with Shuffler design pattern. The quality of these software projects is assessed

[email protected]
374 G. Priyalakshmi et al.

using an Eclipse Plug-in for Object-Oriented Metrics, named Metrics. The impact
of Shuffler Design Pattern on the above-mentioned projects are recorded below.
Poker is a gambling card game in which the dealer distributes the shuffled cards to
the players. This game is designed using object-oriented paradigm and implemented
in Java. The same game is rejuvenated using Shuffler Design Pattern and analyzed
using Metrics. The assessment of various object-oriented metrics for the project
before and after Shuffler Design Pattern are shown in Table 5. The improved metrics
are highlighted.
Jigsaw Puzzle is a tiling puzzle game that displays shuffled pieces of an image.
Scramble is a word game that jumbles the letters of a word. The results of similar
metric analysis on Jigsaw Puzzle and Scramble are tabulated in Tables 6 and 7,
respectively.

Table 5 Object-oriented metrics for Poker

Metrics Original Design Redesign
Total Mean Std. Total Mean Std.
Dev. Dev.
Number of overridden methods 8 1 0.707 9 0.9 0.7
Number of attributes 26 3.25 4.918 26 2.6 4.587
Number of children 0 0 0 1 0.1 0.4
Number of classes 8 8 0 10 10 0
Method lines of code 326 5.094 8.148 331 4.868 7.959
Number of methods 64 8 3.969 68 6.8 4.285
Nested block depth 1.484 0.77 1.456 0.756
Depth of inheritance tree 1.75 0.968 1.7 0.9
Afferent coupling 0 0 0 0
Number of interfaces 0 0 0 0 0 0
McCabe Cyclomatic Complexity 2.125 2.684 2.059 2.617
Total lines of code 506 524
Instability 1 0 1 0
Number of parameters 0.156 0.404 0.206 0.439
Lack of cohesion of methods 0.292 0.321 0.234 0.31
Efferent coupling 3 0 3 0
Number of static methods 0 0 0 0 0 0
Normalized distance 0 0 0 0
Abstractness 0 0 0 0
Specialization index 0.258 0.324 0.306 0.378
Weighted methods per class 136 17 17.183 140 14 16.498
Number of static attributes 0 0 0 0 0 0

[email protected]
Impact of Shuffler Design Pattern on Software Quality 375

Table 6 Object-oriented metrics for Jigsaw puzzle

Metrics Original Design Redesign
Total Mean Std. Total Mean Std.
Dev. Dev.
Number of overridden methods 0 0 0 2 0.4 0.8
Number of attributes 11 3.667 4.497 11 2.2 3.919
Number of children 0 0 0 1 0.2 0.4
Number of classes 3 3 0 5 5 0
Method lines of code 125 7.812 13.749 130 6.5 12.808
Number of methods 14 4.667 1.247 18 3.6 1.625
Nested block depth 1.438 0.788 1.35 0.726
Depth of inheritance tree 4.667 1.886 3.4 2.154
Afferent coupling 0 0 0 0
Number of interfaces 0 0 0 0 0 0
McCabe Cyclomatic Complexity 1.688 1.685 1.55 1.532
Total lines of code 232 251
Instability 1 0 1 0
Number of parameters 0.75 1.09 0.8 0.98
Lack of cohesion of methods 0.267 0.377 0.16 0.32
Efferent coupling 0 0 0 0
Number of static methods 2 0.667 0.943 2 0.4 0.8
Normalized distance 0 0 0 0
Abstractness 0 0 0 0
Specialization index 0 0 0.4 0.8
Weighted methods per class 27 9 4.32 31 6.2 4.792
Number of static attributes 0 0 0 0 0 0

The NOM metric in each class of redesigned Scramble project is depicted in

Table 8 for clarification. The total number of methods in that project is 6 since there
are 2 methods in WordShuffler.java file, 2 methods in Word.java file and 2 in Scramble
Shuffler.java. The number of files and the number of methods in that project are 4 and
6, respectively. The average number of methods in the package ScrambleWithPattern
is 6/4, which is 1.5. The standard deviation is 0.886 from the mean.
The three highlighted metrics from the above tables has been plotted in the bar
graphs. Figures 1, 2 and 3 represents the changes in McCabe Cyclomatic Complexity,
Lack of Cohesion of methods and Specialization Index, respectively, for all the above-
mentioned projects.
According to the CBD model [4], the measurement values of all five metrics are
always normalized to a number between 0 and 1. The confidence co-efficient of
the confidence interval of each metric is 95%. The metric quality is believed to be

[email protected]
376 G. Priyalakshmi et al.

Table 7 Object-oriented metrics for Scramble

Metrics Original Design Redesign
Total Mean Std. Total Mean Std.
Dev. Dev.
Number of overridden methods 0 0 0 2 0.5 0.866
Number of attributes 1 0.5 0.5 1 0.25 0.433
Number of children 0 0 0 1 0.25 0.433
Number of classes 2 2 0 4 4 0
Method lines of code 12 4 2.16 17 2.429 2.129
Number of methods 2 1 1 6 1.5 0.866
Nested block depth 1.333 0.471 1.143 0.35
Depth of inheritance tree 1 0 1.25 0.433
Afferent coupling 0 0 0 0
Number of interfaces 0 0 0 0 0 0
McCabe Cyclomatic Complexity 1.333 0.471 1.143 0.35
Total lines of code 35 53
Instability 1 0 1 0
Number of parameters 0.667 0.471 0.857 0.35
Lack of cohesion of methods 0 0 0 0
Efferent coupling 0 0 0 0
Number of static methods 1 0.5 0.5 1 0.25 0.433
Normalized distance 0 0 0 0
Abstractness 0 0 0 0
Specialization index 0 0 0.5 0.866
Weighted methods per class 4 2 0 8 2 0
Number of static attributes 0 0 0 0 0 0

enormous, if the metric value M is in a confidence interval [LLM, ULM], where

LLM is the lower confidence limit and ULM is the upper confidence limit.
Figure 4 shows the measurement of component-based metrics for Poker module.
Out of the five metrics, EMI and SCCp are within the 95% confidence interval, thus
the quality factor corresponding to these metrics is suitably high. The remaining
metrics RCO, RCC and SCCr illustrate lesser measures in their corresponding quality
factors.
The five metrics applied to the Scramble module redesigned using Shuffler design
pattern are shown in Fig. 5. The Scramble software shows higher values only in the
metric EMI. The quality factors corresponding to RCO, RCC, SCCr and SCCp
metrics show lowered measures.
The measurement of metrics for Jigsaw module is shown in Fig. 6. Out of the
five metrics, EMI, RCO and SCCr are within the 95% confidence interval, thus the

[email protected]
Impact of Shuffler Design Pattern on Software Quality 377

Table 8 NOM metric for scramble

Number of Methods 6 1.5 0.866 2 /ScrambleWithPattern/src/Scramble
(avg/max per type) WithPattern/WordShuffler.java
src 6 1.5 0.866 2 /ScrambleWithPattern/src/Scramble
WithPattern/WordShuffler.java
ScrambleWithPattern 6 1.5 0.866 2 /ScrambleWithPatternAro/Scramble
WithPattern/WordShuffler.java
WordShuffler.java 2 2 0 2 /ScrambleWithPattern/src/Scramble
WithPattemMordShuffler.java
Word Shuffler 2
Word.java 2 2 0 2 /ScrambleWithPattern/src/Scramble
WithPattern/Word.java
Word 2
SorambleShuffler.java 2 2 0 2 /ScrambleWithPattern/5ro/Scramble
WithPattern/ScrambleShuffler.java
ScrambleShuffler 2
WordPrintjava 0 0 0 0 /ScrambleWithPattern/src/Scramble
WithPattern/WordPrint.java
WordPrint 0

Fig. 1 Comparison of McCabe Cyclomatic complexity metric before and after the Shuffler pattern

[email protected]
378 G. Priyalakshmi et al.

Fig. 2 Comparison of Lack of Cohesion of Methods Metric before and after the Shuffler pattern

Fig. 3 Comparison of specialization index metric before and after the Shuffler pattern

quality factor corresponding to these metrics is suitably high. The remaining metrics
RCC and SCCp illustrate lesser measures in their corresponding quality factors.
In order to obtain a clear understanding of how various object-oriented metrics
impact the software quality, various software quality measures are assessed against
the object-oriented metrics and are tabulated in Table 9. The impact of McCabe
Cyclomatic Complexity, Lack of Cohesion of Methods, Specialization Index on var-
ious software quality attributes are shown in the table. All the metrics have an impact
on maintainability, understandability, and reliability of the software. Lack of Cohe-

[email protected]
Impact of Shuffler Design Pattern on Software Quality 379

Card Shuffler Deck

deck : ArrayList

Poker Shuffler
Deck()
Shuffler() : void
Draw() : void
Equals(Deck) : boolean
EMI(Deck) = 1 getDeck() : Arraylist
RCI(Deck) = 1 toString() : String
RCC(Deck) = 0
SCCr(Deck) = 0.25
SCCp(Deck) = 0.75

Fig. 4 Component based metrics for Poker

Word Shuffler Word

Generator: Random

Scramble Shuffler
Word()
Scramble(List) : List

EMI (Word) = 1
RCI (Word) = 0
RCC(Word) = 0
SCCr (Word) = 0
SCCp (Word) = 0

Fig. 5 Component based metrics for scramble

sion of Methods impacts most attributes. The table depicting the fourteen software
quality attributes and three design metrics is tabulated to infer the dependency of
high-level quality attributes and low-level design metrics. The decision table entries
for Cyclomatic Complexity are completed from the observations in the hierarchical
model [8]. Similarly, from the comparisons and analyses in various research studies
[4, 6, 8, 13], the other table entries are filled. It is evident from the table that the
assessed object-oriented metrics do have a positive impact on the reusability of the
software and other software quality attributes.
The paper interprets that McCabe Complexity influences software quality 42%
and Specialization index has 50% impact on quality. The Lack of Cohesion of Meth-

[email protected]
380 G. Priyalakshmi et al.

PuzzleEx
Puzzle Shuffler

panel: JPanel source : BufferedImage

height : int solution : ArrayList[]
img : Image button : Button
width : int button : ArrayList
widefix : int resize:BufferedImage
Jigsaw Shuffler

PuzzleEx()
EMI (PuzzleEx) = 1 initUI() : void
RCI (PuzzleEx) = 0.3 getNewHeight(int,int) : int
RCC(PuzzleEx) = 0 loadImage() : BufferedImage
SCCr (PuzzleEx) = 1 resizeImage(BufferedImage,int,int,int) : BufferedImage
SCCp (PuzzleEx) = 1

Fig. 6 Component based metrics for Jigsaw

Table 9 Decision table showing the interdependence between software quality measures and
object-oriented metrics
McCabe Cyclomatic Lack of cohesion of Specialization index
Complexity methods
Expandability No Yes No
Reusability No Yes Yes
Flexibility Yes Yes No
Modularity Yes Yes No
Generality No No No
Scalability No No No
Robustness No No No
Stability No Yes Yes
Adaptability No No Yes
Maintainability Yes Yes Yes
Reliability Yes Yes Yes
Simplicity Yes No No
Learnability No Yes Yes
Understandability Yes Yes Yes

[email protected]
Impact of Shuffler Design Pattern on Software Quality 381

ods has the maximum hit of 64%. On the other hand, the study concludes that
Maintainability, Reliability, and Understandability are the qualities, which are highly
influenced by the Shuffler design pattern. The other quality features may or may not
show improvement with Shuffler pattern.

5.2 Analysis of Metrics—Shuffler Combined with Related

Patterns

A combination of various related design patterns were applied to the existing projects.
Along with the Shuffler design pattern, Singleton pattern was applied to Jigsaw
project and the impact was measured by the Eclipse plug-in Metrics and the results
are provided in Table 10. There was no significant impact of the combined patterns
when compared to that of the project implementation with only Shuffler pattern. The
Shuffler pattern was combined with Singleton pattern and the results are tabulated
in Table 11. The results were similar to the results in Table 7 with no significant
improvement in any of the metrics. It was also observed that it had negative impact
when the Scramble project was implemented with Prototype design pattern. This
impact is evident from the highlighted metrics in Table 12.

6 Conclusion

The most vital state of the art research on design patterns is; the impact of design
patterns on quality features of a software. This paper aims at emphasizing the effect
of pattern application on quality attributes. The first research effort made by the
authors in the direction of design patterns was the identification of Shuffler design
pattern and more importantly its known uses. After a thorough literature survey on
papers related to GOF design patterns, the impact of Shuffler pattern on three open
source projects was studied. With this study on the impact of Shuffler design pattern
on the software quality metrics tested for gaming applications, it has been analyzed
that the software quality metrics such as McCabe Cyclomatic Complexity, Lack of
Cohesion of Methods and Specialization Index are improved. The software quality
attributes that got affected by the pattern were listed in the Results section.
In addition to the conventional metrics, this study also focuses on reusability
metric to measure the reuse of the black-box components of the three gaming appli-
cations. The advantage of this analysis is, it would be helpful to designers when the
source code of components is not available. The CBD analysis carried out for the
projects supported to prove that the use of Shuffler pattern increases software quality.
The paper also relates high-level software quality measures to the three low-
level object-oriented metrics, McCabe Cyclomatic Complexity, Lack of Cohesion of
Methods and Specialization Index. The study finally measures the software quality

[email protected]
382 G. Priyalakshmi et al.

Table 10 Object-oriented metrics for Jigsaw with Singleton

Metrics With Shuffler pattern only With Shuffler and Singleton
pattern
Total Mean Std. Total Mean Std.
Dev. Dev.
Number of overridden methods 2 0.4 0.8 2 0.4 0.8
Number of attributes 11 2.2 3.919 11 2.2 3.919
Number of children 1 0.2 0.4 1 0.2 0.4
Number of classes 5 5 0 5
Method lines of code 130 6.5 12.808 130 6.5 12.808
Number of methods 18 3.6 1.625 18 3.6 1.625
Nested block depth 1.35 0.726 1.35 0.726
Depth of inheritance tree 3.4 2.154 3.4 2.154
Afferent coupling 0 0 0 0
Number of interfaces 0 0 0 0 0 0
McCabe cyclomatic complexity 1.55 1.532 1.55 1.532
Total lines of code 251 251
Instability 1 0 1 0
Number of parameters 0.8 0.98 0.8 0.98
Lack of cohesion of methods 0.16 0.32 0.16 0.32
Efferent coupling 0 0 0 0
Number of static methods 2 0.4 0.8 2 0.4 0.8
Normalized distance 0 0 0 0
Abstractness 0 0 0 0
Specialization index 0.4 0.8 0.4 0.8
Weighted methods per class 31 6.2 4.792 31 6.2 4.792
Number of static attributes 0 0 0 0 0 0

of the gaming applications with Shuffler and related patterns like Singleton and
Prototype.

[email protected]
Impact of Shuffler Design Pattern on Software Quality 383

Table 11 Object-oriented metrics for Scramble with Singleton

Metrics With Shuffler pattern only With Shuffler and
Singleton pattern
Total Mean Std. Total Mean Std.
Dev. Dev.
Number of overridden methods 2 0.5 0.866 2 0.5 0.866
Number of attributes 1 0.25 0.433 1 0.25 0.433
Number of children 1 0.25 0.433 1 0.25 0.433
Number of classes 4 4 0 4 0
Method lines of code 17 2.429 2.129 17 2.429 2.129
Number of methods 6 1.5 0.866 6 1.5 0.866
Nested block depth 1.143 0.35 1.143 0.35
Depth of inheritance tree 1.25 0.433 1.25 0.433
Afferent coupling 0 0 0 0
Number of interfaces 0 0 0 0 0 0
McCabe Cyclomatic Complexity 1.143 0.35 1.143 0.35
Total lines of code 53 53
Instability 1 0 1 0
Number of parameters 0.857 0.35 0.857 0.35
Lack of cohesion of methods 0 0 0 0
Efferent coupling 0 0 0 0
Number of static methods 1 0.25 0.433 1 0.25 0.433
Normalized distance 0 0 0 0
Abstractness 0 0 0 0
Specialization index 0.5 0.866 0.5 0.866
Weighted methods per class 8 2 0 8 2 0
Number of static attributes 0 0 0 0 0 0

[email protected]
384 G. Priyalakshmi et al.

Table 12 Object-oriented metrics for Scramble with Prototype

Metrics With Shuffler pattern only With Shuffler and
Prototype pattern
Total Mean Std. Total Mean Std.
Dev. Dev.
Number of overridden methods 2 0.5 0.866 3 0.5 0.764
Number of attributes 1 0.25 0.433 2 0.333 0.471
Number of children 1 0.25 0.433 2 0.333 0.471
Number of classes 4 4 0 6 6 0
Method lines of code 17 2.429 2.129 40 4 7.085
Number of methods 6 1.5 0.866 9 1.5 0.764
Nested block depth 1.143 0.35 1.3 0.9
Depth of inheritance tree 1.25 0.433 1.333 0.471
Afferent coupling 0 0 0 0
Number of interfaces 0 0 0 0 0 0
McCabe Cyclomatic Complexity 1.143 0.35 1.5 1.5
Total lines of code 53 85
Instability 1 0 1 0
Number of parameters 0.857 0.35 0.6 0.49
Lack of cohesion of methods 0 0 0 0
Efferent coupling 0 0 0 0
Number of static methods 1 0.25 0.433 1 0.167 0.373
Normalized distance 0 0 0 0
Abstractness 0 0 0 0
Specialization index 0.5 0.866 0.5 0.764
Weighted methods per class 8 2 0 15 2.5 2.062
Number of static attributes 0 0 0 0 0 0

[email protected]
Impact of Shuffler Design Pattern on Software Quality 385

References

1. Gamma, E., Helm, R., Johnson, R., & Vlissides, J. (1995). Design patterns: Elements of
reusable object-oriented software. ACM Digital Library.
2. Ampatzoglou, A., Charalampidou, S., & Stamelos, I. (2013). Research state of the art on GoF
design patterns a mapping study. Journal of Systems and Software.
3. Priyalakshmi, G., Nadarajan, R., & Anandhi, S. Software reuse with shuffler design pattern.
SEAT.
4. Washizaki, H., Yamamoto, H., & Fukazawa, Y. (2003). A metrics suite for measuring reusability
of software components. In Proceedings of 9th International Software Metrics Symposium,
METRICS’ 03.
5. Bieman, J. M., & Wang, H. (2006). Design pattern coupling, change proneness, and change
coupling: A pilot study. Computer Science Technical Report.
6. Lange, D. B., & Nakamura, Y. (1995). Interactive visualization of design patterns can help in
framework understanding. New York: ACM Press.
7. Khomh, F., & Gueheneuc, Y.-G. (2008). Do design patterns impact software quality positively?.
New York: IEEE.
8. Bansiya, J., & Davis, C. G. (2002, January). A hierarchical model for object-oriented design
quality assessment. IEEE Transactions on Software Engineering.
9. Wendorff, P. (2001, March). Assessment of design patterns during software reengineering:
Lessons learned from a large commercial project. In Proceedings of 5th Conference on Software
Maintenance and Reengineering (pp. 77–84).
10. Cohesion metrics: Retrieved March 21, 2018. http://www.aivosto.com/project/help/pm-oo-
cohesion.html.
11. Software metrics in eclipse: September 2, 2005, Retrieved March 21, 2018. http://
realsearchgroup.org/SEMaterials/tutorials/metrics.
12. Wand, A. J. A. (2002). Reuse metrics and assessment in component-based development. In Pro-
ceedings of 6th IASTED International Conference on Software Engineering and Applications,
IASTED (pp. 583–588).
13. Henderson-Sellers, B. (1996). Object-oriented metrics, measures of complexity. Englewood
Cliffs: Prentice Hall.
14. Jabangwe, R., Orstler, J. B., Smite, D., & Wohlin, C. Empirical evidence on the link between
object-oriented measures and external quality attributes: A systematic literature review.
15. Eclipse metrics plug-in: December 27, 2002, Retrieved March 21, 2018. https://sourceforge.
net/projects/metrics.
16. Analyst4j: Provides Java code search using software metrics, Retrieved March 21, 2018. https://
www.theserverside.com/discussions/thread/44693.html.
17. Alternative tools LOC metrics, Retrieved March 21, 2018. http://www.locmetrics.com/
alternatives.html.
18. Chidamber and Kemerer Java metrics: May 22, 2010, Retrieved March 21, 2018. http://www.
spinellis.gr/sw/ckjm.
19. Dependency Finder: Tool to analyse Java codes, Retrieved March 21, 2018. http://depfind.
sourceforge.net.
20. Testwell: Complexity measures tool for Java, September 20, 2012, Retrieved March 21, 2018.
http://www.testwell.fi/cmtjdesc.html.
21. Resource Standard Metrics: Quality analysis tool, Retrieved March 21, 2018. http://www.
msquaredtechnologies.com/index.html.
22. CodePro AnalytiX: Java software testing tool for Eclipse developers, Retrieved March 21,
2018. https://dzone.com/articles/codepro-integration-eclipse.
23. Automated tools for software engineering: Retrieved March 21, 2018. http://www.semdesigns.
com.
24. JDepend: Generate design quality metrics for given Java package, Retrieved March 21, 2018.
http://clarkware.com/software/JDepend.html.

[email protected]
386 G. Priyalakshmi et al.

25. Object oriented software metrics—A short guide: Retrieved March 21, 2018. http://www.
virtualmachinery.com/jhawkmetrics.htm.
26. jMetra: Output project metrics to XML files, May 12, 2006, Retrieved March 21, 2018. http://
www.loribel.com/java/tools/jmetra.html.
27. JMetric: OO-metrics tool, June 19, 2001, Retrieved March 21, 2018. https://sourceforge.net/
projects/jmetric.
28. Essential metrics: Command line metrics for C, C++, and Java projects, February 01, 2015,
Retrieved March 21, 2018. http://www.powersoftware.com/em.
29. Refactorit: Automated refactorings, April 21, 2016, Retrieved March 21, 2018. https://
sourceforge.net/projects/refactorit.
30. SonarJ: Visualize and analyze structure of Java code, March 21, 2015, Retrieved March 21,
2018. http://sonarj.software.informer.com.
31. Alghamdi, J. S., Rufai, R. A., & Khan, S. M. (2005). OOMeter: A software quality assurance
tool. In Proceedings of 9th European Conference on Software Maintenance and Reengineering.
32. SemmleCode: Eclipse plugin to query Java code, June 01, 2007, Retrieved March
21, 2018. http://www.developerfusion.com/thread/47560/semmlecode-free-eclipse-plugin-to-
query-java-code.
33. Understand: Static code analysis tool, Retrieved March 21, 2018. https://scitools.com.
34. Löwe, W., Ericsson, M., Lundberg, J., Panas, T., & Pettersson, N. (2003, October).
Vizzanalyzer-a software comprehension framework. In Third Conference on Software Engi-
neering Research and Practise in Sweden. Sweden: Lund University.

[email protected]
Efficient Algorithms for Text Lines
and Words Segmentation for Recognition
of Arabic Handwritten Script

Amani Ali Ahmed Ali and M. Suresha

Abstract A new methodology for Arabic handwritten document images segmenta-

tion is done in this paper to segment the documents into distinct entities as words and
text lines. Based on features of Arabic scripts, the document images are divided into
three main subsets of connected components where the Hough transform method is
applied to them to achieve text lines segmentation. To enhance the result by avoiding
the Hough transform text line detection failure, the authors used a method in postpro-
cessing stage based on skeletonization that covers the possible false correction alarms
to create proficiency vertical connected characters’ segmentation. The segmentation
of the Arabic words is pointed as a two-class problem. The authors used fusion of
convex and Euclidean distance metrics to calculate the distance between neighbor-
ing overlapped components, which in the Gaussian mixture modeling framework is
classified as a distance of an intra-word or as an inter-word. The proposed method
performance is depended on a constant and particular evaluation method that appro-
priate measures of the performance used to compare the segmentation of our result
against the other strong researcher result. The proposed method showed higher effi-
ciency and accuracy in the experimentation, which was conducted on two various
Arabic handwriting datasets that are IFN/ENIT and AHDB.

Keywords Gaussian mixture modeling · Handwritten document images · Hough

transform · Word segmentation · Text line segmentation

1 Introduction

In the area of recognition of handwritten document to segment, segmenting the

document images into their fundamental entities such as words and text lines is a
very complex problem in Arabic handwritten document images because various kinds

A. A. A. Ali (B) · M. Suresha

Department of Computer Science, Kuvempu University, Shimoga 577451, Karnataka, India
e-mail: [email protected]
A. A. A. Ali
Department of Computer Science, Taiz University, Taiz, Yemen
© Springer Nature Singapore Pte Ltd. 2019 387
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_32

[email protected]
388 A. A. A. Ali and M. Suresha

of difficulties are faced in both process of words and text line segmentation. These
complexes make them a challenging task. The main difficulties in the segmentation
process of Arabic text line are overlapping words, neighboring text lines touching,
and over same text line or on the page between lines the variances angle of the skew.
The main difficulties in segmentation procedure of Arabic words are nonuniform
spacing which is popular in handwritten words, the punctuation marks within the
text line, and the slant and skew in the text line.
A new methodology is introduced in this paper for Arabic handwritten document
images segmentation, segmenting the documents into their distinct entities such as
text lines and words. Basically, an improved method is included in the proposed
method for the vertically connected separation for the text lines, and the new method
for word segmentation depends on an efficient recognition of intra-word and inter-
word gaps using combination of two various distance metrics, which are convex
hull-based and Euclidean distance. The recognition of two classes treated as cluster-
ing problem of an unsupervised, Gaussian mixture theory used for model these two
classes. The efficiency of the proposed method on two various datasets with various
collections of gap metrics is proved in the experimentation. The Arabic handwriting
datasets used are IFN/ENIT and AHDB. A fixed benchmarking enables the compar-
ison between existing researchers’ methods and the proposed methods. This paper
is organized as follows: the related work is presented in Sect. 2 for the segmentation
of text line and words with text line of Arabic handwritten document images. The
proposed words and text lines segmentation methods are detailed in Sect. 3. The
experimental results are presented in Sect. 4, and Sect. 5 describes the conclusions
with the future work.

2 Related Work

In the Arabic handwritten document images literature, broad variety methods of

segmentation have been reported. The authors classified these methods based on
whether they refer to the segmentation of word or text line or both word and text line.

2.1 Text Line Segmentation

In [10], the authors used both morphological dilatation and projection profile. To
rate skew of the line, the horizontal projection profile method is used. In every zone
for smearing, the slope is used, using dilation with adaptive structuring element to
do the changes according to the zone, the slope, and the size. The big blobs are
detected in the second phase with a recursive function that search for the cut point
because segment the components and match them with their lines following repulsion
and attraction criterion. Using erosion recursively to do the thinness portion touching
components detection which matches the cut point these are detected and segmented.

[email protected]
Efficient Algorithms for Text Lines and Words Segmentation … 389

In [2], the authors used projection profile method after joining broken characters
and removing small components to determine the point of separation within the hori-
zontal projection profile; the curve of Fourier fitting is used. To segment the baseline
of the connected component which allows determining the cut point between various
neighbor lines, the contour is used. Regarding the nearest line, the components are
determined, which are approximated by the curve of polynomial that suits in the
baselines the pixels. According to the components of the closest connected nearest
neighbor in four directions, the lifted of small size components are reassigned to
their line.
In [11], the authors begin the algorithm by deleting diacritics, and then, the sparse
similarity graph is built based on the local orientation of the component. By breadth
first search (BFS) via set of disjoint of text line are performed. To assign to the lines
of text the blobs, the affinity propagation clustering method is used.
In [9], the algorithm in this technique begins by eliminating outlier components
by using the value of threshold; after that, the letters related to two lines at the half
distance are detected and segmented horizontally. For line detection, a rectangular
neighborhood on a current component is centered and increases to contain which
satisfy specific conditions. At the beginning from their bounding box with respect to
distance, the filtered components to the corresponding lines are allocated.
In [8], the algorithm in this technique shows a new algorithm for automatic line of
the text segmentation from Arabic handwritten documents, presenting the problems
of a multi-touching and an overlapping characters. Using unsupervised method, their
method depended on analysis of a block covering. The algorithm in this method in
the first step performs the analysis of a statistical block which calculates the typical
number into vertical strips of document decomposition. Their method in the next
point used fuzzy C means technique which achieves line detection based on fuzzy
base. In the final phase to its corresponding lines, blocks are assigned. Experiment
results of this approach showed high accuracy, around 95% for detecting lines in
Arabic handwritten document images which were written with various document
images.

2.2 Word Segmentation

In [3], the authors realized a static threshold for gaps classification as “within-a
word” and “between-words”, by analyzing more than 250 words that were taken
from the IFN/ENIT database through 200 document images. For the errors of the
minimum classification, the Bayesian criterion was employed. This method achieved
an accuracy rate of around 85%.
In [22], the authors have proposed a new technique for words segmentation on
off-line Arabic handwritten. It segmented the connected characters into smaller com-
ponents where every one of them does not contain more than three characters. Every
character may be at most segmented into five parts. Additionally, developing of
Arabic words recognition, another possible application of the segmentation of the

[email protected]
390 A. A. A. Ali and M. Suresha

proposed method is to build of small size lexicon, including no combinations more

than three characters. Generally because lot number of word in Arabic language that
make produce lexicon for the unconstraint handwritten Arabic document images
recognition too complex.

2.3 Text Line and Words Segmentation

Al-Dmour and Fraij [4] realized dynamic gaps as “within-a word” and “between-
words”, which were taken from the input document image itself. Experiments using
various clustering methods were performed with an accuracy rate of 84.8%.

3 Methodology

The following challenges are faced by the proposed method for Arabic handwritten
segmentation from document images treated with:
• overlapping and touching,
• skewed or slanting, and
• short lines.

3.1 Text Line Segmentation

The authors used the method of connected components to segment the document
images into a group of CCs which help us to join probably connecting to each other
with the same line which is connected. To join all the components that belong to the
same line:
Step 1: In the beginning, the authors used a rating threshold to remove the diacritical
components from the documents; those will add them to their corresponding lines in
the final phase.
Step 2: Based on the average height M(H) of every connected component for the
whole document image, the average of the character height is calculated. The authors
suppose the average of the character width M(W ) and the average of character height
is equal.
Step 3: For every component meet the majority of characters size “Subset 1” is
predicted to comprise them where the following constraint is satisfied:

(0.5 ∗ M(H ) ≤ H < 3 ∗ M(H )) and (0.5 ∗ M(W ) ≤ M(W )) (1)

[email protected]
Efficient Algorithms for Text Lines and Words Segmentation … 391

where W and H, respectively, are the component’s width and height, and M(W ) and
M(H), respectively, the average width and the average height of the character.
Step 4: For all large connected components, “Subset 2” is expected to comprise
them where the characters from neighbor text lines which touch each other are the
large components. The following equation defined the height of these connected
components:

H ≥ 3 ∗ M(H ) (2)

Step 5: Punctuation marks, accents, diacritic marks, Tanween, Hamza, Sha’ada, and
the dots of characters, all these characters should include in “Subset 3”. The following
equation defined this:

((H < 3 ∗ M(H )) and (0.5 ∗ M(W ) > W ))

or
((H < 0.5 ∗ M(H )) and (0.5 ∗ M(W ) < W )) (3)

Step 6: The gravity center is calculated in this step which contained in every block
of connected component after creating the blocks.
Step 7: A line to transform the Cartesian space into the Polar coordinate is Hough
transform. The following equation defined the line in the space of the Cartesian
coordinate:

x cos(θ ) + y sin(θ )  p (4)

The line in the space of Cartesian is defined via designating in the space of
coordinate of the Polar where their coordinates are p and θ . In the subset, there are
matches between the accumulator array and every gravity center to a group of cells
of the domain ( p, θ ). The resolution along the direction of θ was set to 1°, where
the value of θ was in the domain of 85–95° and over direction (p) their solution set
to 0.2* M(H) to build the domain of Hough as shown in Eq. (4).
Step 8: The cell ( pi , θi ) is detected by the authors with maximum contribution and
assigned every point which vote to the line of the text ( pi , θi ) in the area ( pi − 5, θi )
… ( pi + 5, θi ). To specify if a connected component belongs to a text line, at least
half of the points which form the corresponding blocks must be reallocated to this
area. All votes from the accumulator array of the Hough transform that matches this
connected component are deleted after the assignment to a connected component text
line. This process is repeated in order to avoid false alarms till cell ( pi , θi ) having the
maximum contribution comprises less than n1 votes. Through the evolution of the
process, the angle of the skew of actually detected lines is calculated. The authors
applied more constraint in the case when the cell ( pi , θi ) with less than n 2 (n 2 > n 1 )
contribution has as a maximum contribution; in that case, the text line is only correct
if the matching of the line skew angle deflects from the skew angle less than 2°.

[email protected]
392 A. A. A. Ali and M. Suresha

Step 9: To improve the result through Hough transform via making correction for
some false alarms, the authors applied a merging method, yi mentioned to the average
y detected line intersection i values are calculated and the bounding box determined
by (i  1, … , n) of connected components are calculated.
Step 10: If the condition at Eq. (5) is not satisfied, the authors exclude the last line
n from the process; all the components of this subset belong to n detected text lines
(n > 1).
xe
 ye
xxs I (x, y)
yyn −(yn −yn−1 )/10
xe > 0.08 (5)
 ye
xxs I (x, y)
yyn−1

Here the image of the component is I which value 1 mentioned for foreground
pixel representation and 0 background pixel representation and are the component
coordinates bounding box. The component location near line n is not because a long
character descender from text line n − 1 and because a vertical character merging
which verified in Eq. (5).
Step 11: The authors define zones Z i for every line i(i  1, …, n − 1), with consid-
eration of the following constraint:

yi+1 − yi
yi + < y < y+1 (6)
2
The connected component skeleton is computed. We remove all points of the
junction and remove them if they are inside zone Z i from the skeleton. In the Z i
zone of the segmentation, the authors remove all points of the skeleton on the zone
center if there does not exist any junction point.
Step 12: With Flag id “1” for all zone Z i , the skeleton portions with line i are inter-
sected. With flag id “2”, all other portions are flagged. In every zone Z i , the connected
component segmentation in the initial stage into various portions is accomplished in
the last by assigning the nearest skeleton pixel id to a pixel.

3.2 Word Segmentation

The segmentation process of the Arabic words is as follows:

• The first phase treats the text line with neighbor components computation distances
of the images of the document, and
• The second phase treats the gaps of the classification as inter-word or inter-
character of previously computed distances.

[email protected]
Efficient Algorithms for Text Lines and Words Segmentation … 393

For the first step twos, the authors proposed different average metrics which are
convex hull-based [15] combined with the Euclidean distance metric [19], utilizing a
well-known technique which is the Gaussian mixtures where the computed distances
classification from the part of unsupervised clustering methods is performed.

3.2.1 Distance Computation

A preprocessing is applied in the text line image to calculate the neighbor components
distance. The preprocessing treats the text line of document image with a dominant
slant angle and the skew angle correction [21]. On only overlapped components
(OCs), the gap metric computation is considered; the OC is a group of connected
components where their projection profiles overlap within the vertical direction and
not considered on (CCs) the connected components. The authors determine the aver-
age value distance based on the convex hull and the Euclidean distance as the distance
of two neighbor OCs. The minimum between of each pairs Euclidean distances of the
two neighbor OCs points is determined the Euclidean distance among two adjacent
OCs. For the Euclidean distance calculation, the authors apply a quick planned that
takes into consideration the pixels’ subset of the right and left OCs only rather than
the whole number of black pixels. This subset comprises the black pixel of the right-
most of every scan line instead of determining the subset with the left pixels of OC.
Including the leftmost black pixel of all scan line is defined the subset of pixels for the
right OC. The minimum Euclidean distances of every pixels pairs are defined finally
the distance of the Euclidean of two OCs. As the following, the authors calculate
the metric based on the convex hull: Ci and Ci+1 , which are a pair of neighbor OCs,
and Hi and Hi+1 , which are their convex hulls are given. The line which linking the
gravity centers (or centroid) of Hi and Hi+1 is denoted by L. The intersection points
of L with the hulls Hi and Hi+1 are denoted by Pi and Pi+1 , respectively. Between
the points Pi and Pi+1 , the Euclidean distance has defined the gap between the two
convex hulls.

3.2.2 Gap Classification

A novel method is used for the gap classification problem. This method is depended
on the classification of the unsupervised the already computed distances into two
distinguishable classes forming, respectively, the word interclass and the word intra-
class. The authors adopt the Gaussian mixtures method to this end. A clustering with
mixture model based on every cluster is mathematically introduced via a parametric
distribution. The authors have a two-cluster problem; hence, every cluster is modeled
with a Gaussian distribution. The EM algorithm is the algorithm that is used to
calculate the parameters for the Gaussians. Because the Gaussian mixture is a well-
known approach for unsupervised clustering, the authors used this method with many
advantages which include the following:

[email protected]
394 A. A. A. Ali and M. Suresha

i. A soft classification is available.

ii. For each cluster, the density rating can be obtained.
iii. The mixture model covers well the data. For the Gaussian mixtures, the detailed
description is given in [16].

4 Evaluation and Experimental Result

4.1 Performance Evaluation Methodology

Observation manually of the result of segmentation is not in every case unbiased

procedure, a high time-consuming, and boring. To avoid user interference, the authors
used an automatic evaluation of the performance method that compared the result of
detected segmentation with a previously annotated ground truth. The performance
evaluation is depended on counting the matches number between the areas inside
the areas which detected by the method and the areas inside the ground truth. There
is a table the authors used it where its values are calculated according to the labeled
pixel sets overlap as the ground and either text lines or words.
Where the set of every image points is I, all points set inside the i word or text line
ground truth region is G i ; all points set inside the j word or text line result region is
R j , a function that counts the set elements s is T (s). As follows, table MS(i, j) forms
the corresponding results of the region j of result and the region i of ground truth:
 
T Gi ∩ R j ∩ I
MS(i, j)     (7)
T Gi ∪ R j ∩ I

where MS refer to MatchScore, along the MatchScore table the performance eval-
uator seeks for one-to-one, many-to-one, or one-to-many corresponding pairs. o2o
denotes a pair of one-to-one correspond if the pair corresponding score is above or
equal to the acceptance threshold θ evaluator. go2m denotes g_one-to-many corre-
spond which is a word or text line ground truth, partially that corresponds with the
detected result of more than one of word or text line. gm2o denotes g_many-to-one
correspond, which corresponds to more than one of the ground truth word or line of
text, partially that corresponds with one word or text line in the detected result. do2m
denotes d_one-to-many correspond which is detected word or text line, partially that
corresponds with two or more words or line in the text in the ground truth. dm2o
finally denotes d_many-to-one correspond, which corresponds with more than one
words or detected text lines, partially that match one word or line of the text in the
ground truth. In the following, the equations of detection rate (DR) and recognition
accuracy (RA) are as follows:

o2o g_o2m g_m2o

DR  w1 + w2 + w3 (8)
N N N

[email protected]
Efficient Algorithms for Text Lines and Words Segmentation … 395

o2o d_o2m d_m2o

RA  w4 + w5 + w6 (9)
M M M
where the ground truth segments words or lines of the text count are denoted by
N, the result segments words or text lines count are denoted by M, and predeter-
mined weights are w1 , w2 , w3 , w4 , w5 , w6 . From Eq. (7), the following o2o, g_m2o,
d_o2m, g_o2m, and d_m2o entities which correspond to the one-to-one, g_many-
to-one, d_one-to-many, g_one-to-many, and d_many-to-one numbers are calculated
following the steps of [18]. If the authors merge the values of RA and DR the metric
of a global performance can be defined. The F-measure (FM) equation is as follows:

2 ∗ DR ∗ RA
FM  (10)
DR + RA

4.2 Experimental Results

The proposed algorithm is tested on various Arabic handwritten databases, which are
IFN/ENIT and AHDB databases. The authors conducted various experiments and
explained them in detail in the following parts. For the segmentation of the text line,
the authors implemented the Hough-based method [12], the fuzzy RLSA [20], and
the projection profiles [7]. For skew correction, the authors implemented the method
based on [1].
In the proposed method of segmentation of the text line, the parameters n 1 and n 2
in (Sect. 3.1) are set with n 1  5, n 2  9 experimentally. Furthermore in (Sect. 4.1)
for the segmentation of word and text line, the acceptance threshold θ evaluator’s is
defined, respectively, as 0.99 and 0.94.

4.2.1 AHDB

The AHDB which denotes Arabic handwriting database [5, 6] is an Arabic hand-
writing database processed with numerous preprocessing. It consists of words; para-
graphs and the words are used to form digits conducted by 100 various writers on
checks of Arabic handwritten. The AHDB contained unconstrained pages’ text. The
proposed methodologies’ efficiency was tested by proceeding experiments on Arabic
handwritten document images.
The authors have the corresponding ground truth for all the required images which
comprise words and text lines. The number of words is 13,311, and the corresponding
number of lines is 1773. As shown in detail in Table 1, the segmentation results of the
text line in the cases of FM, RA, and DR and the number of corresponds are given. The
authors combined the distance metrics defined with two gap classification techniques
in Sect. 3.2.1 for concerning word segmentation. These include the following:
i. The classification method of Gaussian mixture.

[email protected]
396 A. A. A. Ali and M. Suresha

Table 1 Experimental results for text line segmentation over AHDB handwritten
N M o2o dm2o go2m gm2o do2m RA DR FM
1095 1093 1082 4 2 4 2 99.1 98.9 99.0

Fig. 1 Words segmentation

of free Arabic handwritten
document images from
AHDB database

Fig. 2 Text lines

segmentation of free Arabic
handwritten document
images from AHDB database

ii. The method in [13] which used the lengths median on the scan line of white
pixels that have to white transitions the maximum number of black.
Table 2 shows the experimental segmentation results of words in the cases of FM,
RA, and DR and gives the matches number. EDM, CDM, GMCM, and LTCM in
Table 2 represent Euclidean distance metric, convex distance metric, LT classifica-
tion methods, and GM classification methods. The proposed result of segmentation
method of the text line took it as input to the segmentation module of the words. The
proposed method is parameter free.
Table 3 shows the results of experimental after combination of every segmentation
methodologies of the words with each segmentation methodologies of the text line.
The segmentation results of word process and text line process visualize, respectively,
are shown in Figs. 1 and 2.

4.2.2 IFN/ENIT

The IFN/ENIT includes words of towns’ names and villages’ names in Tunisia and
postal code, which are all represented in the Arabic handwritten database IFN/ENIT
[17]. By filling specified form, the dataset of 411 volunteers has been developed.
In the database, the total names words city and town reached to 26,400 including
210,000 characters. The data of ground truth with the database contains information
on the writer sequence, baseline, and character shapes. All filled forms are stored as
binary and digital images at 300 dpi.

[email protected]
 Table 2 Word segmentation experimental results with fusion of gap classification methodologies and distance metrics over AHDB handwritten
EDM CDM LTCM GMCM N M o2o gm2o go2m dm2o do2m DR RA FM
0 0 1 0 510 1546 235 8080 6655 697 8576 81.1 87.8 84.3
0 1 0 1 1190 956 554 8810 6895 445 8576 85.1 82.5 83.8
1 0 0 1 733 1419 342 8282 6674 643 8576 96.8 92.9 94.8
0 0 0 1 630 1280 292 8305 6854 586 8576 83.4 87.2 85.3

[email protected]
1 0 1 0 1050 843 485 8809 7087 397 8576 86.8 84.2 85.5
Efficient Algorithms for Text Lines and Words Segmentation …
397
398 A. A. A. Ali and M. Suresha

Table 3 Comparison between well-known methods and proposed method

Method Measures Accuracy (%)
[3] Word segmentation 85
[4] Text and word segmentation 84.8
[8] Text line segmentation 95
Proposed work Text line segmentation 98.9
Proposed work Word segmentation 96.8

Table 4 Experimental results for text line segmentation over IFN/ENIT handwritten
N M o2o dm2o go2m gm2o do2m RA DR FM
1773 1770 1717 34 6 13 17 97.4 97.4 97.4

Fig. 3 Words segmentation of free Arabic handwritten document images from IFN/ENIT database

Fig. 4 Text lines segmentation of free Arabic handwritten document images from IFN/ENIT
database

The authors have the corresponding ground truth for all the required images which
comprise words and text lines. The number of words is 8576, and the corresponding
number of lines is 1095. As shown in detail in Table 4, the segmentation results of
the text line in the cases of FM, RA, and DR and the number of corresponds are
given. For word segmentation problem, the experiment that is applied to these sets
is the same which is applied to the AHDB set.
Table 5 shows the experimental segmentation results of words in the cases of FM,
RA, and DR and gives the matches number. EDM, CDM, GMCM, and LTCM in
Table 2 represent Euclidean distance metric, convex distance metric, LT classifica-
tion methods, and GM classification methods. The proposed result of segmentation
method of the text line took it as input to the segmentation module of the words. The
proposed method is parameter free. Table 6 shows the results of experimental after
combination of every segmentation methodologies of the words with each segmen-
tation methodologies of the text line. The segmentation results of word process and
text line process visualize, respectively, are shown in Figs. 3 and 4.

[email protected]
 Table 5 Word segmentation experimental results with fusion of gap classification methodologies and distance metrics over IFN/ENIT handwritten
EDM CDM LTCM GMCM N M o2o gm2o go2m dm2o do2m DR RA FM
0 0 1 0 13,311 13,249 11,953 779 334 764 367 91.8 92.3 92.0
0 1 0 1 13,311 13,622 12,093 1082 206 454 514 93.2 90.5 91.8
1 0 0 1 13,311 13,655 12,190 1062 175 371 503 94.9 90.8 92.8
0 0 0 1 13,311 13,322 11,933 869 326 732 410 91.8 91.7 91.7

[email protected]
1 0 1 0 13,311 13,334 12,018 828 302 673 390 92.4 92.1 92.2
Efficient Algorithms for Text Lines and Words Segmentation …
399
400 A. A. A. Ali and M. Suresha

Table 6 Comparison Method Measures Accuracy (%)

between well-known methods
and proposed method [3] Word segmentation 85
[4] Text and word 84.8
segmentation
[8] Text line segmentation 95
Proposed work Text line segmentation 97.4
Proposed work Word segmentation 94.9

5 Conclusion

A methodology of Arabic handwritten documents segmentation which segments the

document into their distinct entities as lines and words of the text is presented. The
proposed approach is composed of (i) an improved method for the vertically con-
nected separation for the lines of text, and (ii) new method for word segmentation
which depends on an efficient recognition of intra-word and inter-word gaps using
combination of two various distance metrics which are convex and Euclidean dis-
tance metrics. The authors utilized the theory of Gaussian mixture for shaping the two
classes where distinguish between these two classes is treated as clustering problem
of an unsupervised. The authors show in the experimental results that the proposed
algorithms are more accurate than the existing state-of-the-art algorithms for seg-
menting of the line and word of the text in Arabic handwritten script. Basically in
the future work, the attention needs to be paid to improve the segmentation of word
using dots, touching lines, and punctuation detection method, and also to improve a
feedback from the modules of the character segmentation and character recognition.

References

1. Amani, A. A. A., & Suresha, M. (2017). A novel approach to correction of a skew at docu-
ment level using an Arabic script. International Journal of Computer Science and Information
Technologies, 8(5), 569–573.
2. Adiguzel, H., Sahin, E., & Duygulu, P. (2012). A hybrid for line segmentation in handwritten
documents. In International Conference on Frontiers in Handwriting Recognition (ICFHR),
(pp. 503–508), September 18–20, 2012.
3. AlKhateeb, J. H., Jiang, J., Ren, J., & Ipson, S. (2009). Interactive knowledge discovery for
baseline estimation and word segmentation in handwritten Arabic text. In M. A. Strangio (Ed.),
Recent Advances in Technologies.
4. Al-Dmour, A., & Fraij, F. (2014). Segmenting Arabic handwritten documents into text lines
and words. International Journal of Advancements in Computing Technology (IJACT), 6(3),
109–119.
5. Al-Maadeed, S., Elliman, D., & Higgins, C. A. (2004). A data base for Arabic handwritten text
recognition research. International Arab Journal of Information Technology, 1, 117–121.

[email protected]
Efficient Algorithms for Text Lines and Words Segmentation … 401

6. Al-Maadeed, S., Elliman, D., & Higgins, C. A. (2002). A database for Arabic handwritten
text recognition research. In Proceedings of the 8th International Workshop on Frontiers in
Handwriting Recognition (pp. 485–489).
7. Ataer, E., & Duygulu, P. (2006). Retrieval of Ottoman documents. In Proceedings of the
Eighth ACM SIGMM International Workshop on Multimedia Information Retrieval, 26–27,
Santa Barbara, CA, USA, October 2006.
8. Boussellaa, W., & Zahour, A., et al. (2010). Unsupervised block covering analysis for text-line
segmentation of Arabic ancient handwritten document images. In 20th International Confer-
ence on Pattern Recognition (ICPR). New York: IEEE.
9. Khandelwal, A., Choudhury, P., Sarkar, R., Basu, S., Nasipuri, M., & Das, N. (2009). Text line
segmentation for unconstrained handwritten document images using neighborhood connected
component analysis. In: Pattern Recognition and Machine Intelligence (pp. 369–374). Berlin:
Springer.
10. Khayyat, M., Lam., L., Suen, C. Y., Yin., F., & Liu, C. L. (2012, March). Arabic handwritten
text line extraction by applying an adaptive mask to morphological dilation. In 10th IAPR
International Workshop on Document Analysis Systems (DAS) (pp. 100–104). New York: IEEE.
11. Kumar, J., Abd-Almageed, W., Kang, L., & Doermann, D. (2010, June). Handwritten Arabic
text line segmentation using affinity propagation. In Proceedings of the 9th IAPR International
Workshop on Document Analysis Systems (pp. 135–142). New York: ACM.
12. Likforman-Sulem, L., Hanimyan, A., & Faure, C. (1995). A Hough based algorithm for extract-
ing text lines in handwritten documents. In Proceedings of the Third International Conference
on Document Analysis and Recognition (pp. 774–777), Montreal, Canada.
13. Louloudis, G., Gatos, B., & Pratikakis, I. (2008, August). Line and word segmentation of
handwritten documents. In International Conference on Frontiers in Handwriting Recognition
(ICFHR’08) (pp. 247–252), Montreal, Canada.
14. Louloudis, G., Halatsis, K., Gatos, B., & Pratikakis, I. (2006, October). A block-based Hough
transform mapping for text line detection in handwritten documents. In The 10th International
Workshop on Frontiers in Handwriting Recognition (IWFHR) (pp. 515–520), La Baule.
15. Mahadevan, U., & Nagabushnam, R. C. (1995). Gap metrics for word separation in hand-
written lines. In The Third International Conference on Document Analysis and Recognition
(pp. 124–127), Montreal, Canada.
16. Marin, J. M., Mengersen, K., & Robert, C. P. (2005). Bayesian modelling and inference on
mixtures of distributions, handbook of statistics (Vol. 25). Amsterdam: Elsevier-Sciences.
17. Pechwitz, M., Maddouri, S. S., Maergner, V., Ellouze, N., & Amiri, H. (2002). IFN/ENIT-
database of handwritten Arabic words. In Proceedings of CIFED (Vol. 2, pp. 127–136).
18. Phillips, I., & Chhabra, A. (1999). Empirical performance evaluation of graphics recognition
systems. IEEE Transaction of Pattern Analysis and Machine Intelligence, 21(9), 849–870.
19. Seni, G., & Cohen, E. (1994). External word segmentation of off-line handwritten text lines.
Pattern Recognition, 27(1), 41–52.
20. Shi, Z., & Govindaraju, V. (2004). Line separation for complex document images using fuzzy
run length. In First International Workshop on Document Image Analysis for Libraries (p. 306).
21. Vinciarelli, A., & Luettin, J. (2001). A new normalization technique for cursive handwritten
words. Pattern Recognition Letters, 2(9), 1043–1050.
22. Wshah, S., & Shi, Z., et al. (2009). Segmentation of Arabic handwriting based on both con-
tour and skeleton segmentation. In 10th International Conference on Document Analysis and
Recognition, ICDAR’09. New York: IEEE.

[email protected]
Privacy-Preserving Lightweight Image
Encryption in Mobile Cloud

M. Sankari and P. Ranjana

Abstract With the rapid development of mobile cloud, multimedia data like image
is a major challenge to maintain security and privacy for mobile users. Encryption is
one of the best solutions to protect data. However, traditional encryption algorithms
are not suitable for images which they were proposed for text data. In this paper, we
propose an image encryption technique called privacy-preserving lightweight image
encryption (PPLiIE) to make an encryption in a simple way, suitable for images and to
maintain privacy. PPLiIE algorithm proceeds with a three-step process to secure the
image data in mobile before storing to the cloud. We implement in Python language
and analyze our results with various file images to conclude that the encryption time
of the PPLiIE is reduced to 50% approximately than the encryption time of AES.
In addition, the measurement of key sensitivity and file with variation of chunk size
have expressed superior performance of PPLiIE. Finally, we review various security
attacks against PPLiIE to express the security level.

Keywords Mobile cloud · Data privacy · Image · AES · Mobile computing ·

Cloud computing

1 Introduction

Due to the increasing development of the mobile cloud, users have a lot of personal
information like photos, images, videos, audio that are stored in cloud [1]. They can
do all computation in cloud than mobile and they believe in third-party cloud vendors
to keep our data secure [2]. It reduces the usage of resources, memory, battery life
in mobile. However, untrusted third party in cloud may be possible to lose our data.
Therefore, maintaining security is one of the greatest challenges for moving mobile

M. Sankari (B) · P. Ranjana

Department of CSE, Hindustan Institute of Technology and Science, Chennai 600025, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 403

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_33

[email protected]
404 M. Sankari and P. Ranjana

data to the cloud. Encryption is the best technique, which is converting to cipher text
from plain text to secure data from unauthorized user. There are various encryption
algorithms developed to secure data in mobile such as AES [3], RSA [4], DES, 3-
DES, Blowfish. It is mainly suitable for text data than image data. In early stages,
mobile data are moved to cloud [5], where computation and encryption take over in
cloud. Later, due to the lack of privacy issues, partial data are encrypted and passed
to cloud [2]. Further, all data are encrypted in mobile before storing it in the cloud.
However, computational overhead is increased. So we can move onto the lightweight
image encryption [6, 7] to reduce the computational overhead and maintain privacy.
The major drawback of encryption is it takes longer time to execute the code. In
addition, various drawbacks are outlined and the following encryption techniques
are used to secure data.
The security mechanism uses RSA algorithm [4] to secure the data by the process
of encryption and decryption. Encryption is used to provide data security during
transmission from mobile to cloud. Users can believe their data are safe and secure
because of storing encrypted file to the cloud. The encrypted file is transferred over
the channel, where data leakage is reduced. Being unknown of data owner’s key, third
person/unauthorized user cannot access the encrypted file. Attribute-based encryp-
tion mechanisms [8] encrypts/decrypts the data based on the user’s attribute. It defines
access policies for other users to utilize data and protect the private data. Encrypted
file is accessed by only those authorized users that satisfy the access policies. A
(M-HABE) modified hierarchical Attribute-Based Encryption [9] method develops
the three-layered structure to provide better security. It is developed to ensure the
users with legal authorities. The user can satisfy the requirements of the algorithm to
access the ciphertext. The three major roles of the methods are data process, storage,
and access.
Biometric authentication [10] is another authentication technique, which is used
to protect the access of unauthorized users and database. It examines with the Sam-
sung Galaxy S3, BlackBerry Z to define the process time for an image taken (Maxi.
Time and Avg. Time are calculated). In [3], the author explains to protect the user
data and the retrieval keywords by existing ciphertext mechanism and ORAM (obliv-
ious) is introduced to protect the privacy security of the cipher access patterns. Also,
the performance of ciphertext retrieval scheme is improved than other security-level
schemes. In [11], the author has explained the solution of privacy-preserving data
storage by integrating PRF-based key management for unlinkability, a secure index-
ing method for privacy-preserving keyword search and access pattern hiding scheme
based on redundancy. It is implemented in private cloud to build privacy for mobile
health systems. It provides auditability for misusing health data. In [12], the author
proposed a secure and efficient index to locate the users with the huge amount of
encrypted image and applied two encryption approaches for reproducing the secure
image in the cloud. This approach produces 90% bandwidth consumption in the
mobile. Table 1 represents the traditional data privacy techniques available in mobile
cloud.

[email protected]
Privacy-Preserving Lightweight Image Encryption in Mobile Cloud 405

Table 1 Existing cryptography data privacy techniques in cloud and mobile cloud
Works Technique/s Encryption Security Issues
used enforcement
An efficient and secure RSA algorithm Yes Confidentiality Utilizes more
data storage in mobile and Hash Data privacy CPU resources,
cloud computing function high memory for
through RSA and Hash Authentication encryption
function [4]
Unauthorized
Data integrity user penetrates
when key is
known
Securing mobile cloud ABE-attribute- Yes Authentication Encryption based
data with personalized based encryption Data confi- only on the
attribute-based meta dentiality attribute
information [8]
Availability Randomized
Data privacy keys only can
find the
decryption
A modified hierarchical M-HABE Model Yes Data privacy Attack of any
attribute-based Authentication layers affect the
encryption access users data
control through
encryption and
Method for mobile cloud Confidentiality
decryption
computing [9]
Securing Mobile Cloud Biometric N/A Authentication High response
Computing Using authentication Data privacy time longer time
Biometric needed with
Authentication database
(SMCBA) [10] comparison
The cloud storage AES, ORAM Yes Data privacy Hierarchical
ciphertext retrieval Data ORAM takes
scheme based on ORAM availability extra time for
[3] execution
Cloud-assisted ABE, search Yes Auditability High memory
mobile-access of health index method, Data storage power required
data with privacy and PRP privacy for key
auditability [11] management and
decryption
Harnessing encrypted Homomorphic Yes Data privacy Increasing
data in cloud for secure Encryption (HE), process time for
and efficient mobile symmetric image storage in
image sharing [12] encryption and cloud
image
reproduction

[email protected]
406 M. Sankari and P. Ranjana

Here, most of the papers focused on the security and privacy plays a major role
for the mobile user. It referred many techniques such as RSA, ABE, Biometric, HE,
AES, PRP, and so on. RSA is used for double encryption, ABE is used for encryption
based on attribute/s, Biometric is used for user’s authentication, HE is used for image
sharing with assuring privacy, AES for maintaining data privacy by encryption, PRP
is used for permutation-based encryption [13] especially for lightweight encryption.
Mobile complexity acts as the great issue for all of these methods and encryption
takes longer time to execute in mobile. Therefore, we propose an image encryp-
tion technique (PPLiIE) [14, 15] to secure data and reduce complexity, improve the
throughput to maintain privacy before storing into the cloud by overcoming the tradi-
tional techniques. It is implemented in Python where Python is a scripting, high-level
language, and easily understandable code. It is an open source and image library is
included such as PyCrypto encryption and decryption library for execution. It is very
suitable for image processing.
The remaining sections are as follows: In Sect. 2, we present the proposed PPLiIE
method deeply with processes and its schema. In Sect. 3, we implement the PPLiIE
method with various metrics in Python language. In Sect. 4, we define the decryption
process of the PPLiIE. In Sect. 5, we describe the security analysis against the PPLiIE
method. Finally, we conclude the implemented method.

2 Proposed PPLiIE Method

The PPLiIE method defines the image split into chunks, distribute to different files
based on pattern and scramble based on file key to maintain the user’s data privacy
rather than traditional technique such as AES. Consider the JPEG image probably
used in mobile and is split as chunks (like 2 bytes as chunk). Chunks is grouped by
pattern and make as files. Key K formed by distributed file and stored in Database.
Finally, we scrambled the files and store it in cloud.

2.1 Layout of the PPLiIE Schema

See Fig. 1.

[email protected]
Privacy-Preserving Lightweight Image Encryption in Mobile Cloud 407

Fig. 1 Layout of the PPLiIE

schema

2.2 Basic Three Steps Required for the PPLiIE Method

2.2.1 Split

The original image file is split into two divisions such as the header and the content
of the image file. Header file contains file type, size, chunk size, date created, width,
height, and resolution. The equation used for a split file is
kmax
Image_filei  Hi + Ci,k (1)
k1

where H i as a header of the image file, C i,k as the number of chunks formed.
Max  [(File_Sizei /Chunk_sizei ) − H_Size] (2)
where File_Sizei denotes the size of the file, Chunk_sizei denotes the size of chunks,
and H_size denotes the header’s size of the original file. All are represented in bytes.

[email protected]
408 M. Sankari and P. Ranjana

File [1,1](odd chunks) File[2,1](Even chunks)

Header Header

File[1,2](odd) File[2,2](Even)
1 3 5 7 2 4 6 8

9 11 13 15 10 12 14 16

17 19 21 23… 18 20 22 24..

Fig. 2 Example of pattern for the PPLiIE method

2.2.2 Distribute (Pattern)

After splitting the image, chunks are grouped into the different files based on the
pattern. A user can act a pattern as a key or it can be selected randomly by a predefined
method. Consider as an example, the proposed PPLiIE method takes a pattern as odd
chunks act as file1, even chunks act as a file2 and continues (Fig. 2).

2.2.3 Scramble

After completion of the distribution of chunks to files, scramble the file within it by
adding Key K1 (first row of the file1) with all rows of the files. Key K i stored in the
database.

K i  Filei,2 (Row1 ) (3)

where Filei,2 represents the distribute file split by pattern, Row1 represents the first
row (first two bytes) of the Filei,2 .


2 
m
 
Filei,2  K i + Filei,2 Row j (4)
i1 j1

where n denotes the number of files split by pattern, Filei,2 denotes the ith file split
by pattern, Row j denotes the jth row of the Filei,2 , m denotes the number of rows in
the Filei,2 .

[email protected]
Privacy-Preserving Lightweight Image Encryption in Mobile Cloud 409

2.3 Algorithm

Start
Input img←”*.jpeg”
Image filei ←bin (Img)
for i←0 to n-1 do
for k←0 to max do
k
max
Image_filei  Hi + Ci,k (Split)
k1
[Header Hi is stored in Database]
where Max=[(File_Sizei /Chunk_sizei )-H_Size]
End for
End for

File1,2 ←Collection of even chunks(Ci.k[even] )

File2,2 ←Collection of odd chunks(Ci,k[odd] )

for i←1 to 2
K i =Filei,2 (Row 1 )
Ki →Database
for j←0 to m-1
Filei,2  Ki + Filei,2 (Row1)
End for
End for
File→cloud storage(Encry_Image)
End

3 Implementation

We have implemented the proposed PPLiIE method in Python language and used
hardware as Windows 7 OS. Consider the size of the chunk as 64, key size of AES
as 16 bytes, key size of proposed PPLiIE method as the first row of the file1 and

Table 2 Various file size Image name Pixel range File size Encryption algorithm
images with execution time [Execution Time
(ms) (ms)]
AES PPLiIE
Cat 244*206 4.52 4000 1000
Penguin 225*225 5.669 4000 1000
Flower 259*194 7.85 4000 2000
Lena 225*225 8.1 4000 2000
Bear 318*159 10.7 5000 2000

[email protected]
410 M. Sankari and P. Ranjana

6000 AES
PPLiIE
5000
Encryption Time(ms)

4000

3000

2000

1000

0
4.52 5.6 7.85 8.1 10.7
Size of the image file(KB)

Fig. 3 A comparison of the proposed PPLiIE method against AES encryption with various file
image sizes

pattern taken as even and odd chunks in Fig. 2. Table 2 represents the various file
size images with execution time by using AES and PPliIE.
We analyzed with the sample jpeg images and concluded that the encryption
time of the proposed PPLiIE method is reduced approximately to 50% which is
represented in a line chart compared with AES encryption in Fig. 3.

3.1 Key Sensitivity

Key sensitivity is an important factor to execute an algorithm. Large key maintains

more security for image and small key reduces the resource used in mobile. AES
have three long keys such as 16, 24, 32 bytes to encrypt the data. PPLiIE method has
taken the key automatically from the first row of the first file (K) in an image. There
is no manual requirement to assign the key. To ensure privacy, storing key in user’s
mobile and never pass it to cloud.

3.2 File with Chunk Size Measurement

In the proposed method, chunk size is used for splitting the image file. The size of
the chunk in AES sets in which the function uses to read and encrypt the image
file. When the size of the chunk is increased, encryption time is also increased in
AES. However, in our proposed method, the encryption time is relatively constant
even with the change of chunk size. It proves to have the better performance such as

[email protected]
Privacy-Preserving Lightweight Image Encryption in Mobile Cloud 411

80000
AES
70000 PPLiIE
Encryption time(ms)

60000

50000

40000

30000

20000

10000

0
2 8 16 32 64
Chunk size(Bytes)

Fig. 4 Encryption time varied with chunk size

speed, the limited resource used, lightweight process for all chunk sizes. Consider
the image .jpeg file as bear with file size as 10.7 MB and variation of chunk size are
taken for the line chart. Figure 4 represents the execution time of the encryption in
AES and PPLiIE method with various chunk sizes.
Our result of the conclusion decided that the lower encryption time in PPLiIE
with the variation of chunk size than AES. Therefore, it have superior performance
and reduced the overall execution time.

4 Decryption Process

We can retrieve the encrypted image data from cloud to mobile. Using the key stored
in mobile, subtract them with the file1 and file2. We can merge the files based on
pattern. Chunks of image data are connected based on chunk size with reference
of the header file. Binary code images are converted to jpeg image. Now, image is
assembled from the disassembled image. Mobile user retrieved the image from the
cloud.

5 Security Analysis

We have stored our image-related data such as key, metadata of the header file,
pattern, and chunk size in mobile to maintain privacy. Assume that the unauthorized

[email protected]
412 M. Sankari and P. Ranjana

user is gathering the image encryption data in the cloud. In addition, it is impossible
to retrieve the original image due to the lack of key, pattern and the chunk size.
We present various scenarios of security attacks in the proposed method. An
attacker mainly requires key k1, scramble method, pattern, header information, and
the size of the chunks to retrieve the original image data. Consider an image as 10*10
bytes, chunk size as 2 bytes.
Scenario 1:

Assume the key K i is known, but the attacker does not know the pattern and
the size of the chunks.

If the attacker knows the key K i , he retrieves the file1 and file2. But he cannot know the
scrambled method. So, he would not move to the distribute process (pattern). Assume
he may apply brute force attack to know the scrambled method. The minimum
probability of checking with known key are 12!  479,001,600 possibilities needed
to retrieve the scrambled data.
Scenario 2:

Assume that the key K i and the scrambled method are known, he does not
known the pattern of file.

An attacker retrieves the file1 and file2. But file1 has approximately 13 parts and
file2 has 12 parts. Assume that the attacker checks with consequent pattern, odd/even
pattern, and random pattern. Consequent pattern such as 3 or 4 or 5 are taken. For
example, first 3 chunks move to file1 next 3 moves to file2 and so on. There are
9 possibilities for 3 chunks taken, 7 possibilities for 4 chunks, 5 possibilities for 3
chunks, and so on. The total probabilities are 9 + 7 + 3  19. The same procedure for
odd/even pattern that is first odd in first1 and second odd in file2 and so on. Randomly
selected pattern are impossible to retrieve it.
Scenario 3:

Assume the key k1, the scramble method and the pattern are known, unknown
the size of the chunks.

An attacker can check the chunks with different sizes such as 2, 3, 4, 5, 6 bytes. 25
chunks formed by 2 bytes 20 by 3, 4 by 15, 5 by 10, 6 by 9. The minimum total
possibilities for identifying chunk size are 25 + 20 + 15 + 10 + 9  79. Even the
chunk size may have 7, 8, 9 and so on. So the attacker is difficult to retrieve the chunk
size.

[email protected]
Privacy-Preserving Lightweight Image Encryption in Mobile Cloud 413

6 Conclusion

The proposed PPLiIE method secured the image data by simple lightweight encryp-
tion and to maintain user’s privacy in mobile. It also provides a clear, effective
performance, data privacy solution for mobile devices which have limited resources.
We have taken the sample images for performance analysis, which is implemented
in Python language. It is proved to reduce more than half of the encryption time
compared to existing method like AES. Finally, we have introduced security attack
scenarios to explain the difficulty to access the proposed method from attackers.

References

1. Zhang, X., Schiffman, J., Gibbs, S., Kunjithapatham, A., & Jeong, S. (2009, November).
Securing elastic applications on mobile devices for cloud computing. In Proceedings of ACM
Workshop on Cloud Computing Security, CCSW ’09, Chicago, IL, USA.
2. Mollah, M. B., Azad, M. A. K., & Vasilakos, A. (2017). Security and privacy challenges in
mobile cloud computing: Survey and way ahead. Journal of Network and Computer Applica-
tions, pp. 38–54.
3. Song, N., & Sun, Y. (2014). The cloud storage ciphertext retrieval scheme based on ORAM.
China Communications, 11(14), 156–165. (IEEE Journals & Magazines).
4. Garg, P., & Sharma, V. (2014). An efficient and secure data storage in mobile cloud computing
through RSA and Hash function. In 2014 International Conference on Issues and Challenges
in Intelligent Computing Techniques (ICICT). New York: IEEE Society.
5. Qin, Z., Weng, J., Cui, Y., & Ren, K. (2018). Privacy-preserving image processing in the cloud.
IEEE Cloud Computing, 99 (IEEE Early Access Articles).
6. Bahrami, M., Li, D., Singhal, M., & Kundu, A. (2016). An efficient parallel implementation
of a light-weight data privacy method for mobile cloud users. In 2016 Seventh International
Workshop on Data-Intensive Computing in the Clouds (DataCloud) (pp. 51–58).
7. Bahrami, M., & Singhal, M. (2016). CloudPDB: A light-weight data privacy schema for cloud-
based databases. In 2016 International Conference on Computing, Networking and Communi-
cations, Cloud Computing and Big Data.
8. Zickau, S., Beierle, F., & Denisow, I. (2015). Securing mobile cloud data with personalized
attribute-based meta information. In 2015 3rd IEEE International Conference on Mobile Cloud
Computing, Services, and Engineering (MobileCloud).
9. Xie, Y., Wen, H., Wu, B., Jiang, Y., & Meng, J. (2014). A modified hierarchical attribute-based
encryption access control method for mobile cloud computing. IEEE Transaction on cloud
Computing.
10. Rassan, A., & AlShaher, H. (2014). Securing mobile cloud computing using biometric authen-
tication (SMCBA). In 2014 International Conference on Computational Science and Compu-
tational Intelligence (CSCI).
11. Tong, Y., Sun, J., Chow, S. S. M., & Li, P. (2014, March). Cloud-assisted mobile-access of
health data with privacy and auditability. IEEE Journal of Biomedical and Health Informatics,
18(2).
12. Cui, H., Yuan, X., & Wang, C. (2017). Harnessing encrypted data in cloud for secure and
efficient mobile image sharing. IEEE Transactions on Mobile Computing, 16(5).
13. Bahrami, M., & Singhal, M. (2015). A light weight permutation based method for data privacy
in mobile cloud computing. In 2015 3rd IEEE International Conference on Mobile Cloud
Computing, Services, and Engineering (MobileCloud).

[email protected]
414 M. Sankari and P. Ranjana

14. Bahrami, M., Khan, A., & Singhal, M. (2016). An energy efficient data privacy scheme for
IoT devices in mobile cloud computing. In 2016 IEEE International Conference on Mobile
Services.
15. Balouch, Z. A., Aslam, M. I., & Ahmed, I. (2017). Energy efficient image encryption algorithm.
In 2017 International Conference on Innovations in Electrical Engineering and Computational
Technologies (ICIEECT).

[email protected]
Performance Evaluation
of Ensemble-Based Machine Learning
Techniques for Prediction of Chronic
Kidney Disease

K. M. Zubair Hasan and Md. Zahid Hasan

Abstract Chronic kidney disease (CKD) is widespread and related with enhanced
risk of cardiovascular disease and end-stage renal disease, which are possibly
escapable through early detection and treatment of individuals at risk. Machine learn-
ing algorithm helps medical experts to diagnose the disease correctly in the earlier
stage. Therefore, machine-predicted analysis has become very popular in recent
decades that can efficiently recognize whether a patient has certain kidney disease or
not. In this regard, we propose an ensemble method based classifier to improve the
decision of the classifiers for kidney disease diagnosis efficiently. Ensemble meth-
ods combine multiple learning algorithms to achieve better predictive performance
than could be obtained from any of the constituent learning algorithms. In addition,
Data is evaluated by using tenfold cross-validation and performance of the system is
assessed on receiver operative characteristic curve. Extensive experiments on CKD
datasets from the UCI machine learning repository show that our ensemble-based
model achieves the state-of-the-art performance.

Keywords Machine learning · Classification · Chronic kidney disease (CKD) ·

Ensemble method · Data mining · Healthcare informatics

1 Introduction

According to the report of the Centers for Disease Control and Prevention (CDC),
kidney disease causes millions to die each year. A good number of people in the
world with kidney damage and slightly reduced kidney function are not conscious of
having CKD. Therefore, strategies for early detection and cure of people with CKD
are consequently required worldwide. A computer-aided diagnosis system based on
sophisticated machine learning techniques is required for healthcare data to mine

K. M. Zubair Hasan · Md. Zahid Hasan (B)

Daffodil International University, 102 Sukrabadh, Mirpur Road, Dhaka 1207, Bangladesh
e-mail: [email protected]
K. M. Zubair Hasan
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2019 415
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_34

[email protected]
416 K. M. Zubair Hasan and Md. Zahid Hasan

hidden pattern from data for effective decision-making. The purpose of our current
study is to evaluate the performance of several ensemble-based machine learning
techniques correctly classifying CKD patients based on clinical datasets. We consider
five machine learning classifiers, namely Adaptive Boosting, Bootstrap Aggregat-
ing, Extra Trees, Gradient Boosting, and Random Forest Classifier. Finally, a large
set of standard performance metrics is used to design the computer-aided diagnosis
system for estimating the performance of each machine learning and artificial intel-
ligence classifier. The metrics we used include Classification Accuracy, Sensitivity,
Precision, Specificity, Negative Predictive Value, False Positive Rate, False Negative
Rate, F1-Score, and Error Rate of Classification.
The paper is organized as follows. Section 2 discusses the related works. Section 3
introduces the materials and methods. Section 4 describes the proposed ensemble-
based machine learning techniques applied to the dataset. Section 5 presents the
classification performance matrices. These comparisons’ study is discussed in Sect. 6.
Finally, we concluded our work in Sect. 7.

2 Related Work

Ho et al. [1] presented a computer-aid diagnosis tool for CKD classification analyzing
ultrasonography images. This system used for detecting and classifying distinctive
various stages of CKD. The K-means clustering was performed for detecting regions
in an ultrasonic image as preprocessing step. Estudillo-Valderrama et al. [2] proposed
the feasibility study of using a distributed system for the management of alarms from
chronic kidney disease patients inside Enefro project. Charleonnan et al. [3], explored
four machine learning approaches including K-nearest neighbors (KNN), support
vector machine (SVM), logistic regression (LR), and decision tree classifiers for
predicting the chronic kidney disease. In Hsieh et al. [4] showed a real-time system
to consider chronic kidney disease by using only ultrasound images. Support vector
machine technique is used to predict and classify CKD stages form the ultrasound
images. The authors in [5] recommended diverse methods to influence the hierarchi-
cal structure in ICD-9 codes to develop the performance of predictive models. A novel
feature engineering approach is proposed in this study to influence this hierarchy,
while immediately diminishing feature dimensionality. An intelligent system based
on artificial neural networks in Chiu et al. [6] for detecting chronic kidney disease
for evaluating the extremity of a patient. Three types of artificial neural networks
have been used in this model including backpropagation network (BPN), modular
neural network (MNN), and generalized feedforward neural networks (GRNN).

[email protected]
Performance Evaluation of Ensemble-Based … 417

3 Materials and Methods

Our research uses a CKD dataset [7], which is openly accessible at UCI machine
learning laboratory. This dataset comprises 400 instances with 150 samples with-
out kidney disease (not presence) and 250 samples with kidney disease (presence).
Missing values in the dataset may effect in the accuracy of disease prediction. So
to avoid this, we consider 158 without missing samples and then apply Extra Tress
algorithm for evaluating feature importance. After getting feature importance, we
compute impute (mean) for missing values of numerical type attributes for increas-
ing the number of samples as preprocessing steps. Thenceforth, for predicting kidney

Hypertension 0.180108
Albumin 0.155358
Specific Gravity 0.095088
Diabetes Mellitus 0.093597
Pus Cell 0.087868
Packed Cell Volume 0.051738
Hemoglobin 0.042891
Serum CreaƟnine 0.042749
Red Blood Cells 0.040695
Pedal Edema 0.036521
AppeƟte 0.03546
Blood Urea 0.01966
Blood Glucose Random 0.019166
Red Blood Cell Count 0.018489
Anemia 0.017574
Pus Cell Clumps 0.012127
Sodium 0.010727
Sugar 0.009852
Coronary Heart Disease 0.008114
Bacteria 0.007844
White Blood Cell Count 0.007043
Blood Pressure 0.006082
Potassium 0.000856
Age 0.000392
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
Feature Importance

Fig. 1 Importance of attributes in CKD datasets

[email protected]
418 K. M. Zubair Hasan and Md. Zahid Hasan

diseases, 13 distinct parameters (Hypertension, Albumin, Specific Gravity, Diabetes

Mellitus, Pus Cell, Packed Cell Volume, Haemoglobin, Serum Creatinine, Red Blood
Cells, Pedal Edema, Appetite, Blood Urea, and Blood Glucose Random) have been
taken out of 24 attributes into account considering the feature importance. Figure 1
shows the importance of attributes in CKD dataset.

4 Machine Learning Techniques

4.1 Adaptive Boosting (AdaBoost)

Adaptive Boosting is an ensemble machine learning meta-algorithm technique that

creates a robust classifier from a number of weak classifiers. For improving perfor-
mance, it incorporated extra copies of the classifier on the same dataset, however
where the weights of incorrectly classified samples are adjusted and adjusts them to
represents the final output of the boosted classifier.
Given M training data {(x1 , y1 ), . . . , (x M , y M )}, xi is a vector corresponding to
an input sample data, associated with Q input attributes, and yi is a target variable
with a class label of either −1 or +1  a set of weak learners {K 1 , . . . , K L } each of
 and
which outputs a classification K j x j ∈ {−1, 1} for each classifier. An initial weight
is set for 1/n. Equal weight is assigned to all the instances in the kidney datasets.
AdaBoost is explained using pseudocode [8].
For t in 1, …, T :
• Choose h t (x) → [−1, 1]:
– Find weak learners that minimize ∈t , the summation of errors in weights for
misclassified points


n
∈t  wi,t
i1
h t (xi  yi )

 
– Choose αt  21 ln 1−∈
∈t
t

– Add to ensemble Ft (x) + αt h t (x)

– Update weights for i  1, …, m using

wt (i)e−αt yi h t (xi ))
wt+1 (i) 
Zt

Where Z t is a normalization factor.

– Final hypothesis:

[email protected]
Performance Evaluation of Ensemble-Based … 419
 T 

H (x)  sign αt h t (x)
t1

H is determined as a weight majority vote of the weak hypothesis h t , where each

classifier is assigned weight αt .

4.2 Bootstrap Aggregating (Bagging)

Bagging is another ensemble meta-algorithm used in statistical classification for

reducing variance and supports to avoid overfitting. At first, an initial weight is set
for kidney datasets [9].
For every classifier as T rounds:
• Normalize the weights
• Train the classifier and evaluate training error
• Choose: lowest error classifier
• Update the weights of the training data
Then, the final classifier is formed as the linear combination of T classifiers.

4.3 Extra Trees Classifier

Extra Trees is another modification of bagging classifier with ordinary tree-based

techniques as far as accuracy and computational efficiency. The main differences
with other tree-based algorithm are that it can split the node by choosing cut-points
randomly and build the trees using the total learning samples [10].
Split_node(S)
Input: Spited node equivalent to local subset S equivalent
Output: a split where ac is greater than a or not anything
• If End_split(S) is true then it will not return anything
• Or else, Choose k attributes {a1 , . . . , ak } among all nonconstant (in S) candidate
attributes.
• Draw k splits {s1 , . . . , sk }, where si  Choice_random_split(s, ai ), ∀i  1, . . . , k;
• Return a split s∗ such that s will be selected from the maximum score of S.
Choice_random_split(S,a)
Input: Local subset S and an attribute a.
Output: a split
• Draw a random cut-point aC considering maximal attributes of subset S and min-
imal attributes of subset S
• Return the split based on [a < aC ]

[email protected]
420 K. M. Zubair Hasan and Md. Zahid Hasan

End_split(S)
Input: Local subset S
Output: a Boolean value
• If the absolute value of S is less than the minimum sample size of the splitting
node, then return TRUE;
• If every property is consistent at that point then return TRUE;
• If the outcome is steady in S, then return TRUE;
• Or else, return FALSE.

4.4 Gradient Boosting

Gradient boosting is an optimization machine learning algorithm on a suitable cost

function for the prediction of kidney disease classification problem. In the process
of gradient boosting, a series of predictor values are iteratively formed. The final
predictor value is generated after iteratively calculating the weighted average values
of the weak predictor. At every sequence, an extra classifier is invoked to boost the
performance of the complete ensemble [11]. The algorithm for the predictive model
is enlisted in Table 2.
Input: A training set of data points (xi , yi ) from the given dataset
Output: A classification tree
• Initialization: Initialize model with a constant low value of T classifier.
• For i to number of iterations M do
• Calculate new weights of (x0 , y0 ) to (xi , yi ) with minimal prediction accuracy rate
of T .
• Draw new weak classifier h i on the pre-weighted prediction accuracy of T .
• Calculate weight βi of new classifier.
• Update the model as h i + βi .
• end for
• return T
Gradient boosting adds weak learners to minimize the bias in addition to the
variance to some degree, thus reducing the error [12].

4.5 Random Forest

Random Forest (RF) is a variant of ensemble classifier consisting of a collection

of tree-structured classifiers h(x, yk ), which is defined as multiple tree predictors
yk such that each tree relies upon the estimations of an arbitrary vector inspected
independently and with a similar distribution for all trees in the forest. The random-
ization works in two ways: an arbitrary sampling of data for bootstrap samples as it

[email protected]
Performance Evaluation of Ensemble-Based … 421

is done in Bootstrap aggregating (bagging) in Sect. 4.2 and random selection of input
attributes for producing individual base decision trees [13]. Random forests become
different in a way from other methods that a modified tree learning algorithm is uti-
lized that chooses the differentiable candidate in the learning procedure, a random
subset of the features. The cause for doing this is the relationship of the trees in a
standard bootstrap sample. For example, if one or a couple of features are extreme
indicators for the response variable (target output), these features will be chosen in
a considerable lot of the decision trees, reasoning them to end up correlated.

5 Classification Performance Measurements

Five ensemble-based supervised machine learning techniques were applied for the
classification of chronic kidney disease attributes. Classification performance was
estimated by tenfold cross-validation. The classification models were assessed on the
nine quality measures [14]. These quality measures for the analysis of classification
were examined closely. Absence attributes of chronic kidney disease were measured
as non-ckd class, and attributes with the presence of chronic kidney disease were
measured as ckd class. Here,
True positive (TP)—Correct positive prediction where samples with ckd predicted
as ckd.
False positive (FP)—Incorrect positive prediction where samples with non-ckd pre-
dicted as ckd.
True negative (TN)—Correct negative prediction where samples with non-ckd pre-
dicted as non-ckd.
False negative (FN)—Incorrect negative prediction where samples with ckd predicted
as non-ckd.
The quality measures are executed to evaluate the efficiency of each machine
learning classifier in the distinction between presence and absence with samples of
chronic kidney disease. They are expressed as follows:
(TN+TP)
Classification Accuracy  (TN+FP+FN+TP)
Sensitivity  (TP+FN)
TP

Precision  (TP+FP)
TP

Specificity  (TN+FP)
TN

Negative Predictive Value  (TN+FN)

TN

False Positive Rate  (FP+TN)

FP

False Negative Rate  (FN+TP)

FN

(FN+FP)
Rate of Misclassification  (TN+FP+FN+TP)
(Precision*Recall)
F1 Score  2 ∗ (Precision+Recall)

[email protected]
422 K. M. Zubair Hasan and Md. Zahid Hasan

6 Result and Discussion

The efficiency of five ensemble-based machine learning approaches was accounted
that all classifiers are trained with all 13 classes discussed in methods and materials
section, following a tenfold cross-validation to illustrate statistically robust results
for the prediction of kidney disease. We employed the Python software to execute
all classification tasks. Predicted results based on the confusion matrix is arranged
in Tables 1, 2, and 3 for Adaptive Boosting (AdaBoost), Bootstrap Aggregating
(Bagging), Extra Trees, Gradient Boosting, and Random Forest, respectively.
Figure 2 shows the classification results of these ensemble machine learning algo-
rithms. It is obvious from the outcomes that Bootstrap Aggregating (Bagging) pre-
dicts most noteworthy of true positives (presence of kidney disease classified as
presence) (Table 2; Fig. 2) and combined Adaptive Boosting (AdaBoost) and Extra
Trees algorithm predicts the most noticeable number of true negatives (absence of
kidney disease classified as absence of kidney disease) (Tables 1 and 3; Fig. 2).
Adaptive and gradient boosting confusion matrix (Tables 1 and 2) indicates that it
has second-best true positives (Fig. 2). Tables 2 and 3 show the confusion matrix of
Extra Trees and Random Forest classifier, which indicates that this classifier gives
the third most elevated number of true positives (Fig. 2).

Table 1 Confusion matrix for tenfold cross-validation using adaptive boosting and bootstrap aggre-
gating
Adaptive boosting Bootstrap aggregating
Predicted class Predicted class
ckd Non- Actual ckd Non- Actual
ckd total ckd total
Actual Actual
class class
ckd 61 1 62 ckd 62 0 62
Non-ckd 0 38 38 Non-ckd 4 34 38
Total 61 39 100 Total 66 34 100
predicted predicted

Table 2 Confusion matrix for tenfold cross-validation using extra trees and gradient boosting
Extra trees Gradient boosting
Predicted class Predicted class
ckd Non- Actual ckd Non- Actual
ckd total ckd total
Actual Actual
class class
ckd 60 2 62 ckd 61 1 62
Non-ckd 0 38 38 Non-ckd 2 36 38
Total 60 40 100 Total 63 37 100
predicted predicted

[email protected]
Performance Evaluation of Ensemble-Based … 423

Table 3 Confusion matrix for tenfold cross-validation using random forest

Random forest
Predictive class
CKD Non-CKD Actual total
Actual class
ckd 60 2 62
Non-ckd 3 35 38
Total predicted 63 37 100

Classifications Results using Machine Learning

70

60

50

40

30

20

10

0
TP FN FP TN
Adaptive Boosting 61 1 0 38
Bootstrap Aggregating 62 0 4 34
Extra Trees 60 2 0 38
Gradient Boosting 61 1 2 36
Random Forest 60 2 3 35

Fig. 2 Classification results of ensemble machine learning techniques

Figure 3 plots nine quality measures for classification performance measurements.

Figure 3 specifies that Adaptive Boosting beats over all other machine learning
algorithms with the most extreme classification accuracy of 99% while the second
most noteworthy characterization exactness is accomplished by Extra Trees 98%.
Furthermore, these two strategies have indicated the most extreme specificity and
precision of 100% that this classifier is most appropriate for distinguishing proof of
patients with chronic kidney disease (absence class). Bootstrap Aggregating achieves
the highest sensitivity, the negative predictive value of 100% and the lowest false
negative rate of zero percent, which indicates that this classifier is most appropriate
for identification of people who are sick with chronic kidney disease (presence class).
But, bagging also has the lowest specificity of 89% and false negative rate of 11% that
is not fit for prediction of sick people with chronic kidney disease (absence class).
Figure 3 associates F1 Score and Error Rate of Classification Adaptive Boosting
shows the highest 99% and the lowest 1%, respectively.

[email protected]
424 K. M. Zubair Hasan and Md. Zahid Hasan

Classification Performance Measurement

1.2
1
0.8
0.6
0.4
0.2
0
CA RMC SN SP PRC NPV FPR FNR F1
Adaptive Boosting 0.99 0.01 0.98 1 1 0.97 0 0.02 0.99
Bootstrap Aggregating 0.96 0.04 1 0.89 0.94 1 0.11 0 0.95
Extra Trees 0.98 0.02 0.97 1 1 0.95 0 0.03 0.98
Gradient Boosting 0.97 0.03 0.98 0.95 0.97 0.97 0.5 0.02 0.97
Random Forest 0.95 0.05 0.97 0.97 0.95 0.95 0.8 0.03 0.95

Fig. 3 Classification performance of ensemble machine learning techniques

Fig. 4 ROC for ensemble machine learning techniques

In agreement to the previously mentioned assessment criteria, a receiver operating

characteristic (ROC) curve [15] is used and the area under the curve (AUC) to evaluate
the advantage and disadvantage of the classifier. ROC is unbiased of the two classes
and significant when the number instances of the two classes change through training
[16]. The region under ROC curve must be close to 1 for the best classifier. Figure 4
shows that AdaBoost classifier beats every single other procedure in expectation of

[email protected]
Performance Evaluation of Ensemble-Based … 425

quality of kidney ailment and other ensemble-based learning techniques perform the
approximate similar result in the prediction of the essence of kidney disease.

7 Conclusion

This prediction of kidney disease may spare the life of people and can have a real
effect on its cure. In this paper, we propose an ensemble method based machine
learning algorithm to improve the performance of the classifier for kidney disease.
In contrast with quite a few classic prediction algorithms, the classification accuracy
of our proposed ensemble learning algorithm achieves 99% classification accuracy.

References

1. Ho, C.-Y., Pai, T.-W., & Peng, Y.-C. Ultrasonography image analysis for detection and classi-
fication of chronic kidney disease. In Sixth International Conference on Complex, Intelligent,
and Software Intensive Systems (pp. 624–629), Palermo.
2. Estudillo-Valderrama, M. A., et al. (2014, November). A distributed approach to alarm man-
agement in chronic kidney disease. IEEE Journal of Biomedical and Health Informatics, 18(6),
1796–1803.
3. Charleonnan, A., Fufaung, T., Niyomwong, T., Chokchueypattanakit, W., Suwanna-wach, S., &
Ninchawee, N. (2016). Predictive analytics for chronic kidney disease using machine learning
techniques. In Management and Innovation Technology International Conference (MITicon)
(pp. 80–83), Bang-San.
4. Hsieh, J.-W., Lee, C.-H., Chen, Y.-C., Lee, W.-S., & Chiang, H.-F. (2014). Stage classification
in chronic kidney disease by ultrasound image. In International Conference on Image and
Vision Computing, New Zealand, ACM, 2014 (pp. 271–276).
5. Singh, A., Nadkarni, G., Guttag, J., & Bottinger, E. (2014). Leveraging hierarchy in medical
codes for predictive modeling. In Proceedings of the 5th ACM Conference on Bioinformatics,
Computational Biology, and Health Informatics (BCB ‘14) (pp. 96–103). ACM, New York,
NY, USA, 2014.
6. Chiu, R. K., Chen, R. Y., Wang, S.-A., & Jian, S.-J. (2012). Intelligent systems on the cloud
for the early detection of chronic kidney disease. In International Conference on Machine
Learning and Cybernetics (pp. 1737–1742), Xian, China.
7. Asuncion, A., & Newman, D. J. UCI Machine Learning Repository [Online]. Available http://
www.ics.uci.edu/~mlearn/MLRepository.html. (2007).
8. Rojas, R. (2009). AdaBoost and the super bowl of classifiers a tutorial introduction to adaptive
boosting. Freie University, Berlin, Technical Report.
9. Breiman, L. (1996). Bagging predictors. Machine Learning, 24(2), 123–140.
10. Geurts, P., Ernst, D., & Wehenkel, L. (2006). Extremely randomized trees. Machine Learning,
63(1), 3–42.
11. Manna, S., Biswas, S., Kundu, R., Rakshit, S., Gupta, P., & Barman, S. (2017). A statistical
approach to predict flight delay using gradient boosted decision tree. In International Confer-
ence on Computational Intelligence in Data Science (ICCIDS) (pp. 1–5), Chennai, India.
12. Friedman, J. H. (2002). Stochastic gradient boosting. Computational Statistics & Data Analysis,
38(4), 367–378.

[email protected]
426 K. M. Zubair Hasan and Md. Zahid Hasan

13. Anbarasi, M. S., & Janani, V. (2017). Ensemble classifier with Random Forest algorithm to deal
with imbalanced healthcare data. In International Conference on Information Communication
and Embedded Systems (ICICES) (pp. 1–7), Chennai.
14. Kanmani, S., Uthariaraj, V. R., Sankaranarayanan, V., & Thambidurai, P. (2007). Object-
oriented software fault prediction using neural networks. Information and Software Technology,
49(5), 483–492.
15. Metz, C. E. (1978). Basic principles of ROC analysis. In L. M. Freeman (Ed.), Seminars in
nuclear medicine (Vol. 4, pp. 283–298). Amsterdam: Elsevier.
16. Dwivedi, A. K. (2016). Performance evaluation of different machine learning techniques for
prediction of heart disease. Neural Computing and Applications, pp. 1–9.

[email protected]
Proficient Cooperative Caching
in SWNET Using Twin Segments
Approach

B. N. Lakshmi Narayan, Prasad N. Hamsavath, Meher Taj, E. G. Satish,

Vivek Bharadwaj and S. Rabendranath

Abstract Web caching is the main practice to scale the Internet. One chief perfor-
mance aspect of web caches is the replacement strategy. The most prominent system
to let an array of distributed caches to cooperate and serve each other using web
requests is cooperative caching. In the existing cooperative caching schemes, most
of them will not provide high network level availability and high node level avail-
ability at the dupe time. The feasibility of the object is ensured by network level in
isolated network partitions and node level in individual nodes when they are com-
pletely detached from all of the networks. To reduce electronic content provisioning
cost in Social Wireless Network (SWNET), it is recommended to use cooperative
caching policies. Social Wireless Networks are designed by mobile devices like
Android, iPhone, Kindle, etc., giving out similar interests in electronic information,
and physically congregating in public areas. In Social Wireless Network, object
caching proves that it reduces the content provisioning cost, which is massively
dependent on the pricing factors within several stakeholders like End Consumers,
Content Service Providers, and Network Service Providers.

Keywords SWNET · Twofold · Ad hoc network

B. N. Lakshmi Narayan (B) · P. N. Hamsavath

Department of MCA, NMIT, Bengaluru, India
e-mail: [email protected]
P. N. Hamsavath
e-mail: [email protected]
M. Taj
Department of Computer Science, GSSS FGC, Mysore, India
E. G. Satish
Department of CSE, NMIT, Bangalore, India
V. Bharadwaj · S. Rabendranath
Department of CSE, AMCEC, Bengaluru, India
e-mail: [email protected]
S. Rabendranath
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 427

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_35

[email protected]
428 B. N. Lakshmi Narayan et al.

1 Introduction

The establishment of mobile ad hoc networks (MANETs) is done with mobile hosts
(MHs), such as notebook, PDAs, and so on. Without the assistance of any network
infrastructure, these mobile devices can form a wireless network dynamically. The
use of multi-hop wireless links helps to move every MH arbitrarily and communicate
with one another. Within the scope of transmission area, the MH which acts as a router,
forwards the data packets to other neighbors. MANET is very useful under certain
environments, such as battles, disaster rescue, earthquake recovery, etc. SWNETs
can be framed with the help of ad hoc wireless connections between the devices. In
the existence of such SWNETs, a distinct way for content access by a device would
be to first look into the local SWNET for the requested content before downloading
from the Content Service Provider’s server. The expected content provisioning cost
of such an approach can be considered lesser, because the download from Content
Service Provider would be avoided since the content is available within the local
SWNET. This advent is named as cooperative caching.
Caching techniques is one of the optimum ways to trim the number of requests
made to the server, which results in increasing the performance communication
between the data. Search Engines, Web Browsers, Content Delivery Networks, and
Web Proxies are some systems which widely cache web files. On the web, cache
placement and replacement in proxy servers reduce the average amount of time
taken for data query and the network traffic. The primary purpose of caching is to
improve the performance of data communication. The data accessed by mobile hosts
will reside in spatial and temporal memory. More the space in temporal and spatial
memory, most of the accesses will go to the data that were recently accessed. Hence,
the performance of data communication can be improved by caching frequently
requested data.
Due to a smaller amount of memory space available in mobile devices all down-
loaded contents cannot be kept for a longer duration. For this reason, the device will
delete cached data after downloading the content from the storage. So introducing
a rebate mechanism to encourage the users to store the content as long as they can.
This makes the popular content available for a long time and reduces the content
provisioning cost.

2 Related Work

2.1 Cache Data

In this approach, recently and frequently accessed objects will be saved in the inter-
mediate nodes. The objects will be saved based on some predefined criteria. By this
way, the requests can be served by the intermediate nodes. But, storing a huge num-
ber of objects in the intermediate nodes is the major drawback of this approach and
is not scalable.

[email protected]
Proficient Cooperative Caching in SWNET Using Twin … 429

2.2 Cache Path

In this approach, instead of storing the recently and frequently accessed objects as
done in CacheData approach, the intermediate nodes store only the object paths of
the nearest node where the objects are available. The primary idea of this approach
is to reduce the average time taken to find the requested object. But, in a huge mobile
network, this strategy will fail due to the replacement of new object paths frequently.

2.3 Hybrid Cache

In this approach, both the CacheData and CachePath approach will be used inter-
changeably based on the traversing objects through the intermediate node.

3 Optimized Object Placing for Caching

3.1 Replacement Using Twin Segments Cache

To put forward the optimal object placement we come up with the Twin Segment
cache method in which the available cache space in each device is categorized into
two segments. First is a Unique segment and second is a Duplicate segment (see
Fig. 1).
As the name indicates, Unique segment can save only distinct objects and the
Duplicate segment can save only the most frequently accessed objects without caring
about the object replication.
In the TwinSegment Caching replacement approach, immediately posterior an
object is downloaded from the Content Provider’s server, it will be considered as a
rare object since there is only a single copy of this object in the network. When a
node downloads the same object from one more SWNET node, that object will be
considered as a duplicate object since there are twin replicate copies of that object
in the network.
The path taken to place a new unique object is based on popularity. The comparison
is done between the new object and the less popular object in the complete cache to

Fig. 1 Cache partition

[email protected]
430 B. N. Lakshmi Narayan et al.

find which one is less popular. If the selected object in the cache is less popular than
the new object then it will be replaced with the new object.
On the other hand for a duplicate object, the candidate is selected only from the
first duplicate segment of the cache.
To simplify, a unique object will never be ignored to save just because to save a
duplicated object. The TwinSegment Cache ensures that the optimal object replace-
ment mechanism is achieved. By making use of this approach, all mobile devices
in their Duplicate segments preserve the same objects, but distinct objects in their
Unique segments. In this way, popular contents will be obtainable for a longer dura-
tion in the network either from Unique segmented cache or Duplicate Segmented
cache. This reduces the content provisioning cost when the content is locally avail-
able within a network. The TwinSegment cache makes it possible to keep track of
popular content, which needs to be kept for a long time.

If (Op is downloaded directly || f == true

Op. benefit = Vp + Uip
Op. l = key
Else
Op. benefit = Uip
Op. l = derived
End
Om = Obj with low benefit
If (Op. benefit > Om. benefit)
replace Om with Op
send the change of status message
end
Algorithm 2: optimized caching policy

3.2 Object Provisioning Cost Along with Split Cache

Local and remote hit rate helps to know the object provisioning cost.
• Local hit rate is the folio of exclusive objects stored with a probability that a device
can find the requested object.
• Remote hit rate is equal to the hit probability contributed by the objects stored in
the unique area of all devices in the partition, minus the unique area of the local
cache.

[email protected]
Proficient Cooperative Caching in SWNET Using Twin … 431

4 Optimized Caching Policy

The caching strategy based on content importance, presented in this section, when
there is less memory in the cache for preserving a new object, the object which
is already present in the cache with the less importance is identified and replaced
with the new object which has more total importance. Based on the source, the
importance of a newly downloaded object is calculated. When an object “p” is directly
downloaded from the Content Provider’s server, the copy is labeled as key. On the
other hand, the copy is labeled as derived, if the object is downloaded from another
node in the SWNET partition. The new object is stored in the cache only if its
importance is greater compared to any of the existing cached objects. Along with
the object replacement based on content importance policy approach as exhibited in
the previous sections, provisioning cost reduction needs that a key object should be
cached within the node that is most possibly to produce requests for that object.

INPUT: Op
Flag=f
If (Op is downloaded directly || f == true
Op. benefit = Vp + Uip
Op. l = key
Else
Op. benefit = Uip
Op. l = derived
End
Om = Obj with low benefit
If (Op. benefit > Om. benefit)
replace Om with Op
send the change of status message
end
Algorithm 2: optimized caching policy

5 Caching Based on Content Benefit with Cost and Rebate

The global popularity of objects with high content demand can be represented as

total requests for object  j 

Global popularity 
total request in network

The local popularity of an object can be represented as

total requests from node  a  for object  j 

Local popularity 
total request from  a 

In comparison with other nodes, most of the time only few nodes will be extra
active and produce extra requests per unit time. Individual node-definite request rates
can have a considerable impact on the average object provisioning cost. Due to this

[email protected]
432 B. N. Lakshmi Narayan et al.

reason, it is important to consider this parameter in object placement algorithm as

presented in Algorithm 2 for the content importance based strategy.
Because of mixed nature enfold proxy can never be a good option for deciding
the optimal solution in TwinSegment Cache. The other best option to decide on the
optimal solution for TwinSegment Cache would be to experimentally run the protocol
for all possible values and then choosing the one that output the minimum cost. This
minimum cost is shown as the best optimal TwinSegment cache.

6 Conclusion

The aspiration of work done is to come up with a cooperative caching strategy

for provisioning cost minimization in Social Wireless Networks. The most excellent
cooperative caching approach for provisioning cost reduction in requires the best way
categorization between distinct and replicated objects. Hereby, we propose a Twin-
Segment replacement which is evaluated for android mobile phones. It is proven that
with the approach based on content importance, it provides improved performance in
comparison with TwinSegment cache which is proposed mainly for content demand.

References

1. Zhao, M., Mason, L., & Wang, W. (2008). Empirical study on human mobility for mobile
wireless networks. In Proceedings of IEEE Military Communications Conference (MILCOM).
2. Cambridge Trace File, Human Interaction Study. (2012). http://www.crawdad.org/download/
cambridge/haggle/Exp6.tar.gz.
3. Cohen, E., Krishnamurthy, B., & Rexford, J. (1998). Evaluating server-assisted cache replace-
ment in the web. In Proceedings of the Sixth Annual European Symposium Algorithms
(pp. 307–319).

[email protected]
Proficient Cooperative Caching in SWNET Using Twin … 433

4. Banerjee, S., & Karforma, S. (2008). A prototype design for DRM based credit card transaction
in e-commerce. Ubiquity, 2008.
5. Breslau, L., Cao, P., Fan, L., & Shenker, S. (1999). Web caching and Zipf-like distributions:
Evidence and implications. In Proceedings of IEEE INFOCOM.
6. Perkins, C., & Royer, E. (1999). Ad-Hoc on-demand distance vector routing. In Proceedings
of IEEE Second Workshop Mobile Systems and Applications.
7. Podlipnig, S., & Boszormenyi, L. (2003). A survey of web cache replacement strategies. ACM
Computing Surveys, 35, 374–398.
8. Chaintreau, A., Hui, P., Crowcroft, J., Diot, C., Gass, R., & Scott, J. (2007). Impact of human
mobility on opportunistic forwarding algorithms. IEEE Transactions on Mobile Computing,
6(6), 606–620.
9. BU-Web-Client—Six Months of Web Client Traces. (2012). http://www.cs.bu.edu/techreports/
1999–011-usertrace-98.gz.
10. Wolman, A., Voelker, M., Karlin, A., & Levy, H. (1999). On the scale and performance of coop-
erative web caching. In Proceedings of 17th ACM Symposium Operating Systems Principles
(pp. 16–31).
11. Dykes, S., & Robbins, K. (2001). A viability analysis of cooperative proxy caching. In Pro-
ceedings of IEEE INFOCOM.
12. Korupolu, M., & Dahlin, M. (2002). Coordinated placement and replacement for large-scale
distributed caches. IEEE Transactions on Knowledge and Data Engineering, 14(6), 1317–1329.

[email protected]
A Particle Swarm
Optimization-Backpropagation
(PSO-BP) Model for the Prediction
of Earthquake in Japan

Abey Abraham and V. Rohini

Abstract Japan is a country that suffers a lot of earthquakes and disasters because it
lies across four major tectonic plates. Subduction zones at the Japanese island curves
are geologically complex and create various earthquakes from various sources. Earth-
quake prediction helps in evacuating areas, which are suspected and could save the
lives of people. Artificial neural network is a computing model inspired by biological
neurons, which learn from examples and can be able to do predictions. In this paper,
we present an artificial neural network with PSO-BP model for the prediction of an
earthquake in Japan. In PSO-BP model, particle swarm optimization method is used
to optimize the input parameters of backpropagation neural network. Information
regarding all major, minor and aftershock earthquake is taken into account for the
input of backpropagation neural network. These parameters are taken from Japan
seismic catalogue provided by USGS (United States Geological Survey) such as
latitude, longitude, magnitude, depth, etc., of earthquake.

Keywords Tectonic plates · Artificial neural network · Particle swarm

optimization · Backpropagation · Seismic catalogue

1 Introduction

The earthquake formation is one of the hectic geological disasters that occur on the
surface of earth having various parameters associated with it. It causes unrecover-
able life harm. Seismic changes, changing in the water’s temperature, etc., are the
factors that are associated with earthquake [1]. The effect of this natural disaster is
high because it happens all of a sudden and unpredictable. High death rates, property
damages, etc., can be caused by an earthquake. Basic living conditions, for instance,
water, sanitation, vitality, correspondence, transportation, and so forth is influenced
by this seismic disaster. Earthquakes smash urban communities and towns, and in
addition, the effects provoke the destabilization of the money related and the com-

A. Abraham (B) · V. Rohini

Department of Computer Science, CHRIST (Deemed to Be University), Bengaluru 560029, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 435

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_36

[email protected]
436 A. Abraham and V. Rohini

munal surface of the nation. The impact of the event is awful in light of the way that
it influences a far-reaching extent to happen all of a sudden and unpredictable [2].
Prediction of the place of event, force, and epicentral zone of future major quakes has
been the subject of different legitimate undertakings with especially differentiating
conclusions in the late years. There are persuaded technique which rely upon either
the examination of precursory phenomena beforehand tremors, for instance, seismic
peacefulness, changes in attractive and electric signals recorded at seismic goals, and
sporadic creature conduct.
Artificial Neural Network (ANN) is interconnected neurons with duplication of
some properties of natural neurons [3]. The neural network structure is an extraordi-
narily self-adaptable nonlinear progression. In this work, neural network models are
shown for foreseeing the magnitude of the greatest seismic event in the following
year in perspective of the examination of eight scientifically characterized seismicity
factors. For a given seismic region, eight seismicity pointers computed from a pre-
described number of vital seismic events are passed as contributions to anticipate the
occurrence or non-occurrence of an earthquake. There has been work done as using
other neural network algorithms such as probabilistic neural network and recurrent
neural network for the prediction of the earthquake [4]. Artificial neural network is a
computational model inspired on the biological neurons in human brain. The artifi-
cial neural network contains interconnected neurons, which are highly self-adaptable
and can be used in various problem-solving scenarios [5].

2 Study Area

Pacific plate, North American plate, Eurasia plate, and the Philippine Sea plate are
the four major tectonic plates in which Japan and its islands inhabit. The Pacific plate
is forced sideways and downwards into the mantle, beneath Hokkaido and northern
Honshu, along the eastern margin of the Okhotsk microplate. Toward south, the
Pacific plate is forced sideways and downwards into the mantle beneath volcanic
islands along the eastern margin of the Philippine Sea plate. Creation of deep offshore
Ogasawara and Japan trenches is due to the 2200 km long zone force of sideways and
downwards into the mantle by pacific plate. Similarly, the Philippine Sea plate is itself
forced sideways and downwards into the mantle under the Eurasia plate. Subduction
zones at japan area with respect to geology are too complex. They produce a lot of
earthquakes from multiple sources of origin (Fig. 1).

3 Particle Swarm Optimization

Particle swarm optimization is an optimization technique based on the social

behaviour of bird flocking. Particle swarm optimization technique has similarities
with genetic algorithm. System starts with population of random solutions and finds

[email protected]
A Particle Swarm Optimization-Backpropagation (PSO-BP) Model … 437

Fig. 1 Tectonic plates in Japan

the optimal solution by updating generations. Every particle stores coordinates in

particle with which best solution is found so far [6]. The values are changed from
iteration to iteration. To find the optimal solution, with respect to the previously best
position (pbest) and global best (gbest) position every particle moves.
 
pbest(i, t)  argk1,..,t min[ f ( pi (k))], i ∈ 1, 2, . . . , Np ,
i1,...,N
gbest(t)  argk1,...,t p min[ f (Pi (k))] (1)

where i denotes the index of the particle, N p is the total number of particles, p is the
position, t is the current iteration and f is the fitness function. The position P and
velocity V is updated by following equations:

Vi (t + 1)  ωVi (t) + c1r1 (pbest(i, t) − Pi (t)) + c2 r2 (gbest(t) − Pi (t)),

Pi (t + 1)  Pi (t) + Vi (t + 1) (2)

where V is the velocity, w is the inertia weight, c1 and c2 are the acceleration
coefficients.

4 Backpropagation Neural Network

The backpropagation algorithm is used as a piece of layered feedforward Artificial

Neural Networks. Backpropagation has multiple layers with a supervised learning
framework in light of gradient descent learning rule [7]. We give the algorithm
instances of the sources of inputs and for outputs we require the framework to enlist,
and the error generated afterward is found out. The likelihood of this algorithm is to

[email protected]
438 A. Abraham and V. Rohini

Fig. 2 Backpropagation neural network model with parameters optimized after PSO

diminish this error, until the point when the moment that the Artificial Neural Network
takes in the training information [8]. The neural network model is demonstrated in
the figure (Fig. 2).
The sum of the inputs is multiplied with corresponding weights W ji for generating
activation function.


n
A j (x, w)  xi w ji (3)
i0

Sigmoidal function is used as the output function for this work

1
O j (x, y)  (4)
1 + e Ai(x,w)
Since the error refinement among genuine and expected results, the error relies
upon the weights and we have to alter the weights so as to limit the error. We can
portray the error function for the output of every neuron:

E j (x, w, d)  (O j (x, w) − d j )2 (5)

Then it is found how error value depends on the inputs outputs and weights.

[email protected]
A Particle Swarm Optimization-Backpropagation (PSO-BP) Model … 439

∂E
W ji  −η (6)
∂ W ji

W ji is the adjustment of each weight and ‘η’ is the constant eta. Now we have
to find out how much error depends on output.

∂E  
 2 Oj − dj (7)
∂Oj

Now, to find out how much the output depends upon the activation and then on
weights we compute the following:

∂Oj ∂Oj ∂ Aj  
  O j 1 − O j xi (8)
∂ W ji ∂ A j ∂ W ji

The difference with respect to each weight will be

   
W ji  −2η O j − d j O j 1 − O j xi (9)

If we have to change Vik , the weights (Vik ) of a past layer, we expect first to
register how the error depends not on the weight, yet in the observation from the past
layer, i.e., supplanting W by x as appeared in the below equation.

∂E ∂ E ∂ xi
Vik  −η  −η (10)
∂ Vik ∂ xi I ∂ Vi k

where

∂E    
 2 O j − d j O j 1 − O j W ji (11)
∂ W ji
∂ xi
 xi (1 − xi )Vik (12)
∂ Vik

5 Computed Parameters

To do the expectation of magnitude in the examination zone, the data source stock
collected in the midst of 2010–2016 is used. The test connection between mag-
nitude, frequency, and vitality of seismic tremor occasions is remarkable as the
Gutenberg–Richter (G–R) connections. The recorded earthquake data for the exam-
ination area are divided into different pre-undefined periods of time, for instance,
occasions in light of magnitude more than 3.5, the input to the neural networks are
eight computational parameters called seismicity input vectors [9].

[email protected]
440 A. Abraham and V. Rohini

b-value
Since Gutenberg and Richter surveyed the parameters a and b, the appraisal of
the parameters has been a great part of the time used as a piece of true figuring of
seismicity. Greatest probability b-value were prepared using the following equation
[10].

1
b loge (13)
Mmean − Mmean

Energy e

Energy E (E in ergs) released during an earthquake can be calculated from the

following equation:

log E  5.8 + 2.4m (14)

Energy j

The Seismic wave energy J (ergs) can be calculated from the following equation:

log10 J  9 + 1.8M (15)

Longitude and latitude

Longitude specifies the east–west position of a point on the Earth’s surface and
latitude specifies north–south position supports several clustering-algorithm imple-
mentations, all written in MapReduce, each with its own set of goals and criteria.

6 Conclusion

This prediction of the earthquake using Particle Swarm Optimization-

Backpropagation model is efficient. As the optimized input parameters are given
as input to the backpropagation neural network, this model is more accurate than a
simple backpropagation network. The accuracy for prediction depends on the rele-
vance of the input parameters. This model allows to clear the area indicated according
to future earthquake prediction by it, moreover allows to give awareness to the people
in that location and to do precautions for the loss of lives and financial misfortunes.

Acknowledgements The authors would like to thank the Department of Computer Science at
Christ University, Bengaluru, India for their wholehearted support.

[email protected]
A Particle Swarm Optimization-Backpropagation (PSO-BP) Model … 441

References

1. Pannakkat, A., & Adeli, H. (2007). Neural network models for earthquake magnitude prediction
using multiple seismicity indicators. International Journal of Neural Systems, 17(1), 13–33.
2. Arjun, C. R., & Kumar, A. (2009, March). Artificial neural network—Based estimation of peak
ground acceleration. ISET Journal of Earthquake Technology, Paper no. 501, 46(1), 19–28.
3. Alarifi A. S., & Alarifi, N. S. (2009). Earthquake magnitude prediction using artificial neural
network in northern red sea area. American Geophysical union, fall meeting 2009.
4. Adeli, H., & Panakkat, A. A probabilistic neural network for earthquake magnitude prediction.
Neural Networks, 22 (in press).
5. Widrow, B., & Lehr, M. A. (1990). 30 years of adaptive neural networks: Perceptron, madaline,
and backpropagation. Proceedings of the IEEE, 78(9), 1415–1442.
6. Kennedy, J., & Eberhart, R. C. (1995). Particle swarm optimization. In Proceeding of the 1995
IEEE International Conference on Neural Networks, Perth, Australia (pp. 1942–1948).
7. Xu, F., & Song, X. (2010). Neural network model for earthquake prediction using DMETER
data and seismic belt information. In SecondWRI Global Congress on Intelligent Systems.
8. Moatti, A., & Amin-Nasseri, M. R. (2013). Pattern recognition on seismic data for earthquake
prediction purpose. In International Conference on Environment, Energy, Ecosystems and
Development.
9. Kanmani, K., Vasavi, J., & Ramya, V. (2016). Forecasting earth quake using back propagation
algorithm based on artificial neural network. International Word Press, IJCTA, 9(37).
10. Narayanakumar, S., & Raja, K. (2016). A BP artificial neural network model for earthquake
magnitude prediction in Himalayas, India. In International Conference on Circuits and Systems.

[email protected]
Analysis and Detection of Diabetes Using
Data Mining Techniques—A Big Data
Application in Health Care

B. G. Mamatha Bai, B. M. Nalini and Jharna Majumdar

Abstract In digitized world, data is growing exponentially and Big Data Analytics
is an emerging trend and a dominant research field. Data mining techniques play
an energetic role in the application of Big Data in healthcare sector. Data mining
algorithms give an exposure to analyse, detect and predict the presence of disease and
help doctors in decision-making by early detection and right management. The main
objective of data mining techniques in healthcare systems is to design an automated
tool which diagnoses the medical data and intimates the patients and doctors about
the intensity of the disease and the type of treatment to be best practiced based on the
symptoms, patient record and treatment history. This paper emphasises on diabetes
medical data where classification and clustering algorithms are implemented and the
efficiency of the same is examined.

Keywords Big data health care · Data mining techniques · Gaussian Naïve
Bayes · OPTICS · BIRCH

1 Introduction

In India, healthcare systems have gained importance in the recent years with the
emergence of Big Data analytics [1]. Diabetes mellitus is posing a unique health
problem in the country today, and hence India ranks top in the world. Diabetes is a
chronic medical condition which can be administered and controlled through changes
in lifestyle at an initial stage. At advanced stages, diabetes can be controlled by timely
detection, right diagnosis and proper medication. Statistics as per today quotes that
approximately 145 million people worldwide are affected by diabetes mellitus and
5% of Indian population contributes towards this rate [2].

B. G. Mamatha Bai (B) · B. M. Nalini · J. Majumdar

Department of CSE, Nitte Meenakshi Institute of Technology, Bangalore, India
e-mail: [email protected]
J. Majumdar
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 443

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_37

[email protected]
444 B. G. Mamatha Bai et al.

Diabetes is a condition in which the human body will not be able to generate the
available amount of insulin which is necessary to balance and monitor the quantity
of sugar in the body. This disease accounts for various diseases such as heart disease,
nerve damage, kidney disease, blindness and blood vessels damage and many more
[3]. Diabetes depends on two reasons:
• Required amount of Insulin is not produced by the pancreas. This specifies Type-1
diabetes and occurs in 5–10% of people.
• In Type-2, cells become inactive to the insulin being produced. Gestational diabetes
occurs in women when they develop high sugar level during the time of pregnancy
and may disappear after that.
Interpretation and analysing the presence of diabetes is a significant problem to
classify. A classifier is essential and intended in such a way that is more cost efficient,
convenient and precise. Big Data and data mining techniques provide a great deal
to human-related application. These methods find the most appropriate space in
the medical diagnosis which is one of the classification phenomena. A physician is
supposed to analyse many factors before actual diagnosis of the diabetes leading to a
difficult task. In recent times, designing of automatic diabetes diagnosis system has
been designed by adopting machine learning and data mining techniques.

2 Literature Survey

The predictive analysis of diabetic patients based on the age and gender using var-
ious data mining classification algorithms techniques are discussed [4]. Diabetes is
characterised by abnormal metabolism, mostly hyperglycaemia, and an associated
risk for specific complications affecting the eyes, kidney, and nervous system and
many more. Classification by decision tree induction and classification by Bayesian
network are used for predicting the diabetes for people of different age groups and
gender.
Ordering points to identify the clustering structure (OPTICS) algorithm is used for
the cluster analysis and it does not explicitly produce the clusters, rather it creates an
improved ordering of the data objects that depicts the structure of the clusters obtained
based on density [5, 6]. This ordering of clusters contains information equal to the
density-based clustering and equivalent range of parameter settings. This suits well
for automatic and interactive analysis of clusters.
OPTICS [7] is a method based on density and an improved version of DBSCAN.
It eliminates the negative aspects of DBSCAN, like forming the clusters in varying
density environment. It specifies the shortest distance from the core comprising the
minimum number of data points and calculates the mean distance among the points
which are placed in and around the core and results in a new distance as the new core
distance.
Balanced iterative reducing and clustering using hierarchies (BIRCH) is an effi-
cient and accountable data clustering method, which is based on a new in-memory

[email protected]
Analysis and Detection of Diabetes Using Data Mining Techniques … 445

data structure termed CF tree that summarises the data distribution [8]. BIRCH
diminishes the problem of identifying the clusters from the initial dataset into one by
clustering the set of summaries, which are much slighter than the original dataset.
BIRCH [9] eventually clusters the inward multidimensional data points and pro-
duces the clustering with best acceptable quality from the available resources such as
memory and time constraints. BIRCH typically finds a better clustering by scanning
the data once and enhances the quality with additional scans ahead [10, 11]. BIRCH
was initially designed to form clusters by handling noise in the dataset.

3 Methodology

Data mining is a new pattern for analysing medical data and achieving useful and
practical patterns. Data mining helps us to predict the type of disease and do not try
to confirm the already identified patterns; rather tries to find already non-identified
patterns. The objective of the proposed methodology is to analyse the medical dataset
and predict whether the patient is suffering from the diabetes disease or not. This
prediction is done using data mining algorithms such as Gaussian Naïve Bayes,
BIRCH and OPTICS.
Naïve Bayes data mining technique is applied to the dataset to predict whether
the patient is diabetic or not. BIRCH and OPTICS clustering algorithms are used to
cluster similar kind of diseased people into one cluster and identify which algorithm
is more efficient by calculating the performance measures.

3.1 Input Dataset

The source of medical dataset on diabetes is obtained from the UCI machine learn-
ing repository from the link https://archive.ics.uci.edu/ml/datasets/Pima+Indians+
Diabetes [12]. The residents of Arizona, USA and those suffering from diabetes are
tested and their data is obtained. It is observed that major percentage of the popula-
tion suffer from diabetes and the reason behind this is overweight. The dataset under
study comprises nine attributes as shown in Table 1.

3.2 Algorithms

3.2.1 Gaussian Naïve Bayes

Naïve Bayes is used for constructing the classifier. Naive Bayes classifiers belong
to a family of simple probabilistic classifiers which works on applying the Bayes’
theorem with complete (naive) independence assumptions amongst the features [13].

[email protected]
446 B. G. Mamatha Bai et al.

Table 1 Attributes of S. No. Attributes

diabetic medical dataset
1. No. of times being pregnant
2. Glucose tolerance test
3. Blood pressure level (mmHg)
4. Body mass index (BMI)
5. Triceps (mm)
6. Insulin (mu U/ml)
7. Age (years)
8. Pedigree function w.r.t diabetes
9. Indicating presence or absence of diabetes

Fig. 1 Core distance and

reachability distance

Gaussian Naïve Bayes is used for the continuous dataset. Detailed steps of the algo-
rithm are given in Appendix 1.

3.2.2 OPTICS

OPTICS algorithm is used to group the data points considering the density of the
objects [14]. It needs two parameters viz. ε-Maximum distance (radius) and MinPts
(number of points necessary for forming clusters). A point ‘P’ is said to be a core
point if it has at least minimum number of points within the neighbourhood. Two
values are needed to be considered for each of the object as shown in Fig. 1.
• Core distance

The core distance refers to the least distance ε between point p and a data object in
its ε-neighbourhood where p becomes a core object.

[email protected]
Analysis and Detection of Diabetes Using Data Mining Techniques … 447

Fig. 2 Overview of BIRCH algorithm

• Reachability distance

It is the distance between the two points p and o or core distance of p, whichever is
bigger.
Detailed algorithm steps are given in Appendix 2.

3.2.3 BIRCH

The BIRCH clustering is a type of hierarchical clustering. The algorithm uses the
clustering feature and it builds a CF tree. It works for large datasets. It is the first
algorithm which is used to handle the noise [15]. There are four phases in BIRCH
algorithm as shown in Fig. 2.
The working description of each phase is as follows:
Phase 1: Scan the dataset and build an initial in-memory CF tree.
Phase 2: Scan all the leaf entities of the CF tree and build a new CF tree which is
smaller in size. Eliminate all the outliers and form the clusters.
Phase 3: Use the clustering algorithm to cluster all the leaf entities. This phase results
in creating a set of clusters.
Phase 4: The cluster centroids obtained in Phase 3 are used as seeds and the data points
are redistributed to its closest neighbour seeds to form new cluster representations.
Finally, each leaf entity signifies each cluster class.
Details of the algorithm steps are given in Appendix 3.

[email protected]
448 B. G. Mamatha Bai et al.

4 Experimental Analysis

4.1 Gaussian Naïve Bayes

The Gaussian Naïve Bayes analyses the input dataset, segments the continuous values
into classes and predicts whether the patient is diabetic or not. The feature value in
our example is considered to be the age attribute and the output is shown in Fig. 3.
The output depicts that the people are affected by diabetes around the age of 20 and
above considering the various factors such as hereditary, lifestyle, food habits, stress,
etc.

4.2 OPTICS

The entire dataset is grouped into three clusters based on different stages of dia-
betes w.r.t glucose tolerance test. The clusters formed are normal, prediabetes and
diabetes. The cluster plot and each cluster representation are shown in Fig. 4a, b,
respectively. The clusters are formed by considering the associated symptoms for
different categories of people such as normal, prediabetes and diabetes for detailed
cluster analysis and the same is shown in Fig. 5. Figure 6 represents the plot of
insulin dependency with the age factor. Analysis of Fig. 3 with 6 shows that people
are affected by diabetes at an early age of 20.

4.3 BIRCH

The BIRCH algorithm builds a CF tree and works with large dataset which is capable
enough to handle noise. By analysing the diabetic dataset, the three clusters are
formed as shown in Fig. 7a, b. Figure 8 refers to the plot of glucose tolerance test
with age to illustrate whether the victim is diabetic or not. Figure 9 represents the

Fig. 3 Predicting diabetic or

non-diabetic

[email protected]
Analysis and Detection of Diabetes Using Data Mining Techniques … 449

(a) Cluster Plot of OPTICS Algorithm (b) Clusters of OPTICS Algorithm

Fig. 4 Results obtained for OPTICS algorithm

Fig. 5 Cluster analysis of OPTICS algorithm

Fig. 6 Plot of insulin dependency with age

[email protected]
450 B. G. Mamatha Bai et al.

(a) Cluster Plot of BIRCH Algorithm (b) Clusters of BIRCH Algorithm

Fig. 7 Results obtained for BIRCH algorithm

Fig. 8 Glucose tolerance test versus age

Fig. 9 Plot of BMI and glucose test

plot of body mass index (BMI) with glucose test. Figure 10 depicts different age
groups of people having comparably good health factor.

[email protected]
Analysis and Detection of Diabetes Using Data Mining Techniques … 451

Fig. 10 Scatter plot of normal health

4.4 Performance Analysis Statistics

4.4.1 Silhouette Method

Silhouette methods are used by the researchers to interpret and validate the consis-
tency of the data within the clusters. This method pictorially represents how well
each data object fits into the cluster [16]. The range of Silhouette width lies between
−1 and +1 where the higher value represents the strong similarity and lower value
represents the weak similarity of data objects within a cluster.
The results of Silhouette method for OPTICS and BIRCH algorithm are as shown
in Figs. 11 and 12, respectively. The average Silhouette width for OPTICS algorithm
is 0.44 with less outliers pointing towards the negative scale of the X-axis, whereas
the Silhouette width is 0.31 in the case of BIRCH algorithm with few outliers. This
clearly indicates that the similarity of the data objects is high in OPTICS.

Fig. 11 Results of
Silhouette method for
OPTICS algorithm

[email protected]
452 B. G. Mamatha Bai et al.

Fig. 12 Results of
Silhouette method for
BIRCH algorithm

Table 2 Performance metric Precision Recall F- Rand

values measure index
OPTICS 0.735849 0.795918 0.7647058 0.7108433
BIRCH 0.523809 0.523809 0.5238095 0.5381062

Fig. 13 Cluster analysis of

various performance metrics

4.4.2 Performance Metrics

Performance metrics are necessary to compare and analyse the clustering quality of
various clustering methods. Higher the performance metric values, superior is the
clustering quality. The performance metrics like precision, recall, F-measure and
Rand Index are used to obtain the best clustering algorithm as shown in Table 2
and Fig. 13. From the analysis, it is found that OPTICS algorithm performs most
efficiently compared to BIRCH algorithm.

5 Conclusion and Future Work

In this paper, we have demonstrated the usefulness of data mining algorithms like
Gaussian Naïve Bayes, BIRCH and OPTICS for the prediction of diabetic dis-

[email protected]
Analysis and Detection of Diabetes Using Data Mining Techniques … 453

ease. Data mining techniques are effective in diagnosing and clustering the diabetic
patients. Gaussian Naïve Bayes classifier is used for predicting based on the probabil-
ities. BIRCH and OPTICS are used to cluster similar kind of people, where BIRCH
is based on the CF tree and OPTICS is based on the ordering of the points in the
cluster.
Analysis and comparison of clustering algorithms are performed by considering
various performance metrics. It is observed that for the same number of clusters
obtained by different clustering techniques, OPTICS is most efficient and is suitable
for diagnosis of diabetes.
This work helps the doctors to diagnose and provide the recommended medicine
at an early stage. The main aim is to reduce the cost and provide better treatment. In
future, this work can be extended with more number of classification algorithms and
their accuracy can be compared to find the optimal one.

Acknowledgements The authors express their sincere gratitude to Prof. N. R. Shetty, Advisor
and Dr. H. C. Nagaraj, Principal, Nitte Meenakshi Institute of Technology for giving constant
encouragement and support to carry out research at NMIT.
The authors extend their thanks to Vision Group on Science and Technology (VGST), Government
of Karnataka to acknowledge our research and providing financial support to set up the infrastructure
required to carry out the research.

Appendix 1

Algorithm: Gaussian Naïve Bayes

Input: Dataset
Output: Classification into different categories
Algorithm Steps:
Step 1. Segment the data by the class.
Step 2. Calculate the probability of each of the class.

Class Count
Class Pr obabilit 
T otal Count
Step 3. Find the average and variance of individual attribute x belonging to a class c.
a. Let μx be the average of the attribute values in x allied with class c.
b. Let σx2 represent the variance of the attribute values in x related with class c.
Step 4. The probability distribution is computed by
−(v−μx )2
1
p(x  v|c)   e 2σx2

2π σx2

Step 5. Calculate the probability of the attribute x

[email protected]
454 B. G. Mamatha Bai et al.

P(x1 , x2......... xn |c)  P(xi |c)
i


Naïve Bayes Classifier  argmax P(c) P(xi |c)
i
Step 6. End.

Appendix 2

Algorithm: OPTICS
Input: Dataset
Output: Clusters
Algorithm Steps:
Step 1. Initially ε and MinPts need to be specified.
Step 2. All the data points in the dataset are marked as unprocessed.
Step 3. Neighbours are found for each point p which is unprocessed.
Step 4. Now mark the point as processed.
Step 5. Initialize the core distance for the data point p.
Step 6. Create an Order file and append point p to the file.
Step 7. If core distance initialization is unsuccessful, return back to Step 3 otherwise
go to Step 8.
Step 8. Calculate the reachability distance for each of the neighbours and update the
order seed with the reference of new values.
Step 9. Find the neighbours for each data point in the order seed and update the point
as processed.
Step 10. Fix the core distance of the point and append to the order file.
Step 11. If undefined core distance exists, go back to Step 9, else continue with Step
12.
Step 12. Repeat Step 8 until there is no change in the order seed.
Step 13. End.

Appendix 3

Algorithm: BIRCH
Input: Dataset
Output: Clusters
Algorithm steps:
Step 1. Set an initial threshold value and insert data points to the CF tree w.r.t the
Insertion algorithm.
Step 2. Increase the threshold value if the size of the tree exceeds the memory limit
assigned to it.

[email protected]
Analysis and Detection of Diabetes Using Data Mining Techniques … 455

Step 3. Reconstruct the partially built tree according to the newly set threshold values
and memory limit.
Step 4. Repeat Step 1 to Step 3 until all the data objects are scanned and form a
complete tree.
Step 5. Smaller CF trees are built by varying the threshold values and eliminating
the Outliers.
Step 6. Considering the leaf entities of the CF tree, the clustering quality is improved
by applying the global clustering algorithm.
Step 7. Redistribution of data objects and labelling each point in the completely built
CF tree.

References

1. Raghupathi, W., & Raghupathi, V. (2014). Big data analytics in healthcare: Promise and poten-
tial. Health Information Science and Systems.
2. Diabetes Mellitus. https://en.wikipedia.org/wiki/Diabetes_mellitus.
3. Agicha, K., et al. Survey on predictive analysis of diabetes in young and old patients. Interna-
tional Journal of Advanced Research in Computer Science and Software Engineering.
4. Iyer, A., Jeyalatha, S., & Sumbaly, R. (2015, January). Diagnosis of diabetes using classification
mining techniques. International Journal of Data Mining & Knowledge Management Process
(IJDKP), 5(1).
5. Ankerst, M., Breunig, M. M., Kriegel, H.-P., & Sander, J. OPTICS: Ordering Points To Identify
the Clustering Structure. Institute for Computer Science, University of Munic.
6. Alzaalan, M. E., & Aldahdooh, R. T. (2012, February). EOPTICS “Enhancement ordering
points to identify the clustering structure”. International Journal of Computer Applications
(0975–8887), 40(17).
7. Senthil kumaran, M., & Rangarajan, R. (2011). Ordering points to identify the clustering struc-
ture (OPTICS) with ant colony optimization for wireless sensor networks. European Journal
of Scientific Research, 59(4), 571–582 (ISSN 1450-216X).
8. Zhang, T., Ramakrishnan, R., & Livny, M. (1997). BIRCH: A new data clustering algorithm
and its applications. Data Mining and Knowledge Discovery, 1, 141–182.
9. Zhang, T., Ramakrishnan, R., & Livny, M. BIRCH: An efficient data clustering method for
very large databases.
10. Du, H. Z., & Li, Y. B. (2010). An improved BIRCH clustering algorithm and application in
thermal power. In 2010 International Conference on Web Information Systems and Mining.
11. Feng, X., & Pan, Q. The algorithm of deviation measure for cluster models based on the
FOCUS framework and BIRCH. In Second International Symposium on Intelligent Information
Technology Application.
12. UCI Machine Learning Repository Pima Indians Diabetes Database https://archive.ics.uci.edu/
ml/datasets/Pima+Indians+Diabetes.
13. Naïve Bayes. https://en.wikipedia.org/wiki/Naive_Bayes_classifier.
14. Optics Algorithm. https://en.wikipedia.org/wiki/OPTICS_algorithm.
15. Birch Algorithm. https://people.eecs.berkeley.edu/~fox/summaries/database/birch.html.
16. Silhouette Method. https://en.wikipedia.org/wiki/Silhouette_(clustering).

[email protected]
Cyclic Scheduling Algorithm

Ravin Kumar

Abstract CPU Scheduling has played a critical role in making efficient systems, It
is the mechanism which allows one process to use the CPU for execution while other
processes are put on hold because of unavailability of any required resource, with
the aim of maximizing the CPU utilization and reducing the waiting time and the
turnaround time. This paper presents a new Scheduling Algorithm, which supports
preemption, reduces the turnaround time and the waiting time. To show its effective-
ness, its comparison is done with other traditional scheduling algorithms including
First Come First Serve, Shortest Job First, and Round Robin Scheduling Algorithm,
and as a result it was found that the proposed algorithm provides a new and effective
scheduling approach which reduces the average waiting time and average turnaround
time in a much better way than the traditional approaches.

Keywords Preemption · Scheduling · Resource utilization · Waiting time

1 Introduction

Computers have become an important part of today’s Information Age, due to their
ability to perform several tasks simultaneously. Inside a computer, processors play
a crucial role in decision-making, and calculation-related tasks. When processor
has more than one process to execute, at that time scheduling algorithms [1] are
required to manage execution efficiency of processes. CPU Scheduling is one of
the fundamental concepts of Operating System, which are designed with the aim of
maximizing the utilization of available system resources, resulting in a better, fair,
and fast use of shared resources among the available processes. There are various
traditional scheduling algorithms which were designed with the aim of better resource
utilization, some of them are FCFS, SJF, and Round Robin scheduling algorithm [2].
There are many factors that require attention during the scheduling process, this

R. Kumar (B)
Department of Computer Science, Meerut Institute of Engineering and Technology, Meerut
250005, Uttar Pradesh, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 457

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_38

[email protected]
458 R. Kumar

includes burst time, arrival time, and priority of the process, with the aim of reducing
the average waiting time and average turnaround time [3]. We have studied various
scheduling algorithms from existing literature.
FCFS (First Come First Serve)—This is one of the most earlier approaches
which was used for scheduling purpose. This follows the ideology of “First Come
First Serve”, i.e., assigning CPU time first to that process, which requested them first
[4].
SJF (Shortest Job First)—In this approach, CPU is allocated to that process
which has the smallest burst time available, i.e., the job with the shortest computation
time is executed first. When a job first arrives, it is put inside the ready queue, and
then the job which has the smallest burst time is chosen for execution [5].
Round Robin Scheduling—In this approach, a preemptive time-slicing is per-
formed with the help of a time quantum or time slice. A time slice or time quantum
is a small unit of time, due to which it can give the effect of all processors sharing
the CPU equally with its given time quantum.
This paper proposes a new CPU scheduling algorithm, which can be used for the
same purpose, and when compared with other approaches, it shows comparatively
good results.

2 Related Work

Kumar et al. [6] designed a method to suggest the length of the next CPU burst in
SJF, and showed that approximate length of the next CPU burst is similar to the
length of the previous request. Khan et al. [7] performed a comparative study over
SJF and FCFS for similar priority jobs. Li et al. [8] improvised the FIFO algorithm
by using fuzzy logic, and showed that it reduces the execution time for tasks and
increases the resource utilization. Indusree et al. [9] suggested an improvement to RR
by calculating the dynamic time quantum each burst time for processes present in the
ready queue. Singh et al. [10] developed a multi-queue-based scheduling approach
for cloud architectures. Lin et al. [11] showed that weighted RR performs better
than FCFS in Hama architecture. Jha et al. [12] proposed a hybrid algorithm of RR
and priority scheduling, which was an improvement over both traditional RR. In
servers based on OpenFlow, Peng et al. [13] suggested a weighted RR algorithm
that can be used for load balancing. Chaturvedi et al. [14] utilized improved RR
in workflow applications in the cloud. Rao et al. [15] proposed a dynamic time
calculation mechanism for RR for improving its performance.

3 Proposed Scheduling Algorithm

Let us assume a set of processes P0 , P1 , P2 ,…, Pn that are needed to be executed

using the proposed scheduling algorithm. QUEUE is representing a queue type data

[email protected]
Cyclic Scheduling Algorithm 459

structure which holds those process which has been executed partially. Now, let us
assume that there is a set ‘L’, which holds all those processes that are not completed
and their Arrival Time is either less than or equal to the current time. Flowchart of
the proposed algorithm is given in Fig. 1.
Cyclic Scheduling Algorithm ( )
Parameters QOLD and QNEW represent the dynamic time quantum, CNT represents an
integer value that is used to decide the dynamic time quantum. RTX is the remaining
time of the selected process. Max and Min are maximum and minimum remaining
time among processes that are either present in L or are in QUEUE. N represents the
total number of active and arrived processes.
Given below is the flowchart of the proposed scheduling algorithm (Fig. 1).

Fig. 1 Flowchart of Cyclic

Scheduling Algorithm

[email protected]
460 R. Kumar

BEGIN
Step 1: Set L={ }, QUEUE = Null, CNT =0 and QNEW = 0.
Step 2: Add all the new process that had arrived, in set L, and update CNT= CNT+1
Step 3: PX = Select_Process( CNT )
Step 4: IF number of process in L == 1
then
IF QUEUE == Null
then
QOLD = QNEW = 1
ELSE
QNEW = QOLD
end IF
ELSE
// calculating new time quantum.
QNEW = mininum ( RTX , ( Max – Min) /2 , ΣRT/N )
QOLD = QNEW .
// in above formula remove all zero values, and calculate QNEW ( >= 1).
end IF
Step 5: After running PX for QNEW time, remove PX from L, and add (PX , NX) to
QUEUE, here NX can be calculated as :
NX= NTEMP + number of ( P , N ) pairs present in QUEUE.
Where,
NTEMP = Number of process in L + Number of process that just + 1
arrived during this execution.
Step 6: Calculate_Jump ( )
Step 7: IF All Processes are completed and L== { } and QUEUE == Null
then
Exit.
Else
Goto step 2.
END

Select_Process ( CNT )
This function is used to decide which process to be chosen for execution from set L.
BEGIN
Step 1: IF CNT % 2 == 1
then
return Process which have minimum remaining time in set L.
ELSE
return Process present in L and have arrival time closer to
( Maximum Arrival time + Minimum Arrival time ) / 2.
Maximum and Minimum are calculated among processes present in L.
END

[email protected]
Cyclic Scheduling Algorithm 461

Calculate_Jump ( )
This function is used to make use of dynamic time quantum to place processes back
from QUEUE to set L.
BEGIN
Step 1: For each ( PT, NT ) present in QUEUE
IF NT <= Number of ( P , N ) pairs present in QUEUE.
Then
For each ( PY, NY ) present in QUEUE
N Y = NY - 1
end For
Remove ( PT, NT ) from QUEUE.
IF Remaining time of PT > 0
then
Put process PT in set L.
end IF
end IF
end For
END

4 Working Demonstration

To better understand the proposed algorithm, we have considered a set of processes

with burst and arrival times, so that the behavior of the new proposed scheduling
algorithm can be easily understood (Table 1).
Initially, we have P  { }, QUEUE  Null, CNT  0 and QNEW  0.
Table 2 shows the status of processes, after applying the proposed CSA.
After applying the proposed scheduling algorithm, the waiting time and
turnaround time of each process are evaluated, it is given in Fig. 2.

Table 1 Sample data to demonstrate the procedure of Cyclic Scheduling Algorithm

Process Arrival time Burst time
P0 0 2
P1 0 3
P2 2 1

Table 2 Applying Cyclic Scheduling Algorithm

Current time CNT QNEW Set L { } QUEUE
0 0 0 {P0, P1} Null
1 1 1 {P1} (P0, 2)
2 2 1 {P0, P2} (P1, 2)
3 3 1 {P1, P2} (P0, 2)
4 4 1 {P1} (P2, 2)
5 5 1 {} (P1, 1)
6 6 1 {} Null

[email protected]
462 R. Kumar

Fig. 2 Representation of
individual turnaround and
waiting time using Cyclic
Scheduling Algorithm

Table 3 Overall performance Parameter Value

of the proposed scheduling
algorithm on sample data Average turnaround time 2.66
Average waiting time 0.66

The overall performance of the algorithm can be determined by calculating the

total waiting time and total turnaround time for the given sample of data. Using
the above values of waiting time and turnaround time of the individual processes,
the average turnaround time and average waiting time can be calculated, so that the
overall performance of the proposed algorithm can be measured. In Table 3, the
average turnaround time and average waiting time of the above processes is shown.

5 Result and Discussion

In the proposed algorithm, the outer loop containing set ‘L’ and queue ‘QUEUE’
makes sure that the algorithm has a cyclic nature for the execution of process, similar
to that of round-robin algorithm. But instead of using the order in which the processes
first arrived, our algorithm has an inner mechanism which helps in improving the
algorithm’s performance by providing dynamic time quantum, and then selecting a
process on the basis of arrival time and remaining time to execute in the outer loop.
Cyclic Scheduling Algorithm is tested with various test cases on real data sets,
and then the obtained result is further compared with FCFS, SJF, and Round Robin
Algorithm. To show its effectiveness, four samples of data are taken for applying the
scheduling algorithm, and then a comparison is done with FCFS, SJF and Round
Robin scheduling algorithms (Table 4).
In Fig. 3, average turnaround and average waiting time are obtained after applying
each algorithm is given.

[email protected]
Cyclic Scheduling Algorithm 463

Table 4 Sample 1—Burst Process Arrival time Burst time

time in increasing order
P0 0 2
P1 1 7
P2 2 8
P3 3 14

Fig. 3 Comparison of
average turnaround and
average waiting time of
FCFS, SJF, RR, and CSA
scheduling algorithms

Table 5 Sample 2—Burst Process Arrival time Burst time

time in decreasing order
P0 0 14
P1 1 8
P2 2 7
P3 3 2

Table 6 Sample 3—Burst Process Arrival time Burst time

time in random order
P0 0 2
P1 1 8
P2 2 7
P3 3 14

Let us consider another sample of data to further check the result, Given below is
the set of assumed data on which proposed scheduling algorithm is applied and then
the comparison is done with other algorithms (Table 5).
Given below is the graphical representation of the average turnaround and average
waiting time obtained after applying each algorithm (Fig. 4).
Similarly, consider another sample of data on which we first apply the proposed
scheduling algorithm, and then a comparison is done with other algorithms (Table 6).
For the above sample, the comparative analysis of performance obtained after
implementing different scheduling algorithms is given in Fig. 5.

[email protected]
464 R. Kumar

Fig. 4 Comparison of
average turnaround and
average waiting time of
various scheduling
algorithms

Fig. 5 Comparison of
average turnaround and
average waiting time of
various scheduling
algorithms

Table 7 Sample 4—Set of Process Arrival time Burst time

ten processes assigned with
random values P0 0 3
P1 1 14
P2 2 7
P3 2 2
P4 3 11
P5 3 5
P6 4 8
P7 5 13
P8 6 9
P9 6 3

The fourth sample is generated by randomly assigning arrival and burst time to a
set of ten processes (Table 7).
Given below is the comparative analysis of different scheduling algorithms over
the above set of processes (Fig. 6).

[email protected]
Cyclic Scheduling Algorithm 465

Fig. 6 Comparison of
average turnaround and
average waiting time of
various scheduling
algorithms

6 Conclusion

In this paper, an advanced method is introduced which helps in the reduction of

average turnaround time, and average waiting time required by a set of processes
(having the same priority) of completing their tasks. Proposed algorithm provides a
better approach for scheduling processes as compared to other traditional approaches.
This algorithm provides a new way of scheduling processes, and also opens a way for
developing new hybrid scheduling algorithms by combining the proposed algorithm
with other traditional algorithms. Future research work can be done on adding a
priority mechanism to the proposed algorithm.

References

1. Silberschatz, A., & Gagne, G. Operating system concepts (8th ed.). WILEY.
2. Mittal, N., Garg, K., & Ashish America (2005). A paper on modified round robin algorithm.
IJLTEMAS, 4(X1). ISSN 2278–2540.
3. Seltzer, M., Chen, P., & Outerhout, J. (1990). Disk scheduling revisited in USENIX. In H. M.
Shamim (ed.), Operating system, DCSA-2302.Winter technical conference.
4. Raman, & Mittal, P. K. (2014). An efficient dynamic round robin CPU scheduling algorithm.
IJARCSSE.ISBN: 2277 128X.
5. Tong, W., & Zhao, J. (2007). Quantum varying deficit round robin scheduling over priority
queues. In International Conference on Computational Intelligence and Security (pp. 252–256).
China.
6. Kumar, M. R., & Renuka Rajendra, B. (2014). Prediction of length of the next CPU burst in SJF
scheduling algorithm using dual simplex method. In 2nd International Conference on current
Trends in Engineering and Technology ICCTET .
7. Khan, R., & Kakhani, G. (2015). Analysis of priority scheduling algorithm on the basis of
fcfs and sjf for similar priority jobs. International Journal of Computer Science and Mobile
Computing, 4, 324–331.
8. Li, J., Ma, T., Tang, M., Shen, W., & Jin, Y. (2017). Improved FIFO scheduling algorithm based
on fuzzy clustering in cloud computing. Information, 8(1), 25.

[email protected]
466 R. Kumar

9. Indusree, J. R., & Prabadevi, B. (2017). Enhanced round robin CPU scheduling with burst time
based time quantum. In IOP Conference Series: Materials Science and Engineering (Vol. 263,
No. 4, p. 042038). IOP Publishing.
10. Singh, J., & Gupta, D. (2017). an smarter multi queue job scheduling policy for cloud comput-
ing. International Journal of Applied Engineering Research, 12(9), 1929–1934.
11. Lin, Z., & Huang, H. (2017). Research on weighted rotation fair scheduling algorithm based on
hama parallel computing framework. DEStech Transactions on Engineering and Technology
Research apop.
12. Jha, H., Chowdhury, S., & Ramya, G. (2017). Survey on various scheduling algorithms. Impe-
rial Journal of Interdisciplinary Research, 3(5).
13. Peng, R., & Ding, L. (2017). Research on virtual network load balancing based on OpenFlow.
In AIP Conference Proceedings (Vol. 1864, No. 1, p. 020014). AIP Publishing.
14. Chaturvedi, P., & Sharma, S. (2017). An provision for workflow scheduling using round robin
algorithm in cloud. International Journal of Advanced Research in Computer Science, 8(5),
667–673.
15. Rao, G. S. N., Srinivasu, N., Srinivasu, S. V. N., & Rao, G. R. K. (2015). Dynamic time slice
calculation for round robin process scheduling using NOC. International Journal of Electrical
and Computer Engineering, 5(6), 1480–1485.

[email protected]
A Framework for Monitoring Clustering
Stability Over Time

K. Namitha and G. Santhosh Kumar

Abstract Mining data streams and arriving at intelligent decisions is becoming

more and more important nowadays as a lot of applications produce large volume
data streams. Data stream clustering has been considered to be very useful for online
analysis of streams. Monitoring the cluster transitions over time provide good insight
into the evolving nature of the data stream. This paper introduces a framework for
monitoring the stability of individual clusters and clusterings over time, along with the
progress of the stream. Tracking the historical evolution of clustering structures is the
main focus of this framework. Two real-world datasets have been used for conducting
the experiments. The results point up the fact that monitoring the stability of clustering
structures will help to get an important hint of the physical events happening in the
environment. This information can be used to predict the future clustering structure
changes and in turn the upcoming events.

Keywords Data streams · Clusters · Evolution tracking

1 Introduction

With the rapid growth in hardware and software technologies, it has become easier
and cheaper to produce data at large scale. Data arriving as continuous, fast stream
makes its online analysis an inevitable need of the era. Classification and clustering
are the important machine learning approach tried in case of data streams. Different
approaches can be seen in the literature for online clustering of the data streams
[1–5].
The concept drift happening along the evolution of the stream is an important con-
cern in data stream mining. Change in the underlying data distribution is the basic
reason for concept drift. In data stream clustering scenarios, concept changes can be
detected only by identifying changes in the data distribution. Since the data distribu-

K. Namitha (B) · G. Santhosh Kumar

Department of Computer Science, Cochin University of Science and Technology,
Kochi, Kerala, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 467

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_39

[email protected]
468 K. Namitha and G. Santhosh Kumar

tion is closely related to the clustering structure, monitoring the clustering structure
has good scope in identification and anticipation of concept change. In addition to
this, understanding the changes happening in the clustering structure can be useful for
studying the characteristics of the domain in detail and especially decision-making.
It has application in many areas like customer relationship management, fraud dis-
covery, healthcare systems, etc. Studying the interrelationship between the clustering
structure changes and the real events can help deriving better business strategies and
decisions.
MONIC [6] and MEC [7] are two important approaches discussed in the literature
for identifying the different category of cluster transitions. Various possible internal
and external cluster transitions are defined in these papers. But they do not assess
the stability of clustering structures. In this paper, we are proposing a framework
that monitors the lifetime of individual clusters and clusterings over time. From the
beginning of the stream, each cluster and clustering are tracked from their birth till
disappearance. This helps to track the evolution of the data stream. Experiments
conducted on two real-world datasets supports the fact that monitoring the clustering
stability and hence tracking the evolution helps to derive a relationship between
the clustering structures and the physical events. This kind of analysis is useful to
understand the changes happening over the stream.
Rest of this paper is organized as follows. Section 2 elaborates the related work.
Section 3 describes the proposed framework in detail. Experiments and discussion
are included in Sect. 4 and finally, Sect. 5 concludes the paper.

2 Related Work

Plenty of algorithms are available for data stream clustering, a brief account of which
can be seen in [8, 9]. But relatively few works can be seen in identifying and tracking
cluster transitions. MONIC [6] is a framework introduced for modelling cluster tran-
sitions. They have categorized cluster transitions to be internal and external. Clusters
generated at two consecutive time points are compared to decide the kind of tran-
sitions happened. Spiliopoulou et al. [10] have listed some category of application
areas which used MONIC framework for cluster transition analysis. This includes
evolution in social networks, change prediction in data stream mining, spatiotem-
poral analysis, topic evolution, etc. MEC [7] is another similar framework which
models different cluster transitions. MEC uses bipartite graphs for tracking cluster
transitions. In addition to the usual enumeration method, cluster representation by
summary is also considered in this framework.
MONIC+ [11] is an extension of the MONIC framework. It defines different types
of clusters, based on which the cluster overlap and cluster transitions are redefined.
Held et al. [12] present the visualization method for dynamic clusterings. In addition
to the cluster transitions discussed in MONIC, this paper introduces one more transi-
tion named ‘rebirth’ of a cluster. ReDSOM method [13] identifies clustering structure
changes by kernel density estimation approach. This paper also defines the cluster

[email protected]
A Framework for Monitoring Clustering Stability Over Time 469

transitions like emergence, disappearance, split, absorption, cluster compaction and

cluster enlargement, etc.

3 Proposed Framework

This section contains the details of the proposed framework which is intended to be
useful for monitoring the stability of clustering structures over time. Here, both the
individual clusters and the whole clusterings are considered for tracking the stability.
Stability refers to the total lifetime—right from its birth till death/disappearance.
To identify and nomenclature the cluster transitions between two consecutive time
points, this framework is using the cluster transition models introduced by MONIC
and MEC. Particularly ‘survival’ of clusters to the next time point decides the stability
of clustering structures. This framework has four components as shown in Fig. 1 and
subsequently discussed in the following sections.

Stream

Stream Clustering

Micro-clustering

k estimation
Macro-clustering

Cluster Transition
Detection

Stability Assessment

Fig. 1 Components of the framework

[email protected]
470 K. Namitha and G. Santhosh Kumar

3.1 Online Clustering

Online clustering of the data stream is done using CluStream algorithm [3]. CluS-
tream is one of the most popular stream clustering algorithms available in the lit-
erature. Like many other stream clustering algorithms, CluStream also processes
the stream by using two components—an online component which performs micro-
clustering of the stream and an offline component which creates macro-clusters out of
the available micro-clusters. Micro-clusters are created and maintained in an online
way, and they represent the summary of the stream. Cluster feature vectors [1] are
used to store micro-clusters in memory. To create an overall clustering of the past
stream, macro-clustering can be performed with the micro-clusters relevant to that
period. The main peculiarity of the CluStream algorithm is that it provides the user
with an opportunity to mention the period over which the offline macro-clustering has
to be performed. The proposed framework utilizes this flexibility and it has employed
fixed-size window based macro-clustering.

3.2 Deciding the Number of Clusters Dynamically

CluStream algorithm uses k-means clustering at two different stages—during ini-

tialization of the micro-clusters at the beginning of the stream and during the offline
clustering phase to create macro-clusters from micro-clusters. The usual assumption
while using k-means clustering is that the number of clusters or the value of k is
fixed and it is provided by the user. But there are limitations while applying this
assumption to data streams, as the streams are highly dynamic in nature and concept
changes are possible. To overcome this difficulty, we are using a method of deciding
the value of k dynamically depending on the current characteristics of the stream.
Whenever k-means clustering algorithm has to be called, it is preceded by a step of
calculating the most suitable value for k.
This framework uses Ordered Multiple Runs of k-Means (OMRk) algorithm [14]
as the approach to decide the number of clusters, k. In this approach,
√ k-means algo-
rithm is run multiple times with value of k varying from 2 to N where N is the
total number of points to be clustered. To validate the relative quality of clusterings
created at each execution of k-means, Simplified Silhouette [15, 16] is used as the
relative clustering validity criteria. Silhouette value of a clustering is measured as
follows. Suppose xi is a data element belonging to the cluster Ca , and a(xi ) is the
average distance of xi to all other data elements within the same cluster. Similarly for
each cluster Cb other than Ca , find the average distance between xi and all the points
in Cb . Let b(xi ) be the smallest among these distances. The cluster with this lowest
average dissimilarity is termed as the neighbouring cluster of xi . Low value for a(xi )
and high value for b(xi ) is the desirable feature of a good clustering. Silhouette is
defined as

[email protected]
A Framework for Monitoring Clustering Stability Over Time 471

b(xi ) − a(xi )
s(xi ) 
max{a(xi ), b(xi )}

Higher values of s(xi ) indicate better clustering. In order to reduce the compu-
tational complexity, instead of calculating distance to each data point in a cluster,
distance to the cluster centre is taken. So a(xi ) is the distance between the data point
and its own cluster’s centroid. Similarly to calculate b(xi ), distances between the
data point and all the other cluster centroids are considered. This variant of silhou-
ette computation is called Simplified Silhouette (SS). Average of all the s(xi ) over i
 1, …, N is taken as the SS value of that partition.

1 
N
SS  s(xi )
N i1

Partition with the highest value of SS is the best clustering, and that particular
value of k is chosen. Experiments supported the fact that keeping the value of k
dynamic helps to adapt easily to the changes or evolution of the stream.

3.3 Transition Model

This framework is motivated by the cluster transition models introduced by MONIC

[6] and MEC [7]. It tracks the five categories of external transitions defined in MONIC
namely—survival, absorption, split, disappearance and creation of new clusters.
Clusters are represented by the enumeration method, i.e. a cluster is characterized
by the data points assigned to it. Cluster characteristics like centre, radius etc., are
calculated from its member elements [17].
Macro-clustering is performed after processing each fixed-size window of samples
from the stream. Macro-clusters created at such consecutive time points are compared
to identify the type of transition each cluster has undergone. To decide the survival
of a cluster, a survival threshold value should be provided by the user. Also, a split
threshold is used to decide if a cluster at a particular time point is split into more
than one cluster in the next time point.

3.4 Stability Monitoring

Cluster stability is decided by checking how long it survives. Any transition other
than survival is considered to be the end of the lifetime of that particular cluster.
That is, cluster identity is strictly followed and even if it gets split or absorbed, its
identity is considered to be lost. If all the clusters in a clustering survive to the next
time point, then it is taken as clustering survival. In this case, none of the clusters

[email protected]
472 K. Namitha and G. Santhosh Kumar

disappears and no new clusters get created in between. Since the number of clusters
is allowed to vary according to the dynamic characteristics of the stream, clustering
as a whole survives only if the clustering structure is perfectly consistent. When the
clustering structure gets disturbed slightly, it might be indicating a new emerging
trend or a new event. Experiments are conducted to check this relationship.

4 Experiments and Discussion

Experiments are conducted to check the capability of the proposed framework for
assessing the stability of clusters and clustering over time in data stream mining sce-
narios. Two real-world datasets have been used to conduct the experiments. Results
are visualized by plotting how long clusters and clusterings survive. Stability of a
cluster/clustering is shown by using thick straight line, length of which represents
the duration of stability.

4.1 Datasets

KDD-CUP’99 intrusion detection dataset:

The KDD-CUP’99 intrusion detection dataset is a standard one, available online
and being used as a benchmark dataset in data stream clustering experiments. 10%
version of this dataset is used in our experiments as it is more concentrated than
the original one [18]. Each entry in the dataset refers either to a normal connection
or an intrusion. Intrusion attacks are classified into 22 categories, making a total of
23 classes including the normal class. The dataset contains 494,020 records in total,
each having 41 attributes, in which 34 attributes are continuous and 7 attributes are
symbolic. In these experiments, we have considered only the continuous attributes.
Min-max normalization is performed to prepare the data for experiments [19]. The
stream is generated from the dataset by keeping the same order as the data records
are in the dataset.

Weather data from the Automatic Weather Station:

The second dataset used in the experiments is a weather dataset that contains the
parameters collected by the Automatic Weather Station (AWS) at Advanced Centre
for Atmospheric Radar Research (ACARR) of Cochin University of Science and
Technology, India. Data collected by the AWS at 1-minute interval is used for this
study. From the original data, the most relevant six parameters are identified with
the help of domain experts and these six parameters are used in our study. These
six parameters include temperature, wind speed, wind direction, solar radiation, net
radiation and cloud radiation. Data collected during the period 4 April 2016–20 June
2016 is used for the experiment. This time period is chosen because the south-west

[email protected]
A Framework for Monitoring Clustering Stability Over Time 473

monsoon enters Kerala during the first week of June and notable changes happen in
the atmospheric conditions during this period.

4.2 Experimental Setup

Experiments mainly focused on tracking cluster transitions and hence, monitoring

the historical evolution of clusters and clusterings along with the progress of the
stream. The lifetime of every single cluster, from its birth to disappearance is being
tracked. The disappearance of a long-lived cluster might be giving a hint on an
important physical event happening in the environment that produces the data. In
addition to individual clusters, the stability of the whole clustering is also being
monitored as part of this framework. Obviously, the stability and decay of a clustering
have physical significances. Experiments are conducted to establish this kind of
relationship between the clustering characteristics of the dataset and the physical
events inferred from the data.
As discussed in Sect. 3.1, CluStream is used as the basic stream clustering algo-
rithm in all the experiments. In the offline clustering phase, whenever it uses k-means
clustering, the value of k is allowed to vary dynamically depending on the current
concept that drives the stream. That is, k is not a fixed parameter supplied by the
user. Clustream is a landmark window based [8] clustering algorithm. It gives the
flexibility that the user can mention the period over which the offline clustering has to
be performed. In our experiments, we used windows of fixed size. This window size
is decided on the basis of the characteristics of the dataset. For KDD-CUP intrusion
detection dataset, the window size is set to 5000 and for the weather dataset, it is set
to 1500. Since the weather dataset contains data at 1-minute interval, approximately
1500 data points constitute one day.
Cluster transitions are identified by using the methodology introduced by MONIC
framework [6]. MONIC needs some threshold values to be provided by the user based
on which the category of transition is decided. In our experiments, different values
from 0.50 to 0.80 are tried for the survival threshold and the split threshold is set to
0.25.

4.3 Results and Discussion

KDD-CUP’99 intrusion detection dataset:

This dataset contains 494,020 records in total and Fig. 2 shows the class labels of
these records. Intrusion classes are numbered 1–22 in the order as they are listed in
the dataset description1 and 23 represents ‘normal’ class.

1 http://kdd.ics.uci.edu/databases/kddcup99/training_attack_types.

[email protected]
474 K. Namitha and G. Santhosh Kumar

A stream is simulated from this dataset by keeping the same order as the data points
are there in the dataset. The timestamp of the data records is generated, starting from
zero and incrementing by one. In Fig. 2, class labels are shown against this timestamp
of the stream. Stream is clustered online by using the CluStream algorithm. Offline
clustering phase of CluStream is performed on processing every 5000 samples from
the stream. Clusterings produced by these offline clustering stages are fed as input
to the cluster/clustering stability monitoring framework proposed here.
Results of cluster and clustering survival monitoring processes applied to KDD-
CUP intrusion detection dataset are shown in Figs. 3 and 4, respectively. By cross-
checking Figs. 2 and 3, it can be inferred that clusters survive quite reasonably
for some classes. For some intrusion categories, clusters appear and disappear too
frequently and for some others, they remain stable for a long time. Figure 3 shows
the survival history and hence the stability of individual clusters.
Survival of the whole clustering in KDD-CUP intrusion detection dataset is
depicted in Fig. 4. When all the clusters in a clustering survive over time, it can
be considered as a period of clustering stability. As evident from the figure, there are
two considerably large periods and three small periods of clustering stability while
processing KDD-CUP intrusion detection dataset. It can be observed from Figs. 2
and 4 that the whole clustering remains stable for long duration for class label 18.
Class label 18 corresponds to the intrusion type ‘smurf’. For this category of attack,

Fig. 2 Class labels of KDD-CUP intrusion detection dataset

Fig. 3 Cluster survival in KDD-CUP intrusion detection dataset

[email protected]
A Framework for Monitoring Clustering Stability Over Time 475

clustering structure is consistent and reappearance of the same clustering structure

indicates the reoccurrence of the same category of attack.
Weather data from the Automatic Weather Station:

Data collected by the AWS contains the rain measurements as well. Rainfall during
the period chosen for analysis is shown in Fig. 5. In 2016, south-west monsoon was
preceded by a pre-monsoon shower which starts during the week of May 14.
Atmospheric conditions changed considerably during this period of pre-monsoon
shower and a corresponding change is seen in the clustering structures as well.
Figure 6 shows the cluster survival for this dataset. Most of the clusters remain fairly
stable from April 15 to May 13. Clusters disappear and new clusters get created
during the week of May 14.
Clustering survival shown in Fig. 7 also supports the fact that the whole clusterings
exist fairly stable from April 23 to May 14, i.e. till the beginning of the pre-monsoon
shower. Once the rain starts, the atmospheric conditions come to stability again after
one week.
Experiments support the fact that stability and the sudden disappearance of clusters
and clusterings have relation with the physical events happening in the environment.

Fig. 4 Periods of clustering survival in KDD-CUP intrusion detection dataset

Fig. 5 Rainfall during the period 8-04-2016 to 20-06-2016. Pre-monsoon shower starts during the
week of May 14

[email protected]
476 K. Namitha and G. Santhosh Kumar

Fig. 6 Cluster survival in weather dataset

Fig. 7 Periods of clustering survival in weather dataset

This leads to the conclusion that monitoring the clustering characteristics can help
the prediction of physical events.

5 Conclusion

This paper proposes a framework for monitoring the stability of clusters and cluster-
ings over time when dealing with data streams. Applications that produce continuous
streams of data are becoming very common nowadays. Monitoring the lifetime of
individual clusters and clusterings over the stream and finding its relationship to the
actual events happening in the environment will help the anticipation of upcoming
changes. Stability of the clustering structure has been studied for two datasets with
different nature. Experiments emphasize the fact that stability and decay of clustering
structures have close relationship to the changes happening in the physical environ-
ment. In future, we are planning to use this framework for prediction of creation
and disappearance of clustering structures and thus predict the upcoming physical
events.

Acknowledgements We would like to thank the faculty and technical staff at Advanced Centre
for Atmospheric Radar Research (ACARR), Cochin University of Science and Technology for
supporting this work.

[email protected]
A Framework for Monitoring Clustering Stability Over Time 477

References

1. Zhang, T., Ramakrishnan, R., & Livny, M. (1996). BIRCH: An efficient data clustering method
for very large databases. In Proceedings of the 1996 ACM SIGMOD International Conference
on Management of Data (Vol. 1, pp. 103–114). http://doi.org/10.1145/233269.233324.
2. Guha, S., & Mishra, N. (2000). Clustering data streams. In Proceedings 41st Annual Sympo-
sium on Foundations of Computer Science (pp. 359–366). http://doi.org/10.1109/SFCS.2000.
892124.
3. Aggarwal, C. C., Watson, T. J., Ctr, R., Han, J., Wang, J., & Yu, P. S. (2003). A framework for
clustering evolving data streams. In Proceedings of the 29th International Conference on Very
Large Data Bases (pp. 81–92). http://doi.org/10.1.1.13.8650.
4. Cao, F., Ester, M., Qian, W., & Zhou, A. (2006). Density-based clustering over an evolving
data stream with noise. Sdm, 326–337. http://doi.org/10.1145/1552303.1552307.
5. Ackermann, M. R., Lammersen, C., Sohler, C., Swierkot, K., & Raupach, C. (2012).
StreamKM++: A clustering algorithm for data streams. Journal of Experimental Algorithmics
(JEA), 17, 173–187. https://doi.org/10.1145/2133803.2184450.
6. Spiliopoulou, M., Ntoutsi, I., & Schult, R. (2006). MONIC—Modeling and monitoring cluster
transitions (pp. 706–711). http://doi.org/10.1145/1150402.1150491.
7. Oliveira, M., & Gama, J. (2010). MEC–Monitoring clusters’ transitions. In Proceedings of
the 2010 Conference on STAIRS 2010: Proceedings of the Fifth Starting AI Researchers’
Symposium (pp. 212–224).
8. Silva, J. A., Faria, E. R., Barros, R. C., Hruschka, E. R., De Carvalho, A. C., & Gama, J.
(2013). Data stream clustering. ACM Computing Surveys, 46(1), 1–31. https://doi.org/10.1145/
2522968.2522981.
9. Ghesmoune, M., Lebbah, M., & Azzag, H. (2016). State-of-the-art on clustering data streams.
Big Data Analytics, 1(1), 13. https://doi.org/10.1186/s41044-016-0011-3.
10. Spiliopoulou, M., Ntoutsi, E., Theodoridis, Y., & Schult, R. (2013). MONIC and followups on
modeling and monitoring cluster transitions. In Lecture Notes in Computer Science (Including
Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 8190
LNAI (PART 3) (pp. 622–626). http://doi.org/10.1007/978-3-642-40994-3_41.
11. Ntoutsi, I., Spiliopoulou, M., & Theodoridis, Y. (2009). Tracing cluster transitions for different
cluster types. Control and Cybernetics, 38(1), 239–259.
12. Held, P., & Kruse, R. (2013). Analysis and visualization of dynamic clusterings. In Proceedings
of the Annual Hawaii International Conference on System Sciences (pp. 1385–1393). http://
doi.org/10.1109/HICSS.2013.93.
13. Wicaksono, P., & Manurung, R. (2014). Automatic detection of cluster structure changes using
relative density self-organizing maps (AusDM) (pp. 9–17).
14. Naldi, M. C., Fontana, A., & Campello, R. J. G. B. (2009). Comparison among methods for
k estimation in k-means. In ISDA 2009—9th International Conference on Intelligent Systems
Design and Applications (pp. 1006–1013). http://doi.org/10.1109/ISDA.2009.78.
15. Rousseeuw, P. J. (1987). Silhouettes: A graphical aid to the interpretation and validation of
cluster analysis. Journal of Computational and Applied Mathematics, 20, 53–65. https://doi.
org/10.1016/0377-0427(87)90125-7.
16. Vendramin, L., Jaskowiak, P. A., & Campello, R. J. G. B. (2013). On the combination of relative
clustering validity criteria. In Proceedings of the 25th International Conference on Scientific and
Statistical Database Management—SSDBM (p. 1). http://doi.org/10.1145/2484838.2484844.
17. Leskovec, J., Rajaraman, A., & Ullman, J. (2014). Clustering. Mining of massive datasets
(pp. 240–280). http://doi.org/10.1007/SpringerReference_34708.
18. Masud, M., Gao, J., Khan, L., Han, J., & Thuraisingham, B. M. (2011). Classification and novel
class detection in concept-drifting data streams under time constraints. IEEE Transactions on
Knowledge and Data Engineering, 23(6), 859–874. https://doi.org/10.1109/TKDE.2010.61.
19. Chen, Y., & Tu, L. (2007). Density-based clustering for real-time stream data. In Proceed-
ings of the 13th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining—KDD ’07 (p. 133). http://doi.org/10.1145/1281192.1281210.

[email protected]
Efficient Dynamic Double Threshold
Energy Detection of Cooperative
Spectrum Sensing in Cognitive Radio

Shahbaz Soofi, Anjali Potnis and Prashant Diwivedy

Abstract With the increasing number of wireless users every day, cognitive radio
serves as an approach to solve the spectrum crunch problem. Spectrum sensing
serves as the heart of cognitive radio with it being able to decide if an unlicensed
user is to be given access to a licensed band or not without causing interference.
Various spectrum sensing techniques have been discussed in the literature. In this
paper, energy detection is considered for spectrum sensing, which conventionally
works on a fixed threshold without requiring any prior knowledge of the signal. It is
found the performance of conventional energy detection falls in regions with noise
uncertainty. In this paper, a dynamic double threshold scheme along with cooperative
spectrum sensing at the fusion center is proposed. The dynamic threshold selection
works on the parameter of noise uncertainty for practical cases by creating a noise
variance history. Also, the fusion center uses a dynamic threshold to make a final
decision compared to a fixed threshold for all the energy values lying between the
two thresholds at the local nodes. A simulation model has been discussed to compare
the proposed scheme with traditional energy detection and other detection schemes
as well. A 20% improvement in probability of detection at −22 dB SNR and 0.5
probability of false alarm is achieved using the proposed scheme.

Keywords Cognitive radio · Spectrum sensing · Fusion center · Noise

uncertainty · Probability of detection · Probability of false alarm

1 Introduction

The ever-increasing number of users in wireless communication has led to the short-
age of existing spectrum. However, it is not the shortage but the underutilization of
the spectrum which has led to this problem based on a report by FCC [1]. Cognitive

S. Soofi (B) · A. Potnis

Department of EEE, NITTTR, Bhopal 462001, Madhya Pradesh, India
e-mail: [email protected]
P. Diwivedy
Department of ECE, Allen House Institute of Technology, Kanpur 208008, India
© Springer Nature Singapore Pte Ltd. 2019 479
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_40

[email protected]
480 S. Soofi et al.

radio (CR) has been proposed to accommodate the increasing number of users by
providing dynamic spectrum access to unlicensed users called secondary user (SU)
to a licensed band. This access is provided based on the fact there are no licensed
users called primary user (PU) using that band in this duration to avoid interference
[2].
The first usage of the term “cognitive radio” was seen in 1999 and 2000 by Joseph
Mitola in [3]. The abilities of cognitive radio allow it to be an intelligent system
aware of its surroundings. To make this technology possible, however, the cognitive
radio has to sense if the spectrum is available or not and is not currently being held
by a PU before giving access to an SU. Thereby placing spectrum sensing at the
heart of a cognitive radio network. Such opportunities in which an SU can transmit
without creating any interference are called “Spectrum Holes” [4]. Hence the job
of CR is to efficiently identify these spectrum holes in which SU can transmit their
signals. Also, this allocated band has to be vacated by the SU immediately after its
utilization. At no cost, the PU should incur any interference in its transmission.
Spectrum sensing hence allows the SU to detect the presence or absence of a
primary user in its band. Various spectrum sensing techniques have been proposed in
the literature [5]. Energy detection among all techniques serves as the most simple and
widely implemented technique owing to its easy implementation and not requiring
any prior knowledge of the signal, unlike matched filter. Cyclostationary and matched
filter are techniques explored in literature requiring exact parameters of primary
signal leading to increased complexity and computation at the local nodes [6]. Also,
this information is always not available.
While considering traditional energy detection for spectrum sensing, the measured
energy is compared to a fixed threshold. If the energy is above a fixed threshold it
indicates the presence of a PU. However, if the energy lies below this set threshold it
indicates the absence of a PU allowing the spectrum to be allocated to an SU. With
its simplicity of implementation however traditional energy detection fails in two
scenarios. First in a low SNR region, the CR will not be able to differentiate between
the presence of signal along with noise or just noise due to a similar PSD. Hence
the presence of a fixed threshold will lead to the failure of CR system irrespective
of its sensing intervals [7]. Also if the threshold is set to a low value the CR will
detect the presence of PUs even in their absence leading to an increased error called
probability of false alarm (Pfa ). This error causes more and more missed opportunities
for the SU. On the contrary, if it is set to a high value the SU will not be able to
detect the presence of a PU leading to an increased error called probability of miss
detection (Pmd ) causing more interference. To mitigate the noise uncertainty problem
as well as the threshold value is set to either of the extremes a technique of double
threshold is proposed in the literature [8]. Second, the energy detection suffers from
hidden node terminal problem and multipath fading. Due to a hidden node, there
might not be a direct LOS between the PU and the SU causing the energy level to
drop suddenly indicating the absence of a PU, whereas, it may be present. In such a
scenario, the SU may start transmitting leading to interference. To overcome this CR
can be made to not rely on the decision made by a single secondary node; instead
combines decisions made my several nodes at a fusion center and then make a final

[email protected]
Efficient Dynamic Double Threshold Energy Detection … 481

decision. In this technique of cooperative spectrum sensing, the fusion center makes
a final decision based on certain rules [9].
Owing to noise uncertainty problem, the threshold as we have seen cannot be
kept fixed. Also, the double threshold techniques proposed in literature choose fixed
thresholds making the spectrum sensing problem more difficult [10]. Defines both
thresholds based on parameter ρ called noise uncertainty parameter. It is observed that
ρ is kept fixed, however and cannot be varied (0.5) making it impractical for real noise
environments. In this paper based on the works of [11, 12] working toward a single
dynamic threshold adaption based on noise variance a dynamic double threshold is
proposed which takes noise variance history for N − 1 nodes and calculates noise
uncertainty parameter according to which both thresholds are varied making them
dynamic.
Using cooperation scheme to make a final decision in double threshold method is
proposed in [8, 13]. Although both schemes show improved detection performance
authors fail to address setting of both individual thresholds for cognitive users [14].
For values lying in between the two thresholds referred to as the confused region in
case of detection using double threshold, various schemes have been proposed. In
[8, 10, 13] if the observed energy values lie in between the two thresholds, then no
decision is made and they are forwarded to the fusion center without making any
decision on it. The FC center then works on these values to make a final decision
based on schemes proposed. For example, in [10], the fusion center combines all
such energy values from local nodes and makes a final decision on the presence or
absence by comparing the average energy with a fixed threshold. The fusion center
will now have K local 1-bit decisions where the SUs were able to make a decision
and one decision of the FC for all the values on which no decision was made by
the SUs. The FC combines these two decisions to make a final decision. However,
these schemes make the use of a fixed threshold on the FC as well. To bring noise
uncertainty parameter into the picture at the FC as well, we further propose a dynamic
threshold on the FC based on average noise variance.
To summarize the contributions of this paper, we propose a framework for spec-
trum sensing using a dynamic double threshold based on noise uncertainty parameter.
This noise uncertainty parameter was found to be fixed in our research and not varied
with different noises obtained at the detector. By being able to vary this parameter,
we were able to make both the decision thresholds dynamic and hence more adaptive
to noise.

2 System Model

2.1 Traditional Energy Detector

A traditional energy detector works as a squaring and integrating device where the
output is observed energy value Oi . This observed value is then compared to a
set threshold. For this, consider a deterministic signal say x(n) [15]. The impulse

[email protected]
482 S. Soofi et al.

response of the channel over which the signal is transmitted is say h(n) and w(n) is
the additive white Gaussian noise with zero mean and variance σn2 . The spectrum
sensing problem is modeled as a binary hypothesis testing problem [16] where the
detector has to make a distinction between the following two hypothesis [17]:

w(n) H0 ; PU Absent
y(n)  (1)
x(n)h(n) + w(n) H1 ; PU Present

H 0 is called the null hypothesis when the spectrum is free from any PU transmis-
sion and H 1 is called alternative hypothesis which indicates transmission from the
PU. The test statistic is given as follows


N
X |y (n)|2 (2)
n1

where N is the number of samples and X is the received energy of the signal. Received
energy is then compared with a single threshold (λ). If it is greater than the threshold
H 1 is decided otherwise H 0 as shown in Fig. 1.
While testing for H 0 and H 1 there are two errors which can occur. The first error
Probability of false alarm (Pf ) will occur when decision H 1 is made but H 0 was
true. In this case, the receiver will detect the spectrum to be occupied when in fact it
was free leading to missed opportunities for transmission. The other error involved is
Probability of miss detection (Pmd ) and will occur if the receiver detects the spectrum
to be free and allows SU to transmit signals when in fact it was occupied by a PU
leading to interference. A high value of Pf will lead to spectrum underutilization and a
high value of Pmd on another hand will lead to interference. To describe performance,
we will consider Probability of detection (Pd ) instead of Pmd .

Pd  1−Pmd (3)

A trade-off between the two probabilities is made as a decrease in Pfa will lead
to a decrease in Pd as well and vice versa. This trade-off is made by using the
Neyman–Pearson criterion by fixing Pfa and focusing on keeping Pd as high as

Fig. 1 Single threshold energy detector

[email protected]
Efficient Dynamic Double Threshold Energy Detection … 483

possible. This is because at no cost the PU can suffer from interference. This selection
process is called constant false alarm rate (CFAR) [18]. Using this approach plot
between Pd and Pfa called receiver operating characteristics (ROC) is plotted for
performance.
The central limit theorem can be applied to approximate the test statistic X as N
is usually large. Hence for a fixed threshold λ, the expressions for Pfa and Pd can be
given as [19]
⎛ ⎞
λ − N σ 2
Pfa  P(X > λ|H0 )  Q ⎝   w ⎠ (4)
2N σw4
λ − N (σn2 + σw2 )
Pd  P(X > λ|H1 )  Q (5)
2N (σn2 + σw2 )2

where Q(.) is the Gaussian Q function. The probability of error (Pe ) will be given by

Pe  Pfa + Pmd (6)

Using Eq. (4), we can solve for λ

 √ 
λ  σw2 Q −1 (Pfa ) 2N + N (7)

2.2 Double Threshold Energy Detector

Two thresholds λ1 and λ2 are considered based on noise variance. If the observed
energy values now say Oi exceed upper thresholds λ2 decision H 1 is reported and if
Oi lies below lower threshold λ1 decision H 0 is reported. For the energy values lying
in between the two thresholds, the SU takes no decision and this value to be reported
to the FC as illustrated in Fig. 2.
To define thresholds λ1 and λ2 noise uncertainty parameter ρ is introduced where
0 < ρ < 1. This parameter is based on the practical estimation of noise variance

Fig. 2 Double threshold energy detector

[email protected]
484 S. Soofi et al.

and exists in the interval [(1 − ρ)σn2 , (1 + ρ)σn2 ] taking the varying levels of noise
power incurred due to changing location and time into consideration. In this paper,
ρ is varied based on average noise variance received at each node, whereas in [10,
14], it is set to 0.5. Using ρ the equations for λ1 and λ2 can be given as follows

λ1  (1−ρ)λ (8)

λ2  (1 + ρ)λ (9)

2.3 Cooperative Spectrum Sensing (CSS)

CSS enables a two-stage process for sensing. The local nodes make their decision
and send it to an FC [11, 20]. The FC then combines the decision of each SU using
either OR rule, AND rule or Majority rule. The optimum decision rule is discussed
in [21] taken as OR rule which decides the presence of PU if even one SU reports its
presence. For cooperative scheme, the probability of detection (Qd ) and probability
of false alarm (Qfa ) based on probability of detection (Pdi ) and probability of false
alarm (Pfai ) for ith SUs and N number of sensors is given in [20] as


N
Qd  1 − (1 − Pdi ) (10)
i1
N
Qf  1 − (1 − Pfi ) (11)
i1

3 Proposed Dynamic Double Threshold Energy Detection

Algorithms discussed in the literature which use fixed threshold need the exact knowl-
edge of noise power which is not possible due to changing noise power. The proposed
algorithm works to take the noise uncertainty effect into account based on which the
SU and FC both make decisions.
From [12] which uses average noise variance to vary noise uncertainty parame-
ter in determining dynamic single threshold two threshold levels are varied in this
scheme. By storing noise variances for L − 1 instances, noise variance history is
created. For ith instance, the noise variance will be σi2 (n). For L − 1 instances, the
average variance can be calculated from

1 2
L
σavg
2
 σ (n) (12)
L i1 i

[email protected]
Efficient Dynamic Double Threshold Energy Detection … 485

Using this stored noise variance history, the maximum noise variance σmax
2
is
calculated from the L − 1 nodes and the Lth node as well.
 
σmax
2
 max σ12 , σ22 , σ32 , σ42 , . . . , σ L2 (13)

Then, the noise uncertainty parameter based on average noise variance is defined
as

σavg
2
ρnew  where 0 < ρ < 1 (14)
σmax
2

The dynamic thresholds λ1 and λ2 are defined using noise uncertainty parameter
ρnew as follows

λ1  (1−ρnew )λ (15)

λ2  (1 + ρnew )λ (16)

Then, the observed energy value Oi is compared to λ1 and λ2. If it is found to

be greater than higher λ2 bit “1” is transmitted to the FC indicating the spectrum
is occupied. On the other hand, if Oi is found less than lower threshold λ2 then bit
“0” is reported. If for instance, Oi lies between the two thresholds, the SU makes
no decision and reports this value to the FC. At the FC, two kinds of information
are hence received the local decisions say “K” as “0”s or “1”s and the observed
“N-K” energy values Oi for N nodes. The FC combines these observed N-K energy
values by taking their average and compares with dynamic threshold λ2 instead of
traditional fixed energy detection threshold λ. Here, the upper threshold λ2 is chosen
to increase the reliability and accuracy as λ2 corresponds to upper noise variance.
The simulation model works with the following algorithm:
• Each observed energy value Oi is received at the local node.
• Noise variance is stored and calculated for every observed energy value as σi2 (n).
• Traditional energy threshold λ is calculated using Eq. (7).
• Calculate σavg
2
and σmax
2
using Eqs. (12) and (13).
• Noise uncertainty parameter based on noise variance is found using Eq. (14).
• Dynamic double thresholds λ1 and λ2 are defined for every sensing instance using
Eqs. (15) and (16).
• Each SU makes a decision (L i ) on the presence or absence of PU by comparing
observed energy Oi with λ1 and λ2.

0 O i < λ1
Li  (17)
1 O i > λ2

• If Oi lies in between λ1 and λ2, then the local node makes no decision and reports
the value to the FC.

[email protected]
486 S. Soofi et al.

• Assuming the FC receives K hard decisions from N nodes, it will receive N-K
energy values. The FC center takes the average of all N-K energy values and
calculates Oavg

−K
N
1
Oavg  Oi (18)
N − K i1

• The FC compares the average energy values of the confused region with upper
threshold λ2. If the decision is denoted by M then,

0 Oavg < λ2
M (19)
1 Oavg > λ2

• The FC makes a final decision based on OR rule defined [22] as follows


1 if D + M > 1
F (20)
0 otherwise

The decision F  1 corresponds to hypothesis H1; PU present and F  0 corre-

sponds to hypothesis H0; PU absent. The FC reports the decision back to the SU in
the backward reporting phase informing the SU if the spectrum is free for transmis-
sion or not. From [8], probability of detection Pdnew and probability of false alarm
Pfnew can be derived as
 
Pdnew  P{Oi > λ2|H 1}  Q 2γ , λ2 (21)

where γ  SNR and Q (a, b) is generalized Marcum function.

 
χ μ, λ22
Pfnew  P{Oi > λ2 |H0 }  (22)
χ (μ)

where χ(a, b) and χ(a) are complete and incomplete gamma functions. Also if
Q new
d and Q new
fa represent the probability for detection and false alarm, respectively,
for cooperative sensing, then we have

Q new
d  P{F  1|H1 } (23)

fa  P{F  1|H0 }
Q new (24)

The performance of the proposed algorithm is based on the changing values

of noise variance for practical environments such as AWGN and Rayleigh Fading
Channel. The system focuses on fixing Pfa and maximizing the value of Pd .

[email protected]
Efficient Dynamic Double Threshold Energy Detection … 487

4 Simulation Results and Analysis

For simulations, AWGN communication channel is considered using BPSK signal.

To obtain various samples of energy levels Monte Carlo simulations for 100 samples
is run on MATLAB. To prove the superiority of the algorithm in low SNR plots
between Pd and Pfa called region receiver operating characteristics (ROC) curves are
plotted.
Figures 3, 4 and 5 compare the ROC of traditional energy detection and proposed
scheme for nodes N  10, N  20 and N  50, respectively, at SNR  −22 dB.
Figures 3, 4, and 5 show how the proposed dynamic double threshold can improve
the probability of detection by 0.7 in case of 10 SU nodes, 0.8 in case of 20 SU nodes,
and 0.9 in case of 50 SU nodes in low SNR  −22 dB for the probability of false
alarm of 0.2. For the purpose of evaluation of the proposed scheme at various SNRs
ROC curve at SNR  −28 and −14 dB is shown in Fig. 6 using 10 SUs. Extra
detection probability of 0.05 at Pf  0.2 is achieved with improvements in SNR.
In Fig. 7, a comparison of the proposed scheme at varying SNR and changing
number of secondary user nodes is made using the ROC curve. Due to prediction
nature of the proposed algorithm, large performance improvements are observed in
the probability of detection for both increases in the number of SUs (50) and SNR
(−4 dB).
In Fig. 8, the effectiveness of the proposed algorithm its performance is compared
with other energy detection schemes including conventional energy detection using
fixed single threshold and fixed double threshold at SNR  −22 dB.

Fig. 3 ROC of traditional energy detection and proposed scheme for nodes N  10

[email protected]
488 S. Soofi et al.

Fig. 4 ROC of traditional energy detection and proposed scheme for nodes N  20

Fig. 5 ROC of traditional energy detection and proposed scheme for nodes N  50

[email protected]
Efficient Dynamic Double Threshold Energy Detection … 489

Fig. 6 Proposed scheme ROC at SNR −14dB and −28dB

Fig. 7 ROC of the proposed scheme at varying SNR and changing number of secondary user nodes

[email protected]
490 S. Soofi et al.

Fig. 8 Comparison of the proposed algorithm with other energy detection schemes at
SNR  −22dB

It is interesting to note the proposed scheme offers improved Pd even at low

SNR scenarios using cooperative scheme. Figure 8 shows high Pd up with 20%
improvement at −22 dB SNR for 0.5 Pfa compared to other conventional schemes
proposed in the literature.

5 Conclusion

In this paper, we have presented a dynamic selection of double thresholds which

are varied using noise variance history. This proposed scheme outperforms other
traditional energy detection schemes offering improvement of up to 20% for 0.5 Pfa
at −22 dB SNR. This improves the performance of spectrum sensing, especially
in low SNR. Also for energy values lying in between the two thresholds, the FC
combines all these values and compares them with higher dynamic double threshold
compared to fixed threshold enabling the FC to vary its threshold with previous noise
instances. The storing and forwarding of energy values to FC will lead to increase
in complexity, however, the sensing performance increases by 3.5 times at Pfa  0.2
for this small increase in complexity.
In future, the impacts of the proposed algorithm are to be evaluated for fading
channels. Furthermore, the tradeoffs between an increase in sensing performance
and complexity will be evaluated using practical implementations.

[email protected]
Efficient Dynamic Double Threshold Energy Detection … 491

References

1. Federal Communications Commission. (2002, November). Spectrum policy task force report,
FCC 02-155.
2. Haykin, S. (2005). Cognitve radio: Brain-empowered wireless communications. IEEE Journal
on Selected Areas in Communications, 23, 201–220.
3. Mitola, J., & Maguire, G. Q. (1999). Cognitive Radio: Making software radios more personal.
IEEE Personal Communications, 6(4), 13–18.
4. Tandra, R., Sahai, A., & Mishra, S. M. (2009). What is a spectrum hole and what does it take
to recognize one. Proceedings of the IEEE, 97(5), 824–848.
5. Yucek, T., & Arslan, H. (2009). A survey of spectrum sensing algorithms for Cognitive radio
applications. IEEE Communications Surveys and Tutorials, 11(1), 116–130 (first quarter).
6. Sutton, P. D., Nolan, K. E., & Doyle, L. E. (2008). Cyclostationary signatures in practical
cognitive radio applications. IEEE Journal on Selected Areas in Communications, 26(1), 13–24.
7. Tandra, R., & Sahai, A. (2008, February). SNR Walls for Signal Detection. IEEE Journal of
Selected Topics in Signal Processing, 2, 4–17.
8. Zhu, J., Xu, Z., & Wang, F. (2008, May). Double threshold energy detection of cooperative
spectrum sensing in cognitive radio. In Proceedings of the 3rd International Conference on
Cognitive Radio Oriented Wireless Networks and Communications (CrownCom 2008).
9. Sun, C. H., Zhang, W., Ben Letaief, K. (2007, March). Cooperative spectrum sensing for
cognitive radios under bandwidth constraints. In IEEE Conference on Wireless Communications
and Networking (Vols. 1–9, pp. 1–5).
10. Plata, D. M. M., & Reátiga, Á. G. A. (2012). Evaluation of energy detection for spectrum
sensing based on the dynamic selection of detection-threshold. Procedia Engineering, 35,
135–143.
11. Joshi, D. R., Popescu, D. C., & Dobre, O. A. (2010, March). Adaptive spectrum sensing with
noise variance estimation for dynamic cognitive radio systems. In Proceedings of the 44th
Annual Conference on Information Sciences and Systems (CISS 2010).
12. Farag, H. M., & Ehab, M. (2014). An efficient dynamic thresholds energy detection technique
for cognitive radio spectrum sensing. In International Conference on Computer Engineering
(pp. 139–144).
13. Vien, Q. T., Nguyen, H. X., & Le-Ngoc, T. (2014). An efficient hybrid double-threshold based
energy detection for cooperative spectrum sensing. In 2014 27th Biennial Symposium on Com-
munications (QBSC), Kingston, ON (pp. 42–46).
14. Urkowitz, H. (1967). Energy detection of unknown deterministic signals. IEEE Proceedings,
55(4), 523–531.
15. Ghasemi, A., Sousa, E.S. (2005, November). Collaborative spectrum sensing for opportunistic
access in fading environment. In Proceedings of IEEE Dyspan 2005 (pp. 131–136).
16. Cabric, D., Tkachenko, A., & Brodersen, R. W. (2006, August). Experimental study of spec-
trum sensing based on energy detection and network cooperation. In Proceedings of the First
International Workshop on Technology and Policy for Accessing Spectrum (TAPAS 2006).
17. Poor, H. V. (1994). An introduction to signal detection and estimation (2nd ed.). Berlin:
Springer.
18. Nuttall, A. H. (1975). Some integrals involving the QM function. IEEE Transactions on Infor-
mation Theory, 21(1), 95–99.
19. Xuping, Z., & Jianguo, P. (2007, December). Energy-detection based spectrum sensing for
cognitive radio. In IET Conference on Wireless, Mobile and Sensor Networks (pp. 944–947).
20. Zhang, W., Mallik, R. K., & Letaief, K. B. (2009). Optimization of cooperative spectrum
sensing with energy detection in cognitive radio networks. IEEE Transactions on Wireless
Communication, 8(12), 5761–5766.
21. Sun, C., Zhang, W., Letaief, K. B. (2007). Cluster-based cooperative spectrum sensing in
cognitive radio systems. In Proceedings of IEEE ICC’07 (pp. 2511–2515).

[email protected]
492 S. Soofi et al.

22. Dean, J., & Ghemawat, S. (2008, January). Mapreduce: Simplified data processing on large
clusters. ACM Communications, 51, 107–113.
23. Verma, P., & Singh, B. (2015). Simulation study of double threshold energy detection method
for cognitive radios. In 2015 2nd International Conference on Signal Processing and Integrated
Networks (SPIN), Noida (pp. 232–236).

[email protected]
ConvFood: A CNN-Based Food
Recognition Mobile Application
for Obese and Diabetic Patients

Kaiz Merchant and Yash Pande

Abstract In recent years, obesity and health of diabetic patients have become major
issues. To address these issues, it is very important to know the intake of calories,
carbohydrates, and sugar. We propose a novel deep learning convolutional neural
network based image recognition system that can run on Android smartphones that
not only provides the appropriate nutritional estimates to users after passing a food
image as input but also suggests alternative food recipes for diabetic patients. We
have implemented transfer learning as well as fine-tuning and our CNN model was
able to achieve comparatively higher accuracy than other approaches that used a
similar setup on the Food-101 dataset. By user experiments and approval from well-
known doctors, effectiveness of the proposed system was confirmed. The future
scope includes expanding to more food categories and optimizing the model for
better results.

Keywords Convolutional neural networks · Deep learning · Obesity · Diabetic

patients · Smartphone

1 Introduction

Due to better standards of living, obesity rates are getting higher and its impact
is increasing [1]. To keep a check on this, people must have a daily estimate of the
calories that they consume in every meal. Also, diabetic patients are assigned specific
quantities of sugar and carbohydrates that they are allowed to eat by their doctors.
However, before eating a meal they often do not know the amount of carbohydrates
and sugar that they are about to consume. Furthermore, even if a diabetic patient
can see the nutritional estimates, learning that he or she cannot consume the meal
may alter his or her sentiment in a negative way. Thus, alternative healthier options

K. Merchant · Y. Pande (B)

Dwarkadas J. Sanghvi College of Engineering, Mumbai, India
e-mail: [email protected]
K. Merchant
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2019 493
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_41

[email protected]
494 K. Merchant and Y. Pande

appropriate for diabetic patients should be made available to them. Due to recent
trends in technology and increased usage of smartphones, many mobile applications
have been developed that record everyday meals. Many of them ask the users to
give several text inputs and assume that the users are ready to undergo a tedious
process before getting the required output such as in [2–4]. Some applications have
aimed to carry out the process using food images. However, most of them have
problems related to usability. Also, they do not provide alternative options for diabetic
patients. Our proposed system aims to correct both the above issues by providing a
mobile application that can estimate calories, sugar, and carbohydrate levels from
food images and provide healthier options for diabetic patients.

2 Background

Many methods have been suggested in the literature for recognizing food from images
using hand-engineered features along with traditional image processing techniques
[5–7]. Yang et al. [5] proposed to learn spatial relationships between ingredients
for 61 food categories using pairwise features followed by feature fusion which is
bound to work only for standardized meals. Matsuda et al. [6] used miscellaneous
ranking based approach for recognizing multiple foods. However, this type of solu-
tion is computationally intensive and may not be practically deployable within the
mobile cloud computing platform. Bossard et al. [8] reported classification accuracy
of 50.76% on test set by mining discriminative components using random forests.
Clustering the superpixels of training dataset was done using random forest approach
to train the component models. Random forest was discarded after mining and during
classification multiclass SVM with spatial pooling predicted the final class. Here,
the main problem was with noisy training images not being cleaned and the accu-
racy of predictions being very low comparatively. Liu et al. [9] successfully applied
Inception-model-based CNN approach to two real-world food image data sets (UEC-
256 and Food-101).
There have been a number of available applications that measure daily food’s
dietary information but most of these approaches involve lots of tedious manual oper-
ations to be performed. One example, which is typical of current clinical approaches,
is the 24-hour dietary recall [10]. The issue with this approach is that people have
to remember what they have eaten for a long period and consult a dietician very
frequently. Pouladzadeh [11] proposed a mobile-based application for measuring
calorie and nutrition. The measurement system used the patient’s thumb for first-
time calibration which was then used to calculate the amount of food consumed.
However, the problem is that the density of the food is equally important for calcu-
lating the calorific value. Thus, although many of these approaches have achieved
highly impressive accuracies and tried to measure nutritional value from fast food
recognition and their volume calculation, none of them has aimed to make life of a
diabetic patient easier.

[email protected]
ConvFood: A CNN-Based Food Recognition Mobile … 495

3 Proposed System

3.1 Dataset and Image Preprocessing

We have used ETHZ Food-101 dataset for training which was newly introduced by
Bossard et al. [8]. It contains 101,000 labeled images of food dishes which belong to
101 food categories. Each category has 1000 images which are already rescaled to
have maximum 512 pixels side length. These are the most 101 popular food categories
all over the world which have been sourced from food picture sharing website named
foodspotting.com [12]. We initially make use of the histogram equalization algorithm
to increase the contrast and luminance of images [13]. We implemented this using
the OpenCV library in Python. Moreover, before giving images as input to the neural
network, image augmentation is carried out for entire dataset which includes rotation,
flipping, and scaling the images to 299 × 299 pixels size. Logarithmic correction
[13] has also been applied to input images to enhance the low pixel values with not
much loss of information in high pixel values.

3.2 Food Segmentation Using CNN

Initially, we were motivated to use the bag-of-words approach [14] for the food recog-
nition task. However, CNNs perform the same task as the bag-of-words approach
using a much deeper understanding of the input data [15]. The working of CNNs
[16] can be compared to the functioning of the human brain. A CNN consists of
convolution, pooling, and ReLU layers. A convolution layer searches for features
everywhere in an image by making use of a filter. The work of a pooling layer is to
reduce the size of the image making sure that important features are preserved. The
rectified linear unit eliminates negative values.
Different model architectures exist under the domain of CNNs. These include
LeNet-5 [17], which successfully carried out digit recognition before other ones
did, AlexNet [16], which is similar to but deeper than LeNet, VGG16 [18], which
uses a very large number of parameters for training, ResNet50 [19], that makes use
of skip connections for a deeper network and Inception [20] that combines different
convolutions and reduces the total number of parameters. In our system, we make use
of the Inception network which is shown in Fig. 1. The Inception module is based
on the concept of pattern recognition. After passing several images as input, the
network starts getting used to the small details. As we can see, the Inception network
allows the model to take advantage of all of the convolutions simultaneously. We
make use of the Inception V3 network specifically, which is a 48 layer deep network
as compared to the Inception V1 network which is only 22 layers deep.

[email protected]
496 K. Merchant and Y. Pande

Fig. 1 Inception module

3.3 Nutritional Estimation

Nutritional values for food categories are fetched from the USDA Food Composition
Database [21]. United States Department of Agriculture Agricultural Research Ser-
vice provides us with lab-tested nutritional estimates for different food groups. First,
it gives us nutrient reports which provide lists of food dishes pertaining to different
brands and their respective nutrient values for a specified set of nutrients. Second,
it provides us with food reports which list nutrient values including calories, fats,
carbohydrates, protein, and sugar values per 100 grams for a queried food dish from
the particular food outlet. The nutrient data is obtained after laboratory tests were
conducted on the USDA branded food products database and is thus a certified trusted
source. Being in the public domain there are no permissions needed to use the USDA
food composition data. The suggested citations are mentioned in the references [22,
23].

3.4 Alternative Recipe Recommendation

If diabetic patients feel that they cannot consume food safely after getting the nutri-
tional estimates, they have an option to view alternative healthier versions of the
identified junk foods. This will satisfy their taste buds without compromising their
health. We obtain multiple low and complex carbohydrate recipe recommendations
from Yummly [24]. Ingredients for these are also listed so that patients know what
option suits them best. Moreover, Yummly’s patent-pending food intelligence tech-

[email protected]
ConvFood: A CNN-Based Food Recognition Mobile … 497

Fig. 2 Overall system flow

nology guides users with images and directions to prepare food dishes easily at home.
The nutritional facts for these alternative recipes are calculated using the USDA
nutrient database and displayed for one serving. Thus, both our initial estimation and
alternative recipe nutrition counts reference the same database, and hence the link
between modules is justified.
As shown in Fig. 2, the user takes an image of a food item from an Android device
having active Internet connection. This image is sent to the server, where our trained
CNN model returns predicted probabilities and the food category is recognized.
The user then views the lab-tested nutritional values of selected brand and diabetic
patients can further see a list of alternative healthy recipes.

4 Experimental Results

In the following subsections, we present our experimental setup, model comparisons,

and running application analysis.

4.1 Experimental Setup

TensorFlow provides us with a publicly available Inception V3 based model that

was pretrained on the large ImageNet dataset that contains 1000 classes. Image
recognition with convolutional neural networks on small datasets works better when
fine-tuned on such existing networks. This helps to generalize and prevent overfitting.
Thus, we fine-tune this network and adjust the dimensions of the last softmax layer to
fit our dataset which consists of 101 food categories. The experiment was conducted
using a learning rate of 0.01, which ensures that the training converges at an optimal
point with a moderate pace. We ran the code for about 100,000 iterations using 80%
of the images for training and the remaining 20% for testing. We selected a batch
size of 32 and used the Adam optimizer, which was chosen using a trial and error
procedure.

[email protected]
498 K. Merchant and Y. Pande

4.2 Results and Analysis

The cross-entropy loss, which is a good pick for classification tasks, decreased
over time from 4.5 to 0.96. Initially, as the number of iterations increased, the loss
decreased significantly; however, above 20,000 iterations the decrease in loss became
steadier. Similarly, the accuracy increased slowly after 20,000 iterations. However, it
still did not converge, which shows that the model would have performed even better
if we ran the code for more number of iterations. We were finally able to achieve a
testing accuracy of about 70% for the food recognition task. The accuracy and loss
curves are presented in Figs. 3 and 4, respectively, with a smoothing of 0.9.
The server corresponding to the Android application responds with either one or
multiple categorical probabilistic predictions within 5 s. The user is then accordingly
provided a display of different food groups and can view the lab-tested nutritional
values instantaneously for each of them. No other manual operations are required
and all this happens with just the click of an image. Furthermore, diabetic patients
discover unprecedented healthier version/recipes of the classified food. The ingre-
dients displayed help diabetic patients to strictly obey complex low carb diets. All
this functionality comes in a small Android application of size 6.9 megabytes and

Fig. 3 Accuracy of each step in fine-tuned Inception V3 network

Fig. 4 Cross-entropy loss of each step in fine-tuned Inception V3 network

[email protected]
ConvFood: A CNN-Based Food Recognition Mobile … 499

can easily work on Android devices supporting versions greater than 4.0 (Ice Cream
Sandwich). Finally, the proposed system was approved and provided with positive
feedback from renowned doctors.

4.3 Model Comparisons

Before convolutional neural networks came into the picture, the various approaches
tried out for this task were not able to reach even a moderately high accuracy. In
fact, as can be seen in the comparison table, most of the accuracies were below 50%.
AlexNet, which contains five convolution layers, was able to achieve a decent high
accuracy on the food recognition problem. Among other CNNs, the Inception and
GoogLeNet models have been able to outperform all other ones.
In our proposed system, we have made use of the Inception module, precisely the
Inception V3 network. The reason why we are able to perform better than most other
methodologies is due to the advantages of the way the Inception network is built.
There are three convolution layers to which the input is fed—the 1 * 1 layer, the 3 * 3
layer, and the 5 * 5 layer. Due to this, the network is able to not only capture the
general features but also the local ones. Also, because of this structural benefit, the
depth of the module increases without increase in the number of dimensions. The
model is able to perform well and achieve a considerably high accuracy. However,
Liu et al. [9] were able to achieve an accuracy of about 77%, which is some percent
higher than the proposed model. This is mainly due to the fact that they were able
to run their program for very large number of iterations. We were not able to do the
same due to the lack of resources available. Similarly, Yu et al. [25] were able to get
an output whose value is very close to ours. However, they were able to reach this
accuracy when retraining all layers of the Inception network, while we have retrained
only the final layer. In comparison to other models that were retrained only on the
final layer, our model performs significantly better as can be seen from Table 1.

Table 1 Comparison of various models with accuracies

Method Top-1 accuracy (%)
Bag-of-words histogram [8] 28.51
Improved fish vectors [8] 38.88
Random forest classification [8] 32.72
Randomized clustering forests [8] 28.46
Mid-level discriminative superpixels [8] 42.63
Discriminative components with random 50.76
forests [8]
(continued)

[email protected]
500 K. Merchant and Y. Pande

Table 1 (continued)
Method Top-1 accuracy (%)
GoogleLeNet [9] (10,000 steps) 70.2
GoogleLeNet [9] (300,000 steps) 77.4
Inception V3 [25] (last layer training) 35.32
Inception V3 [25] (full layer training) 70.60
Inception-Resnet [25] (last layer training) 42.69
Inception-Resnet [25] (full layer training) 72.55
AlexNet [26] 66.40
MLDS [27] 42.63
Inception V3 (our model) (last layer training) 70

5 Conclusion

Obesity and diabetes are two of the most common health issues in today’s world. To
tackle these issues, we need a technical assistant that can help us get the nutritional
content of everyday meals that we eat. To solve this problem, we first develop a novel
algorithm for food category recognition which is based on convolutional neural
networks. We perform our experiment on the challenging Food-101 dataset and
are able to achieve a comparatively high accuracy with lesser model parameters
and retraining. However, preprocessing the image with logarithmic correction and
histogram equalization did not help in improving accuracy, which was mainly due to
the fact that most of the feature recognition task had already been carried out by the
starting layers of the pretrained network. Second, to make the life of users easier, we
incorporate our model into a mobile application that provides nutritional quantities
for each image of food clicked and helps assist diabetic patients by providing them
with alternative food recipes. In the future, we hope to incorporate more datasets into
our system in order to increase the number of food categories. We also aim to further
improve the model accuracy and expand our mobile application to provide options
for patients of other health disorders.

Acknowledgements We would like to thank Doctor Harshal Joshi, M. D. and Doctor Sanjay
Gulhane, M. D. for their support and guidance.

References

1. Hruby, A., & Hu, F. B. (2015). The epidemiology of obesity: A big picture. PharmacoEco-
nomics, 33, 673–689. https://doi.org/10.1007/s40273-014-0243-x.
2. MyFitnessPal.com: Free Calorie Counter, Diet & Exercise Tracker, available at http://www.
myfitnesspal.com/.

[email protected]
ConvFood: A CNN-Based Food Recognition Mobile … 501

3. MyNetDiary: The easiest and smartest free calorie counter and free food diary for iPhone, iPad,
Android, and BlackBerry applications, available at http://www.mynetdiary.com/.
4. FatSecret: All Things Food and Diet, available at http://www.fatsecret.com/.
5. Yang, S., Chen, M., Pomerleau, D., & Sukthankar, R. (2010). Food recognition using statistics of
pairwise local features. In 2010 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR). https://doi.org/10.1109/cvpr.2010.5539907.
6. Matsuda, Y., & Yanai, K. (2012). Multiple-food recognition considering co-occurrence employ-
ing manifold ranking. In Proceedings of the 21st International Conference on Pattern Recog-
nition (ICPR 2012), Tsukuba (pp. 2017–2020).
7. TADA: Technology Assisted Dietary Assessment at Purdue University, West Lafayette, Indiana,
USA, available at http://www.tadaproject.org/.
8. Bossard, L., Guillaumin, M., & Van Gool, L. (2014). Food-101—Mining discriminative compo-
nents with random forests. In D. Fleet, T. Pajdla, B. Schiele & T. Tuytelaars (Eds.), Computer
Vision—ECCV 2014. ECCV 2014. Lecture Notes in Computer Science (Vol. 8694). Cham:
Springer. https://doi.org/10.1007/978-3-319-10599-4_29.
9. Liu, C., Cao, Y., Luo, Y., Chen, G., Vokkarane, V., & Ma, Y. (2016, May). DeepFood: Deep
learning-based food image recognition for computer-aided dietary assessment. In International
Conference on Smart Homes and Health Telematics (pp. 37–48). Springer International Pub-
lishing. https://doi.org/10.1007/978-3-319-39601-9_4.
10. Livingstone, M., Robson, P., & Wallace, J. (2004). Issues in dietary intake assessment of
children and adolescents. British Journal of Nutrition, 92(S2), S213–S222. https://doi.org/10.
1079/BJN20041169.
11. Pouladzadeh, P., Shirmohammadi, S., & Almaghrabi, R. (2014, August). Measuring calorie
and nutrition from food image. IEEE Transactions on Instrumentation and Measurement, 63,
1974–1956. https://doi.org/10.1109/tim.2014.2303533.
12. Foodspotting: find and share great dishes, available at http://www.foodspotting.com/find.
13. Chaudhury, S., Raw, S., Biswas, A., & Gautam, A. (2015). An integrated approach of logarith-
mic transformation and histogram equalization for image enhancement. In K. Das, K. Deep,
M. Pant, J. Bansal, & A. Nagar (Eds.), Proceedings of Fourth International Conference on Soft
Computing for Problem Solving. Advances in Intelligent Systems and Computing (Vol. 335).
New Delhi: Springer. https://doi.org/10.1007/978-81-322-2217-0_6.
14. Gao, H., Dou, L., Chen, W., & Sun, J. (2013). Image classification with Bag-of-Words model
based on improved SIFT algorithm. In 2013 9th Asian Control Conference (ASCC), Istanbul
(pp. 1–6). https://doi.org/10.1109/ascc.2013.6606268.
15. Okafor, E., et al. (2016). Comparative study between deep learning and bag of visual words
for wild-animal recognition. In 2016 IEEE Symposium Series on Computational Intelligence
(SSCI), Athens (pp. 1–8). https://doi.org/10.1109/ssci.2016.7850111.
16. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep con-
volutional neural networks. In NIPS (p. 4).
17. Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 86(11), 2278–2324. https://doi.org/10.1109/
5.726791.
18. Simonyan, K., & Zisserman, A. (2015). Very deep convolutional networks for large-scale image
recognition. In ICLR. arXiv:1409.1556.
19. He, K., Zhang, X., Ren, S., Sun, J. (2016). Deep residual learning for image recognition. In
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV
(pp. 770–778). https://doi.org/10.1109/cvpr.2016.90.
20. Szegedy, C., et al. (2015). Going deeper with convolutions. In 2015 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), Boston, MA (pp. 1–9). https://doi.org/10.
1109/cvpr.2015.7298594.
21. United States Department of Agriculture Agricultural Research Service Food Composition
Databases: available at https://ndb.nal.usda.gov/ndb/.
22. U.S. Department of Agriculture, Agricultural Research Service. 20xx. USDA National Nutrient
Database for Standard Reference, Release. Nutrient Data Laboratory Home Page, http://www.
ars.usda.gov/nutrientdata.

[email protected]
502 K. Merchant and Y. Pande

23. U.S. Department of Agriculture, Agricultural Research Service. 20xx. USDA Branded Food
Products Database. Nutrient Data Laboratory Home Page, http://ndb.nal.usda.gov.
24. Yummly: personalized recipe recommendation and search, available at https://www.yummly.
com/recipes.
25. Yu, Q., Mao, D., & Wang, J. Deep Learning Based Food Recognition.
26. Ao, S., & Ling, C. X. (2015, November). Adapting new categories for food recognition with
deep representation. In Proceedings of the IEEE International Conference on Data Mining
Workshop (pp. 1196–1203). https://doi.org/10.1109/icdmw.2015.203.
27. Singh, S., Gupta, A., & Efros, A. A. (2012). Unsupervised discovery of mid-level discriminative
patches. In Computer Vision–ECCV 2012 (pp. 73–86). Berlin: Springer. arXiv:1205.3137.

[email protected]
Segmentation and Recognition of E. coli
Bacteria Cell in Digital Microscopic
Images Based on Enhanced Particle
Filtering Framework

Manjunatha Hiremath

Abstract Image processing and pattern recognitions play an important role in

biomedical image analysis. Using these techniques, one can aid biomedical experts
to identify the microbial particles in electron microscopy images. So far, many algo-
rithms and methods are proposed in the state-of-the-art literature. But still, the exact
identification of region of interest in biomedical image is a research topic. In this
paper, E. coli bacteria particle segmentation and classification is proposed. For the
current research work, the hybrid algorithm is developed based on sequential impor-
tance sampling (SIS) framework, particle filtering, and Chan–Vese level set method.
The proposed research work produces 95.50% of average classification accuracy.

Keywords Image segmentation · Sequential importance sampling · Particle

filtering · Chan–Vese level set method · Minimum distance classifier

1 Introduction

The image processing and pattern recognitions are related to algorithmic develop-
ment and putting abstract region of interest into categories. Generally, the categories
are expected to be known in advance as supervised learning. There are methods
and methodologies to understand the clusters. Current days, the pattern recogni-
tion is useful in so many scenarios (Applications) such as information processing,
document image exploration and identification, forensics, biometric analysis, and
bioinformatics analysis.
Present days, so many real-world problems are automated with pattern recogni-
tion system, where tedious tasks are made easy. For example, in a chemical industry,
checking and verifying the chemical product by human beings are too dangerous
for some extent. By automating the process by computer vision and pattern recogni-
tion overcomes such problems. Biomedical imaging is current buzzword in medical
industry.

M. Hiremath (B)
Department of Computer Science, CHRIST (Deemed to be University), Bengaluru, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 503

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_42

[email protected]
504 M. Hiremath

Fig. 1 Microscopic (SEM)

image of E. coli [11]

As a part and partial of biomedical imaging and processing, microscopy images

are also used. In the current study, the E. coli microscopy images are considered.
Escherichia coli are a gram-negative, facultative anaerobic, rod-shaped and coliform
bacterium of the genus Escherichia that is commonly found in the lower intestine
of warm-blooded organisms (endotherms). Figure 1 shows the generic microscopic
image of E. coli.

2 Materials and Methods

Before proceeding to the computational model, it is necessary to develop a dataset of

stained E. coli images. This section presents abstract technique of staining method
of bacterial gram staining method. This technique is widely used bacteriological
preparation system for microscopic vision. These particular staining methods are
proposed by Dr. Christian Gram in the period of 1884.
In the current research work, Gram-negative staining of E. coli bacteria is con-
sidered. Figure 2 presents the typical staining methodologies.

3 Proposed Method

• Sequential Importance Sampling (SIS) Framework:

The framework requires only one user interaction on the slice where vessel tracking
begins. In the initialization step, the particles and their weights are initialized using
the seed. In the proposed framework, the user needs to roughly select a region

[email protected]
Segmentation and Recognition of E. coli Bacteria Cell … 505

Fig. 2 The staining general method used for microorganisms

Fig. 3 SIS PF framework with four steps: initialization, prediction, update and particle selection

around the vessel on the first slice. Then, the particles are initialized based on the
contour obtained by exploiting the CV algorithm, and the particle weights are set
uniformly. In the prediction step, each particle (contour) moves toward the vessel
boundary on the consecutive slice based on the dynamics. In the update step, the
particle weights are updated based on the particles predicted for that slice. In the
proposed method, the CV algorithm is used to predict the particle changes on the
slice. By combining the updated weights and predicted particles, a particle with the
maximum a posterior (MAP) probability is selected in the particle selection step.
These prediction, update and particle selection steps are repeated for the following
consecutive slices to track along the vessel until vessel tracking terminates [1–9]
(Fig. 3).

[email protected]
506 M. Hiremath

• Particle Filtering:
The particle filtering is a progressive Monte Carlo framework in which the crucial
indication is to characterize the essential subsequent density function by a group
of random samples with related loads (weights). Here, let X t and Z t be the state
and measurement, respectively, at time index ‘t’. The unknown state is usually
estimated by the posterior density function p(X t |Z 1:t ) which can be calculated
recursively using the following Bayesian formula:

p(X t |Z 1:t ) ∝ p(Z t |X t )Z p (X t | X t−1 ) p(X t−1 |Z 1:t−1 )dX t−1 .

In the PF framework, it is approximated recursively using a set of particles. Let Ns

be the number of particles, and nX t (n), ωt (n)oN n  1 s be the set of particles and
corresponding weights describing their relevance [10]. Then, it is approximated
as follows:


Na  
p(X t |Z 1:t )  ωt(n) δ X t − X t(n)
n1

where δ(·) denotes the Dirac delta measure. With this particle approximation,
object tracking is performed by the following procedure:
Prediction: move the swarm of particles based on the dynamics as
 
X t(n) ∼ q X t |X t−1
(n)
, Zt .

Update: calculate the particle weights based on the likelihood as

 
ωt(n) ∝ ωt−1
(n)
· p Z t |X t(n)

where q(·) is the importance density. Note that the selection of importance density
significantly affects PF tracking accuracy. The above-mentioned PF is known as the
sequential importance sampling (SIS) algorithm [11]. If the importance function is
set to q(X t |X t (−i)1, Z t )  p(X t |X t (−n) 1), the SIS algorithm becomes a bootstrap
filter, which is the most common choice. The main advantage of a bootstrap filter is
its simple implementation because it only requires sampling from the distribution
p(X t |X t (−n) 1) and the evaluation of p(Z t |X t (n)). However, it incurs the following
disadvantages.
(1) The prediction step uses only the previous state and does not consider the
observed information, and
(2) A degeneracy problem can arise. In other words, after a few iterations, most
particles have negligible weights; thus significant computational effort is
required to update particles with a very small contribution to posterior density.

[email protected]
Segmentation and Recognition of E. coli Bacteria Cell … 507

• Chan–Vese Level Set Method:

For a variational level set formulation, Chan and Vese proposed region-based
image segmentation [12]. This approach has grown into very widespread in the
field of digital image processing community principally due to its capability to
identify objects not necessarily defined by a gradient. The basic concept of the
segmentation method is to divide a particular partition of a given image I into
two regions, where one region represents the objects to be detected and the other
region represents the background. Then, the contour of the object is defined as
the boundary between these two regions [13]. For a given image I, they proposed
minimizing the following energy functional:
 
E CV (, I )  λ1 (I − c1 )2 H ()dxdy + λ2 (I − c2 )2 (1 − H ())dxdy
 

+v |∇ H ()|dxdy


where  is the image domain, and λ1, λ2 and ν are positive, user-defined eights.
In addition, c1 , c2 and H() are defined as follows:
 
I (x, y)H ()dxdy I (x, y)(1 − H ())dxdy
c1   , c2   ,
H ()dxdy (1 − H ())dxdy


1≥0
H () 
0 Otherwise

where I(x, y) is the pixel intensity and  is the level set function. If we regularize
H(x, y) and δ(x, y) using suitable smooth functions, such as H ε (·) and δ ε (·), (5)
can be minimized using a calculus of variations. The resulting Euler–Lagrange
equation is given as follows:

∂ ∇
 δε () μ · div − ν − λ1 (I − c1 )2 + λ2 (I − c2 )2 .
∂t |∇|

Then, the contour is deformed to the desired boundary using repetitive iterations
of (7) until the energy reaches its local minimum point, or its iteration number
(L) does not exceed the predefined maximum iteration number (L max ). The CV
algorithm has many advantages, such as the ability to perform topology variation
of the contour automatically and to stabilize global region information responses to
local variations such as weak edges and noise [13–16]. However, the performance
of the level set method is limited due to the computational cost of embedding
the contour in higher dimensional space. In particular, for object tracking, it is
difficult to predict drastic shape changes of the object because the algorithm does
not incorporate motion dynamics between frames into the tracking frameworks.

[email protected]
508 M. Hiremath

4 Experimental Results and Discussion

The proposed experiment is carried out using 50 digital images of different types of
E. coli particles of transmission electron microscopy. The experimentation is done
using an Intel quadcore processor @ 2.03 GHz machine. Figure 4 presents the block
diagram of proposed model.
The training and testing algorithms of proposed method are given below:

Algorithm 1: Training phase:

Step 1. Acquire the input training set images (E. coli electron microscope image
set).
Step 2. Transform the present input image into grayscale image.
Step 3. Accomplish preprocessing by using morphological procedures, namely,
erosion, reconstruction and dilation.
Step 4. Achieve segmentation of the image of step 3 using proposed SIS and particle
filtering with Chan–Vese level set method and obtain resulting binary image.
Step 5. Identify and eliminate the border touched cells to obtain binary image of
ROI and then perform labelling the segmented binary image.
Step 6. Calculate geometric shape features for each labelled segment (Major axis,
Minor Axis, Area, Eccentricity, Perimeter, Length/Width ratio, Compact-
ness) and store them.
Step 7. Iterate the steps 1–6 for all the training images.
Step 8. Calculate the minimum feature vector and maximum feature vector of E.
coli particle and store them as knowledge base.

Algorithm 2: Testing phase:

Step 1. Acquire the input: E. coli electron microscope image set.

Step 2. Transform the present input image into grayscale image.

Training Phase

Training Set Image Pre - Proposed Feature Set Generation of

Image processing Hybrid Image Extraction Knowledge
Acquisition Segmentation base
Model

Testing Phase

Test Image Pre - Proposed Feature

Image processing Hybrid Image Selection
Acquisition Segmentation (Minimum Recognition
Model distance)

Fig. 4 Generic block diagram of the proposed method

[email protected]
Segmentation and Recognition of E. coli Bacteria Cell … 509

Step 3. Accomplish preprocessing using morphological procedures, namely, ero-

sion, reconstruction and dilation.
Step 4. Achieve segmentation of the image of step 3 using proposed SIS and particle
filtering with Chan–Vese level set method and obtain resulting binary image.
Step 5. Identify and eliminate the border touched cells to obtain binary image of
ROI and then perform labelling the segmented binary image.
Step 6. Calculate geometric shape features for each labelled segment (Major axis,
Minor Axis, Area, Eccentricity, Perimeter, Length/Width ratio, Compact-
ness) and store them.
Step 7. Apply rule for classification of the E. coli particles: A segmented region is
of E. coli, if its feature values lie in min-max range (3 Sigma classification
rule).
Step 8. Iterate the steps 7 and 8 for all labelled segments and output the identified
E. coli particles.

Cross-validation and Resampling Methods:

For cross-validation and resampling model, the principal necessity for suggested
method is to obtain a number of training data set and validation data set pairs from
the acquired imageset X (remaining some part as the test set). If the dataset sample
“X” is huge enough, then randomly divide it into K parts. After this, randomly
subdivide every part into two and use one half for training and the other half for
validation. So, given a dataset X, generate K training/validation set pairs, {T i , V i }K
i  1, from this dataset. Figure 5 shows the sample training dataset (Figs. 6 and 7;
Tables 1 and 2).

(a) (b) (c) (d)

Fig. 5 Sample training images of E. coli

(a) (b) (c) (d)

Fig. 6 Segmented outcome of region of interest

[email protected]
510 M. Hiremath

Fig. 7 Plot of Receiver Operating Characteristic (ROC) curve for the proposed model

Table 1 The feature vectors of the E. coli particles

E. coli Image pixel Region Particle Particle Length/width
features area eccentricity perimeter circularity ratio of the
region
Particle 1 3165 0.52 367 0.41 1.17
Particle 2 3498 0.67 390 0.47 1.03
Particle 3 3434 0.70 326 0.50 1.09
Particle 4 3456 0.58 311 0.41 1.04
Particle 5 3684 0.59 375 0.39 1.04
Particle 6 3459 0.63 325 0.52 1.07
Particle 7 3698 0.68 300 0.38 1.07
Particle 8 3147 0.59 340 0.44 1.04
Particle 9 3021 0.66 339 0.49 1.17
Particle 10 3089 0.63 378 0.41 1.11

Table 2 The minimum and maximum geometric feature values of knowledgebase

Feature set Image pixel Region Particle Particle Length/width
area eccentricity perimeter circularity ratio of the
region
Minimum 3021 0.52 300 0.38 1.03
Maximum 3698 0.7 390 0.52 1.17

[email protected]
Segmentation and Recognition of E. coli Bacteria Cell … 511

5 Conclusion

Image processing and pattern recognition methods presently used in larger scale to
aid medical experts. The current technologies such as machine learning also inte-
grated to achieve better accurate results. In this research article, we have proposed
a computer-aided E. coli particle segmentation and recognition based on improved
method which comprises SIS, particle filter and Chan–Vese level set method. The
investigational outcomes are matched with the manual outcomes gained by microbi-
ological professionals. The projected model is further optimum and computationally
less complex. It produces a classification rate of 95.50% for E. coli particles. The cur-
rent research work can be enhanced further by improved preprocessing techniques,
feature sets and classifiers, which will be taken up in our future work.

References

1. Caselles, V., Kimmel, R., & Sapiro, G. (1997). Geodesic Active Contours. International Journal
of Computer Vision, 22(1), 61–79.
2. Mumford, D. (1989). Optimal approximation by piecewise smooth functions and associated
variational problems. Communications on Pure Applied Mathematics, 42(5), 577–685.
3. Chan, T., & Vese, L. (2001). Active contours without edges. IEEE Transactions on Image
Processing, 10(2), 266–277. https://doi.org/10.1109/83.902291.
4. Tsai, A., Yezzi, A., & Willsky, A. S. (2001). Curve evolution implementation of the Mumford-
Shah functional for image segmentation, denoising, interpolation, and magnification. IEEE
Transactions on Image Processing, 10(8), 1169–1186. https://doi.org/10.1109/83.935033.
5. Cremers, D., & Soatto, S. (2004). Motion competition: A variational approach to piecewise
parametric motion segmentation. International Journal of Computer Vision, 62(3), 249–265.
6. Goldenberg, R., Kimmel, R., Rivlin, E., & Rudzsky, M. (2001). Fast geodesic active con-
tours. IEEE Transactions on Image Processing, 10(10), 1467–1475. https://doi.org/10.1109/
83.951533.
7. Niethammer, M., & Tannenbaum, A. (2004). Dynamic geodesic snakes for visual tracking.
In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pat-
tern Recognition, 2004. CVPR 2004 (Vol. 1, pp. 660–667). https://doi.org/10.1109/cvpr.2004.
1315095.
8. Paragios, N., & Deriche, R. (2002). Geodesic active regions: A new framework to deal with
frame partition problems in computer vision. Journal of Visual Communication and Image
Representation, 13(12), 249–268. https://doi.org/10.1006/jvci.2001.0475.
9. Avenel, C., Mémin, E., & Pérez, P. (2009). Tracking closed curves with non-linear stochastic
filters. In X.-C. Tai, K. Mrken, M. Lysaker, & K.-A. Lie (Eds.), Scale space and variational
methods in computer vision, no. 5567 in Lecture Notes in Computer Science (pp. 576–587).
Berlin: Springer.
10. Lesage, D., Angelini, E. D., Bloch, I., & Funka-Lea, G. (2009). A review of 3D vessel lumen
segmentation techniques: Models, features and extraction schemes. Medical Image Analysis,
13(6), 819–845. https://doi.org/10.1016/j.media.2009.07.011.
11. Image courtesy: from http://en.wikipedia.org/wiki/Image:EscherichiaColi_NIAID.
jpg—Escherichia coli: Scanning electron micrograph of Escherichia coli, grown in cul-
ture and adhered to a cover slip. Credit: Rocky Mountain Laboratories, NIAID.
12. Blake, A., Curwen, R., & Zisserman, A. (1993). A framework for spatiotemporal control in
the tracking of visual contours. International Journal of Computer Vision, 11(2), 127–145.

[email protected]
512 M. Hiremath

13. Doucet, A., Godsill, S., & Andrieu, C. (2000). On sequential Monte Carlo sampling methods
for Bayesian filtering. Statistics and Computing, 10(3), 197–208.
14. Isard, M., & Blake, A. (1998). CONDENSATION–Conditional density propagation for visual
tracking. International Journal of Computer Vision, 29(1), 5–28.
15. Yilmaz, A., Li, X., & Shah, M. (2004). Contour-based object tracking with occlusion handling
in video acquired using mobile cameras. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 26(11), 1531–1536. https://doi.org/10.1109/TPAMI.2004.96.
16. Shao, J., Porikli, F., & Chellappa, R. (2007). Estimation of contour motion and deformation
for nonrigid object tracking. JOSA A, 24(8), 2109–2121.
17. Pham, D. L., Xu, C., & Prince, J. L. (2000). A survey of current methods in medical image
segmentation. Annual Review of Biomedical Engineering, 2, 315–338.
18. Zaitoun, N. M., & Aqel, M. J. (2015). Survey on image segmentation techniques. Procedia
Computer Science, 65, 797–806.
19. Iglesias, J. E., & Sabuncu, M. R. (2015, August). Multi-atlas segmentation of biomedical
images: A survey. Medical Image Analysis, 24(1), 205–219.
20. Hamuda, E., Glavin, M., & Jones, E. (2016, July). A survey of image processing techniques
for plant extraction and segmentation in the field. Computers and Electronics in Agriculture,
125, 184–199.
21. Khairuzzaman, A. K. M., & Chaudhury, S. (2017). Multilevel thresholding using grey wolf
optimizer for image segmentation. Expert Systems with Applications, 86(15), 64–76.
22. Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A. (2017, December). A survey on deep
learning in medical image analysis. Medical Image Analysis, 42, 60–88.
23. Mesejo, P., Ibáñez, Ó., Cordón, Ó., Cagnoni, S. (2016, July). A survey on image segmentation
using metaheuristic-based deformable models: State of the art and critical analysis. Applied
Soft Computing, 44, 1–29.
24. Sridevi, M., & Mala, C. (2012). A survey on monochrome image segmentation methods.
Procedia Technology, 6, 548–555.
25. Comaniciu, D., Ramesh, V., Meer, P. (2000). Real-time tracking of non-rigid objects using mean
shift. In IEEE Conference on Computer Vision and Pattern Recognition, 2000. Proceedings
(Vol. 2, pp. 142–149). IEEE.
26. Sethian, J. A. (1999). Level set methods and fast marching methods: Evolving interfaces in
computational geometry, fluid mechanics, computer vision, and materials science (Vol. 3).
Cambridge University Press.
27. Kass, M., Witkin, A., & Terzopoulos, D. (1988). Snakes: Active contour models. International
Journal of Computer Vision, 1(4), 321–331.
28. https://www.cliffsnotes.com/study-guides/biology/microbiology/microscopy/staining-
techniques.

[email protected]
Automatic Bengali Document
Categorization Based on Deep
Convolution Nets

Md. Rajib Hossain and Mohammed Moshiul Hoque

Abstract Automatic document categorization has gained much attention by natu-

ral language processing (NLP) researches due to the enormous availability of text
resources in digital form in recent years. It is the process of assigning a document
into one or more categories that help the document manipulate and sort quickly.
An efficient information processing system is required due to the rapid growth of
Bengali text contents in digital form for searching, organizing, and retrieving tasks.
In this paper, we proposed a framework for classifying Bengali text documents us-
ing deep convolution nets. The proposed framework consists of word embedding
and document classifier models. Experiments with more than 1 million Bengali text
documents reveals that the proposed system worthy of classifying documents with
94.96% accuracy.

Keywords Bengali language processing · Document categorization · Word

embedding · Deep convolution neural nets

1 Introduction

Automatic document categorization is a challenging task in the field of NLP where

a text document or a sequence of text documents automatically assigned into a set
of predefined categories. Bengali language is spoken by about 245 million people
in all over the world and is being considered the seventh most spoken language in
the world [1]. Number of Bengali text documents in digital form have grown rapidly
day by day in size and variety due to the increased usability of the Internet. It is very
difficult to manage such huge amount of text documents for human expert manually

Md. R. Hossain · M. M. Hoque (B)

Department of Computer Science & Engineering, Chittagong University of Engineering
& Technology (CUET), Chittagong 4349, Bangladesh
e-mail: [email protected]
Md. R. Hossain
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 513

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_43

[email protected]
514 Md. R. Hossain and M. M. Hoque

that also consume a lot of time and cost of money. Therefore, an automatic document
categorization system will be developed to handle a large amount of text data so that
documents can be organized, manipulated, or sorted easily and quickly.
Although very few research activities have been conducted on Bangla language
processing (BLP) such as syntax analysis, English to Bangla MT, Bangla OCR, and
so on. Bengali text document categorization is also an important research issue that
needs to be solved. There are many linguistic and statistical approaches that have
been developed for automatic documents categorization of English and European
languages. However, no usable and effective system is developed for classifying the
Bangla texts still today. Bangla document categorization system may be used by
security agency to identify the suspected streamed web data or spam detection, the
daily newspapers to organize news by subject categories, the library to classify papers
or books, the hospitals to categorize patient based on diagnosis reports, archiving,
or clustering the government/nongovernment organization data, improve Bengali
content searching, retrieving or mining specific web data, and so on.
In this paper, we proposed a framework for automatic Bengali documents cat-
egorization that works with word embedding model and deep convolution neural
networks (DCNNs) [2, 3]. The proposed model will be able to overcome the tra-
ditional document classification shortcomings and also hardware cost. The word
embedding algorithm such as Word2Vec extracts semantic feature for each word and
represents as 1D vector. The semantic feature carries the actual meaning with respect
to surrounding words and word order. Each document represents a 2D feature vector
where the rows represent the word and columns represent the feature values. There
are lots of hyperparameters in Word2Vec algorithm and we tune these parameters
for Bengali corpus and achieved the better accuracy with respect to other language
corpora. In the recent year, the convolution neural networks (CNNs) achieved the
very good result for English and some other languages [4]. We design a DCNNs
architecture where each hyperparameters will be well trained for Bengali large-scale
data set and generate a classifier model. The model obtained the better accuracy for
categorization of Bengali text documents.

2 Related Work

A significant amount of researches has been conducted on document categoriza-

tion in English and European languages. In the recent year, the Word2Vec [5–7] and
Glove [8] algorithms achieved the state-of-the-art word embedding result for English,
French, Arabic, and Turkish languages. The CNN- and RNN-based documents classi-
fier approaches achieved 84.00 and 85.60% accuracy from English text [9]. Conneau
et al. [3] introduced very deep CNN and achieved 96–98% accuracy from different
English data set classification. In hierarchical, CNN (HCNN)-based and decision-
tree-based CNN also have been achieved higher accuracy for English text [10, 11].
Stochastic gradient descent (SGD) based classifier perform lower accuracy due to
feature scaling and lack of huge hyperparameters tuning [12]. The character-level

[email protected]
Automatic Bengali Document Categorization Based on Deep … 515

CNN is performing slower embedding system and memory consuming [13]. There
are very few researches have been conducted on Bengali text document classifica-
tion. The TF-IDF-based features are the traditional technique which only depends on
documents term frequency or BoW model [1, 12, 14]. Lexical feature only carries
the limited number of information such as average sentence length, average number
of words length, number of different words, and so on [15]. TF-IDF feature performs
lower accuracy due to absence of semantic information. The TF-IDF feature not con-
tained the word position and correlation of the word. The lexical and TF-IDF feature
are not working properly for Bengali language due to its large inflectional diversity
in verbs, tense, noun, etc. In our work, we use DCNNs for Bengali documents cate-
gorization. This approach showed better performance than the previous Bengali text
classification methods due to the hyperparameter tuning and deep network training
architecture.

3 Methodology

The proposed document categorization architecture consists of three main modules:

text to feature extraction module, documents classifier training module, and docu-
ments classifier projection module.

3.1 Text to Feature Extraction Module

Word2Vec with skip-gram algorithm is used to train a Bengali word embedding

model. In order to train the model, we tune the window size and embedding dimen-
sion. The tuned hyperparameter shows the better result for Bengali text classification
purpose. The embedding model row numbers represent the number of distinct words
and columns represent the feature dimension. First split the sentence to word list for
each text document and for each word feature vector is extracted from embedding
model. Therefore, for each document, a 2D feature vector is generated.

3.2 Documents Classifier Training Module

Figure 1 shows the overall architecture of Bengali text document classifier model.
The DCNNs is consisting of input layer, convolution layer, rectified linear units
(ReLU) layer, pooling layer, and fully connected (FC) layer. S(R, C, F) represents
the input and output tensor where R represents the number of rows, C for the number
of columns, and F for the depth or the number of filters.
Input layer: The input document has sequentially passed each word through the
look-up table (Wv·d ) and generates a n × d representation vector, where n represents

[email protected]
516 Md. R. Hossain and M. M. Hoque

Fig. 1 DCNNs layered architecture

the number of word in document and d represents the feature dimension. In our case,
n is fixed to 2200 and d to 300. If the number of word shorter than 2200 then a
padding operation will be imposed. The 2D representation S(2200, 300, 1) is fed to
the input tensor that will be known by the input layer.
Convolution layer: The convolution layer extracts the local feature from the input
tensor and trained the filter weights. The ith layer height and width are calculated by
the following Eqs. (1)–(3).
Dil = D (1)

Hil = Hi−1
l−1
− Dhl + 1 (2)

Wil = Wi−1
l−1
− Dwl + 1 (3)

Here, padding = 0 and stride = 1. Hil , Wil represents the output layer height and
l−1 l−1
width, Hi−1 , Wi−1 represents the input layer height and width, respectively. The D
denoted the number of filter in the current layer. Dhl ,Dwl represents the filter height
and width. Each of the convolution operation follow a Eq. (4).
l l l
 
Hi Wi Di
Sill j l ,d l = j,d × K i, j,d l
Si,l−1 l
(4)
i=0 j=0 d l =0

where K i,l j,d l represents the kernel of each filter. In this architecture, convolution
operation is imposed in three different layers. The first operation is applied at in-
put tensor S(2200, 300, 1) with 16 filters and kernel size (2, 2) with stride is to

[email protected]
Automatic Bengali Document Categorization Based on Deep … 517

one, and output tensor is to S(2199, 299, 16). The second operation is applied with
32 filter and kernel size (3, 3) with stride is to one, and the output tensor is to
S(1097, 147, 32). The last operation is applied with four different kernel sizes such
as Conv1(2, 73), Conv2(3, 73), Conv3(5, 73), and Conv4(7, 73) with zero padding
and stride is to one only for heightwise and 64 filters. The output produces four
tensors with S1(446, 1, 64), S2(445, 1, 64), S3(444, 1, 64), and S4(443, 1, 64).
Rectified Linear Units (ReLU) layer: The rectified linear units (ReLUs) are a
special operation that combines nonlinearity and rectification layers in DCNNs. The
input and output volumes remain same in this layer only changing the elementwise
value.
Sil = max(0, Sil−1 ) (5)

Here, max(0, Sil−1 ) denotes the elementwise operation. ReLU propagates the gradient
efficiently, and therefore reduces the likelihood of a vanishing gradient problem
which in turn reduces the time complexity. The ReLU sets negative values to zero,
and therefore solved the cancellation problem.
Pooling layer: The pooling or feature reduction layer is responsible for reducing
the volume of activation maps. They are used for reducing the computational re-
quirements progressively through the network as well as minimizing the likelihood
of overfitting. In the DCNNs architecture, the max polling operation is imposed in
three different layer. In the pooling layer, the input and output are calculated by Eqs.
(6)–(7).
(H l−1 − K hl + 2P l )
Hl = +1 (6)
Tl

(W l−1 − K wl + 2P l )
Wl = +1 (7)
Tl

where H l and W l are the output tensor height and width of pooling layer. K hl , K wl ,
and T l are the pooling layer polling height, width, and stride size. In the DCNNs
architecture, the padding parameter P is to zero and the first pooling layer input
tensor S(2199, 299, 16) with max pool(2, 2), stride is to two, and the output tensor
size is to S(1099, 149, 16). The second polling layer input tensor is S(1097, 147, 32)
with max pool(2, 2), stride is to 2 and the output tensor size is S(548, 73, 32). In
the last pooling layer, the notation 2-max pool from each region means the pool 2
nonoverlapping feature from each feature map. The output tensor of the pooling layer
having 2 ∗ 4 ∗ 64 = 512 feature.
Fully connected (FC) layer: The fully connected or dense layer is a last layer of
DCNNs. It follows the convolution, ReLU, and max pooling layer. The main goal of
the layer is to design a flatten tensor which is input tensor activation map in three-
dimensional volume. The transformation from input to output tensor of FC layer
is S(R, C, F) → S(R ∗ C ∗ F, 1), where R, C, and F denote the input tensor rows
,columns, and number of filters. In a FC layer, every node in the layer is connected
to each other in the preceding layer. In Fig. 1, the first FC(512, 512) means that the
input tensor S(512, 1) is connected to S(512, 1). The second FC layer F(512, #Class)

[email protected]
518 Md. R. Hossain and M. M. Hoque

means that each of the class got a classification score using Softmax classifier. The
value of the each node in FC layer is calculated by Eq. (8).


M
θil = Wi,l j Al−1
i (8)
j=1

where Wi,l j represents the weight of current layer l and Al−1

i represents the previous
layer activation maps. The θil calculates the projected value.

Ali = F(θil ) (9)

Here, Ali represents the current layer activation maps.

3.3 Documents Classifier Projection Module

In this module, unlabeled text data is taken as input and it is divided into word
and if the input is less then 2200 words then padding is added. For each word is
look-up by embedding model and generate a 2D feature vector where each row
represents the word and corresponding column represents the feature vector. The
DCNNs architecture trained a classifier model (pretrained model) which saves the
layerwise different filter data. The DCNNs is initialized by the trained model and 2D
feature is projected by the DCNNs architecture, and finally produces a score vector
which means that by classwise expected score. The output of FC layer is projected by
a weight matrix W with C distinct linear classification, and the predicted probability
for the ith class given by Eq. (10).
T
(e X Wi
)
P(class = i|X ) = C (10)
c=1 e X T Wc

where X is a ith class feature vector and W is the trained weight matrix. For each
unlabeled text data, the system provides 12 significant score, from this score we
select the max value which is desired class value.

4 Experiments

We implemented DCNNs algorithm in Python with TensorFlow and ran the exper-
iments on a Nvidia GeForce GTX 1070 GPU. This GPU has 8 GB of GPU RAM,
which helps us for extending the networks and batch size. A GPU-based computer
with 32 GB physical memory and Intel Core i7-7700K CPU with 256 GB SSD is

[email protected]
Automatic Bengali Document Categorization Based on Deep … 519

used for experiments. The DCNNs architecture has lots of hyperparameters. The pro-
posed system is suited for better performance with the following parameters: number
of batch size = 32, L2 regularization lambda = 0.0001, number of training epochs=
200, word embedding = Word2Vec, decay coefficient = 1.5, dropout keep probability
= 0.50, save model after this many steps = 100, evaluate model on development set
after this many steps = 100, and convolutions filter initialization method = Xavier
initialization.

4.1 Corpus

Resource acquisition is one of the challenging hurdles to work with electronically

low resource languages like Bengali. Due to the lack of available Bengali corpus
for text categorization, we have built our own corpus and data in the corpus have
extracted from the available online Bengali resources such as blogs and newspapers
[16–19]. The unlabeled data is stored in the corpus for word embedding purposes. We
have collected 836412 Bengali unique words. If a word is not found in the embedding
table, then applying zero-value padding technique. Table 1 shows the summary of
the dataset.
For the classification tasks, handcrafted labeled data are collected from Bengali
online newspapers. Table 2 shows the statistics of labeled dataset. The number of
training and testing documents are varied in each category with variable word length.
The label of the training data is assigned by the human expert that help to achieve
better accuracy. In the label corpus, if the number of words is lower than 2200 then
zero padding is added to each of the input document.

4.2 Evaluation Measures

In order to evaluate the proposed Bengali document categorization system, we used

the following measures: development/training phases loss versus iteration, develop-

Table 1 Embedding data summary

Number of documents 113082
Number of sentence 220000
Number of words 33407511
Number of unique words 836412
Embedding type Word2Vec
Contextual window size 12
Word embedding dimension 300

[email protected]
520 Md. R. Hossain and M. M. Hoque

Table 2 Summary of the categorical data

Category name #Training documents #Testing documents
Accident (AC) 6069 402
Art (AR) 1264 146
Crime (CR) 11,322 1312
Economics (EC) 4526 743
Education (ED) 5159 865
Entertainment (ET) 8558 1734
Environment (EV) 923 110
Health (HE) 2004 580
Opinion (OP) 6479 1248
Politics (PL) 24,136 1834
Science & Technology (ST) 3106 694
Sports (SP) 12,653 1039
Total 86,199 10,707

ment phase/training phases accuracy versus iteration. The testing phase statistical
analysis is shown by precision, recall, F1 -measure, and accuracy.
Training and development phase evaluation: The loss and accuracy of training
and development imply the model beauty or better fitness. The loss and accuracy
is the summation of the errors made for each example in training and development
sets. In each training iteration, the loss is calculated by the following equation:


C
e Wc ∗X i
T
Lossi = −W ∗ X iT + (11)
c=1

where the Lossi means the ith iteration loss mean value and W represents the Softmax
layer weight matrix, where rows and column represent the feature and category val-
ues, respectively. The X i represents the ith example feature vector. The C represents
the total category of our system. The training accuracy is calculated by Eq. (12).

Pi
Acci = (12)
Mi

Here, Acci , Pi , and Mi represent the ith iteration accuracy value, total number of
correctly predicted category, and total number of sample data point in that iteration.
The development loss and accuracy are also calculated using Eqs. (11) and (12).
Testing phase evaluation: In order to measure the overall performance of the pro-
posed system, we used precision, recall, accuracy, and F1 -measure [Eqs. (13)–(16)].
Tp
Precision = (13)
T p + Fp

[email protected]
Automatic Bengali Document Categorization Based on Deep … 521

Tp
Recall = (14)
T p + Fn

T p + Tn
Accuracy = (15)
T p + F p + Tn + Fn

2 ∗ Precision ∗ Recall
F1 -measure = (16)
Recall + Precision

where T p , Tn , F p , and Fn represent the true positive, true negative, false positive, and
false negative, respectively.

5 Results

In order to check the effect of size of training data on the performance, we plotted
two learning curves: loss versus iterations and accuracy versus iterations. Moreover,
we applied the ROC measure for the testing phase accuracy.
– Development and training loss: Figure 2 shows the impact of number of iterations
on loss in development and training phases. In the development phase at iteration
number 1 loss is more than 2.5 and loss is decreased with the increase of iteration.
The decreasing rate of loss is stable at iteration number 20,000 (Fig. 2a). In that
case, we trained on the training data only and test on the tested data. In the training
phase, at iteration number 1 the loss is more than 35 and the loss is decreasing
from the iteration number 5000 (Fig. 2b). This decreasing rate remains stable after

(a) Development loss (b) Training loss

Fig. 2 Losses in development and training phases

[email protected]
522 Md. R. Hossain and M. M. Hoque

(a) Development accuracy (b) Training accuracy

Fig. 3 Accuracy with iteration in development and training phases

10,000 iterations. In this case, we trained on a set consisting of both, training and
development data, and test on the test data.
– Impact of number of iteration on accuracy: Figure 3 illustrates the impact of
number of iterations on accuracy. Figure 3a shows that the development accuracy
is increasing with respect to iteration number and varies from 87.00% (iteration
no. 8000) to 93.00% (iteration no. 65000). However, the accuracy is almost stable
from iteration number 68,000 and remains constant with accuracy about 94.50%.
Figure 3b shows that the training accuracy is increasing with respect to the iteration
numbers.

5.1 Testing Performance

Table 3 shows the precision, recall, and F1 -measure of the proposed system. The
maximum accuracy is achieved by the entertainment (ET) class and the minimum
accuracy is achieved by the environment (EV) class.

5.2 Comparison with Existing Approaches

To evaluate the effectiveness, we compare the proposed system with existing ap-
proaches [1, 5]. Table 4 summarizes the comparison performance. This result shows
that the proposed system outperforms the previous work in terms of accuracy.

– ROC measures: Figure 4 shows the ROC curve for multiclass Bengali documents
categorization. This curve reveals that the maximum area under the curve covered
by both classes SP and ST, whereas the minimum area covered by the EV class

[email protected]
Automatic Bengali Document Categorization Based on Deep … 523

Table 3 Summary of the analysis

Category name Precision Recall F1 -score Support
HE 0.90 0.88 0.89 580
AC 0.90 0.89 0.89 402
AR 0.81 0.83 0.82 146
CR 0.99 0.92 0.95 1312
EC 0.93 0.94 0.94 743
ED 0.98 0.93 0.95 865
ET 0.99 0.99 0.99 1734
EV 0.85 0.62 0.72 110
OP 0.96 0.96 0.96 1248
PL 0.93 0.98 0.96 1834
ST 0.93 0.98 0.95 694
SP 0.94 0.97 0.96 1039
Avg./total 0.95 0.95 0.95 10,707

Table 4 Performance comparison

Method #Training #Testing #Category Accuracy (%)
documents documents
TF-IDF + SVM [1] 1000 118 5 89.14
Word2Vec+K-NN+SVM [5] 19,750 4713 7 91.02
Proposed 86,199 10,707 12 94.96

(0.87). According to area under cover value, ten categories go to excellent class
and only one class (EV) falls into good class and there is no category into bad
class.

6 Conclusion

Text document classification is a well-studied problem for the highly resourced lan-
guages like English. However, it is a relatively new problem for an under-resourced
language, especially in Bengali. Bengali text document categorization is a challeng-
ing task due to the lack amount of digitized text, and scarcity of available corpora. In
this work, we introduce a new technique for Bengali document categorization system
based on DCNNs. In this system, semantic features are extracted by Word2Vec al-
gorithm and classifier is trained by DCNNs. We used 86,199 documents for training
and 10,707 documents for testing and both sets have been handcrafted from online
newspapers. The system obtains 94.96% accuracy for text document classification

[email protected]
524 Md. R. Hossain and M. M. Hoque

Fig. 4 Receiver operating

characteristic (ROC)

and shows the better performance compared to the existing techniques. For future
research, we will extend our framework for other forms of text classification such
as books, blogs, tweets, and online forum threads. Moreover, we will also include
more classes with more data to improve the overall performance of the system.

References

1. Mandal, A. K., & Sen, R. (2014). Supervised learning methods for Bangla web document
categorization. International Journal of Artificial Intelligence and Applications (IJAIA), 5(5),
93–105.
2. Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word represen-
tations in vector space. Journal of CoRR.
3. Conneau, A., Schwenk, H., & Cun, Y. L. (2017): Very deep convolutional networks for text
classification. In The 15th Conference of the European Chapter of the Association for Compu-
tational Linguistics (Vol. 1, pp. 1107–1116), Valencia, Spain.
4. Xu, K., Feng, Y., Huang, S., & Zhao, D. (2015). Semantic relation classification via con-
volutional neural networks with simple negative sampling. In Empirical Methods in Natural
Language Processing (pp. 536–540), Lisbon, Portugal.
5. Ahmad, A., & Amin, M. R. (2016). Bengali word embeddings and its application in solving
document classification problem. In 19th International Conference on Computer and Informa-
tion Technology (pp. 425–430).
6. Johnson, R., & Zhang, T. (2017). Deep pyramid convolutional neural networks for text cate-
gorization. In Proceedings of the 55th Annual Meeting of the Association for Computational
Linguistics (ACL) (Vol. 1, pp. 562–570).
7. Tang, D., Qi, B., & Liu, T. (2015). Document modeling with gated recurrent neural network
for sentiment classification. In Empirical Methods in Natural Language Processing (pp. 1422–
1432), Lisbon, Portugal.

[email protected]
Automatic Bengali Document Categorization Based on Deep … 525

8. Pennington, J., Socher, R., & Manning, C. D. (2014). GloVe: Global vectors for word represen-
tation. In Conference on Empirical Methods in Natural Language Processing (EMNLP) (pp.
1532–1543).
9. Lee, J. Y., & Dernoncourt, F. (2016). Sequential short-text classification with recurrent and con-
volutional neural networks. In North American Chapter of the Association for Computational
Linguistics: Human Language Technologies (NAACL) (pp. 515–520).
10. Kowsari, K., Brown, D. E., Heidarysafa, M., Meimandi, K. J., & Barnes, L. E. (2017). Hierar-
chical deep learning for text classification. In 16th IEEE International Conference on Machine
Learning and Applications (ICMLA) (pp. 364–371).
11. Bahassine, S., Madani, A., & Kissi, M. (2017). Arabic text classification using new stemmer
for feature selection and decision trees. Journal of Engineering Science and Technology, 12,
1475–1487.
12. Kabir, F., Siddique, S., Kotwal, M., & Huda, M. (2015, March). Bangla text document catego-
rization using stochastic gradient descent (SGD) classifier. In 2015 International Conference
on Cognitive Computing and Information Processing (CCIP) (pp. 1–4).
13. Zhang, X., Zhao, J., & LeCun, Y. (2015). Character-level convolution networks for text classi-
fication. In NIPS’15 28th International Conference on Neural Information Processing Systems
(Vol. 1, pp. 649–657).
14. Krendzelak, M., & Jakab, F. (2015). Text categorization with machine learning and hierar-
chical structures. In Proceedings of 13th International Conference on Emerging eLearning
Technologies and Applications (pp. 1–5).
15. Liebeskind, C., Kotlerman, L., & Dagan, I. (2015). Text Categorization from category name
in an industry motivated scenario. Journal of Language Resources and Evaluation, 49(2),
227–261.
16. The Daily Prothom Alo, Online, http://www.prothom-alo.com.
17. The Daily Jugantor, Online, https://www.jugantor.com.
18. The Daily Ittefaq, Online, http://www.ittefaq.com.bd.
19. The Daily Manobkantha, Online, http://www.manobkantha.com.

[email protected]
Artist Recommendation System Using
Hybrid Method: A Novel Approach

Ajay Dhruv, Aastha Kamath, Anuja Powar and Karan Gaikwad

Abstract Recommendation systems have a wide range of applications today in the

digital world. The recommender system must be able to accurately predict the users’
tastes as well as broaden their horizon about the available products. There are various
dimensions in which recommendation systems are created and evaluated. Accuracy
and diversity play an important role in the recommendation systems and a trade-off
must be identified between the two parameters to suit the business requirements. The
proposed system makes use of various recommendation approaches to give a wide
range of recommendations to users. The recommendations are provided based on
similarity of the selected artist, top artists in a genre, using a hybrid model and artists
listened by users’ friends. Some recommendations would include the most popular
ones, and some would be randomly picked for a diverse range of recommendations.

Keywords Collaborative filtering · Diversity · Hybrid model · R · Recommender

systems

1 Introduction

In today’s world of Internet and large-scale growth of e-commerce business, recom-

mender systems play an important role in the growth of the business. This is because
if the website provides products as per the users’ requirements, the customer is bound
to be loyal to that company for a long period of time. Hence, effective recommen-
dations play an important role [1, 2]. The main objective of a recommender system
from the customer point of view is to provide suggestions to online users to make

A. Dhruv (B) · A. Kamath · A. Powar · K. Gaikwad

Department of Information Technology, Vidyalankar Institute of Technology,
University of Mumbai, Mumbai 400037, India
e-mail: [email protected]
A. Kamath
e-mail: [email protected]
A. Powar
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2019 527
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_44

[email protected]
528 A. Dhruv et al.

better decisions from a variety of options available. The main goal from the business
point is to increase the sales and profit [3]. Recommending popular items involves
lower risk in terms of accuracy. However, it is not always advisable to recommend
only the popular ones. Increasing the diversity of recommendations is essential in
businesses. Increasing the diversity and accuracy is not in the same direction, thus
making it important to make ideal decisions between the two parameters [4]. Appro-
priate recommendation model must be chosen so as to maximize the benefits and
maintain customer loyalty by creating value-added relationships. The user content-
ment plays an important parameter in measuring the success of the recommendations
provided [5]. A recommender system must consider diverse factors while provid-
ing recommendations. Some of them include vendor’s target market, data sources,
data availability, scalability, algorithms to be used, and interfaces to manage the
recommendations and data access.
Recommender systems are broadly classified as content-based filtering, collabo-
rative filtering (CF), hybrid approach, and knowledge-based recommender systems.

1.1 Content-Based Filtering

The content-based filtering method (Fig. 1) makes use of content and features of
the item already purchased by the user to recommend items. Content-based filtering
considers how do the items ia and ib relate to each other [6]. The similarity between
two items can be calculated using various methods like Pearson, Jaccard similarity,
and Cosine similarity. The problem with content-based filtering is overspecialization.
Overspecialization is taking into account only those items that are very similar to
each other and giving least importance to the interests of the users. Also, it offers
only partial information, generally text information, and the visual and semantic
information is a challenging task [7].

1.2 Collaborative Filtering

The collaborative filtering (CF) (Fig. 2) approach makes recommendations for a user
with the help of collective preferences of other users [3, 6]. Collaborative filtering

Fig. 1 Content-based
filtering

[email protected]
Artist Recommendation System Using Hybrid Method … 529

Fig. 2 Collaborative
filtering

models are grouped into two types—neighborhood-based and model-based methods.

Neighborhood-based methods include user-based collaborative filtering (UBCF) and
item-based collaborative filtering (IBCF). The UBCF recommends to a user those
items that are most preferred by similar users [1]. The IBCF recommends to a user the
items that are most similar to the users’ purchases. Model-based approaches include
Bayesian clustering, adaptive learning, linear classification, and neural networks
[8]. The user feedback can either be explicit indications or implicit indications.
Implicit feedback gathering does not require active user requirement. Feedback can
be derived from monitoring users’ activities and click-throughs. Explicit feedback
can be gathered by letting users rate the items. Knowing whether user like/disliked
a particular item and text comments can be analyzed. CF, however, suffers from
data sparsity problems, particularly when dealing with large datasets. The cold start
problem occurs when a new user or item has just entered the system. Unless some
information becomes available from reliable sources, it is difficult to find similar
ones. New items cannot be recommended until some users’ rate it. To alleviate the
data sparsity problem, many approaches have been proposed. Some of them are
dimensionality reduction techniques, such as singular value decomposition (SVD),
removing unrepresentative or insignificant users (users who have not given much
ratings/reviews), or items to reduce the dimensionalities of the user–item matrix
directly [9].

1.3 Hybrid

The hybrid system (Fig. 3) uses information from previous user–item interactions and
contents of purchased items. There are seven types of hybrid methods—Weighted,
switching, mixed, feature combination, cascade, feature augmentation, and meta-
level [10].

[email protected]
530 A. Dhruv et al.

Fig. 3 Hybrid model

1.4 Knowledge-Based

There might arise some conditions when collaborative and content-based filtering
does not work. The knowledge-based filtering uses explicit knowledge about the
users and the products [11]. Such systems require knowledge about the group of
items and knowledge about the users [7].

2 Investigating Related Work

Table 1 shows the comparative study of the existing recommender systems.

3 Proposed System

The proposed system aims to implement a multi-featured approach to musical artist

recommendation through the use of UBCF algorithm, similarity matrices, content-
based filtering, and hybrid filtering. The proposed system will address the diversity
and accuracy problems of recommender system which was missing in the earlier
systems. Accuracy and diversity are not in the same direction; hence, the proposed
system will implement a trade-off between these two quantities [4]. The flowchart
(Fig. 4) of the proposed system is as shown.

4 Implementation

The system was implemented using the R language. The “Last.fm” dataset was used
for implementation.

[email protected]
Artist Recommendation System Using Hybrid Method … 531

4.1 Data Acquisition and Cleaning

The artist dataset has been taken for research. The dataset has information about the
artists, users, friends of users, and artists tagged by the users along with its timestamp.
Additional attributes like age and frequency of user listening to a particular artist
were added to improve the quality of the recommendations. The data cleaning is
the process of filling missing values, smoothing the data, removing any outliers, and
resolving any inconsistencies [15]. Thus, the dataset was cleaned and made suitable
for analysis.

4.2 Loading the Data

The data was loaded to RStudio to build a recommender model.

Table 1 Comparison of recommender systems

S. No. Ref. No. Title of paper Technology used Advantages
1 [6] Building a sporting Single value 1. Predicts next
goods decomposition purchase of user
recommendation (SVD), with matrix 2. Considers the
system factorization, purchase
ALS-WR algorithm frequency of items
3. Uses item-based
recommendation
to solve cold start
problem
2 [10] Recommender system SVD algorithm 1. The system
for news articles using considered various
supervised learning parameters like
relevancy,
readability,
novelty, and
time-on-page to
recommend
articles
2. Category of
articles is
considered to
improve
recommendations
(continued)

[email protected]
532 A. Dhruv et al.

Table 1 (continued)
S. No. Ref. No. Title of paper Technology used Advantages
3 [12] Research paper Collaborative filtering 1. Generate
recommendation with algorithm satisfactory
topic analysis recommendations
with few ratings
2. Alleviates cold
start problem
4 [7] Development of a Artificial intelligence 1. The system plans
tourism recommender mechanism and combines
system itineraries with
other cultural and
leisure activities
according to user’s
preferences
2. Considers factors
like place of origin
of user, travel
group, date of trip,
and activities to
recommend places
of interest
5 [8] A personalized Support vector 1. Personalized
electronic movie machine (SVM) recommendation
recommendation classification, 2. Evolution speed
system based on improved particle factor and
support vector swarm optimization aggregation degree
machine and (IPSO) factor used to
improved particle optimize
swarm optimization parameters of the
model
6 [13] A new recommender Collaborative filtering 1. Matrix
system for the algorithm factorization tech
interactive radio to increase
network Fmhost scalability
7 [14] An innovative tour Collaborative 1. Suggests optimal
recommendation filtering, greedy touring plans
system for tourists in algorithm composed of
Japan various points of
interest

4.3 Developing Recommender Engine

The RStudio aids development of recommender systems. It supports various algo-

rithms like UBCF, IBCF, SVD, POPULAR, RANDOM, and Hybrid recommenda-
tions [16]. The dataset was tested using various methods on similarity. Analysis using
different models was also performed.

[email protected]
Artist Recommendation System Using Hybrid Method … 533
Fig. 4 Flowchart of the
proposed system

Figure 5 shows the comparison of ROC curves using three models—UBCF, RAN-
DOM, and POPULAR. ROC analysis helps to pick the best model. The greater the
area under the ROC curve, the better is the model performance. As seen in the graph,
the UBCF outperforms the other models. Though UBCF performs better, due weigh-
tage has been given to RANDOM and POPULAR methods to include a diverse range
of recommendations.
The recommender model provides recommendations by implementing various
algorithms. Many factors are considered for recommendation. A user can get top 10
similar artists by selecting a particular artist (Fig. 6). The similarity was calculated
using similarity() function, and cosine distance was used for measuring similarity
(Fig. 7). The greater the value of the cosine distance, the greater the similarity between

[email protected]
534 A. Dhruv et al.

Fig. 5 Comparison of
various models

Fig. 6 Similar artists to a

selected artist

two artists. The rows and columns represent the unique artists, and the cell value
corresponds to the degree of similarity. The characteristics of the artist are considered
to recommend to the user [17].
The user can get the top 5 recommendations on a particular genre (Fig. 8). The
user tagged data is used to classify an artist in a particular genre. The rows represent
artist IDs, and the columns represent the tag ID (Fig. 9). The cell value shows the
frequency of artist tagged to a particular genre. As the data is sparse in the matrix,
the matrix is binarized to give recommendations.
The user-based collaborative filtering (UBCF) model (Fig. 10) provides recom-
mendations by finding users that are similar to the selected user. From such users,
the top 20 artists are determined by the frequency of tagging of the artists. The artists
that are previously listened by the artist are removed. From the remaining artists, the
top seven are selected for recommendation. The remaining three are selected from

[email protected]
Artist Recommendation System Using Hybrid Method … 535

Fig. 7 Artist similarity

matrix

Fig. 8 Top artists in a genre

the remaining list of artists. These ten recommendations are shuffled and presented
to the user.
The hybrid model uses the linear-weighted method of recommendation. A linear-
weighted hybrid is composed of recommendation components κ1 through κk, whose
output is combined by computing a weighted sum. A linear-weighted hybrid of this
style has a number of advantages like; the recommendations are specialized in a
particular dimension of the data. Thus, the linear-weighted hybrid offers a way to
construct algorithms that take all dimensions of a system into account without requir-
ing mathematically complex and computationally intensive dimensionality reduction
techniques, which are less extensible and flexible [18].
In the hybrid model (Fig. 11), the weights are assigned to POPULAR, RAN-
DOM, and RERECOMMEND methods to provide recommendations. The POPU-
LAR method recommends the trending artists. The RANDOM method selects a
random list of artists. It is used to improve diversity of the recommender system. The
RERECOMMEND method recommends artists from the user’s history. The weights

[email protected]
536 A. Dhruv et al.

Fig. 9 Artist–genre matrix

Fig. 10 Artist recommendation using UBCF

[email protected]
Artist Recommendation System Using Hybrid Method … 537

Fig. 11 Artists recommendation using hybrid model

assigned to POPULAR, RANDOM, and RERECOMMEND are 0.3, 0.6, and 0.1,
respectively. Recommendations are also given considering artists tagged by friends
of the users (Fig. 12). The friends are the ones with whom the user has confidence in
and have similar likings as that of the user [19]. The friends are filtered considering
their age. The artists that are tagged most frequently by those friends and not yet
tagged by the user are recommended.

4.4 Model Evaluation

The most popular measure of effectiveness is accuracy metrics, which is used to

evaluate the ratings a particular user would give to a particular item [3]. The true
positives (TP) are the recommendations provided to the user and liked by the user.
False positives (FP) are the recommendations provided but not liked by the user. The
False Negatives (FN) are those which the system fails to recommend, though the
user would like them. The true negatives (TN) are those which the user does not like
and are not recommended [19].
The Top-N recommendation lists are typically evaluated in terms of their preci-
sion and recall. Precision measures the percentage of recommended products that
are relevant, whereas recall measures the percentage of relevant products that are
recommended [18, 20].
The accuracy of the recommender system is calculated as follows [8]:

The correctly classified samples

Accuracy  (1)
The total number of samples

[email protected]
538 A. Dhruv et al.

Fig. 12 Artist recommendation from friends’ tags

The performance of the UBCF model is shown in Table 2 and that of the hybrid
model is shown in Table 3. From Formula (1), the accuracy of the system is calculated.
The accuracy of the system using the UBCF method is 95%, and using the hybrid
model is 69%.

5 Conclusion

In this paper, a recommender system is implemented using various recommendation

approaches. The system also gives recommendations based on artist similarity, top
artists in a genre, and artists listened by users’ friends. An interactive user interface
is developed to display the recommendations. It is observed that the accuracy of the
system decreases on increasing the randomness of the recommendations. An attempt
has been made to provide a list of accurate and diverse range of recommendations
to the user, using various algorithms.

[email protected]
 Table 2 Performance of the system using the UBCF method
TP FP FN TN Precision Recall TPR FPR
1.29144385 18.70855615 15.55080214 778.44919786 0.06457219 0.08461227 0.08461227 0.02345738
Artist Recommendation System Using Hybrid Method …

[email protected]
539
 540

Table 3 Performance of the system using the hybrid method

TP FP FN TN Precision Recall TPR FPR
1.1149733 8.8850267 24.9919786 770.0080214 0.1114973 0.0428570 0.0428570 0.0113972

[email protected]
A. Dhruv et al.
Artist Recommendation System Using Hybrid Method … 541

6 Future Work

Possible future work with the recommender model could include assessing in an
online environment whether or not the suggested “Artists You Might Enjoy” lists
lead users to explore artists they have not listened to previously. Such an assessment
could be done by tracking click-through rates for those lists. The system could include
URLs of the artists which, upon clicking on the artist recommended, will directly
lead the users on their respective websites. The system could be made dynamic
by including user feedback for the recommendations provided and replacing those
recommendations which the user did not like with new ones. The recommendations
liked by the user could be used for recommending further items.

References

1. Bhumichitr, K., Channarukul, S., Saejiem, N., Jiamthapthaksin, R., & Nongpong, K. (2017,
July). Recommender Systems for university elective course recommendation. Presented at the
14th International Joint Conference on Computer Science and Software Engineering (JCSSE),
Nakhon Si Thammarat, Thailand. https://doi.org/10.1109/jcsse.2017.8025933.
2. Ekvall, N. (2012). Movie recommendation system based on clustered low-rank approximation.
Master’s Thesis, Department of Mathematics, Linkoping University, Sweden.
3. Shakirova, E. (2017). Collaborative filtering for music recommender system. Presented at the
IEEE Conference of Russian, Young Researchers in Electrical and Electronic Engineering
(EIConRus), St. Petersburg, Russia. https://doi.org/10.1109/elconrus.2017.7910613.
4. Javari, A., & Jalili, M. (2014). A probabilistic model to resolve diversity–accuracy challenge
of recommendation systems. In: Knowledge and Information Systems, London (pp. 609–627).
https://doi.org/10.1007/s10115-014-0779-2.
5. Ge, M., & Persia, F. (2017). Research challenges in multimedia recommender systems. Pre-
sented at the IEEE 11th International Conference on Semantic Computing, San Diego, CA,
USA. https://doi.org/10.1109/icsc.2017.31.
6. Flodman, M. (2015). Building a sporting goods recommendation system. Degree project in
Computer Science, Second Level KTH Royal Institute of Technology 26.
7. Ciurana Simó, E. R. (2012). Development of a Tourism recommender system. Master of Science
Thesis, Master in Artificial Intelligence, (UPC-URV-UB).
8. Wang, X., Luo, F., Qian, Y., & Ranzi, G. (2016). A personalized electronic movie recommen-
dation system based on support vector machine and improved particle swarm optimization.
PLoS ONE, 11, e0165868. https://doi.org/10.1371/journal.pone.0165868.
9. Su, X., & Khoshgoftaar, T. M. (2009). A survey of collaborative filtering techniques. Advances
in Artificial Intelligence, 2009, 1–19. https://doi.org/10.1155/2009/421425.
10. Chaturvedi, A. K. (2017). Recommender system for news articles using supervised learning.
MISS Master Thesis, Department of Information and Communication Technologies—UPF,
Barcelona, Spain.
11. Gorakala, S. K., & Usuelli, M. (2015). Building a recommendation system with R: Learn the
art of building robust and powerful recommendation engines using R. UK: Packt Publishing.
12. Pan, C., & Li, W. (2010, June). Research paper recommendation with topic analysis. Presented
at the International Conference on Computer Design and Applications, Beijing, China. https://
doi.org/10.1109/iccda.20105541170.
13. Zaharchuk, V., Ignatov, D. I., & Konstantinov, A. (2012). A new recommender system for the
interactive radio network FMhost. National Research University Higher School of Economic
12.

[email protected]
542 A. Dhruv et al.

14. Le, T. Q., & Pishva, D. (2016, January). An innovative tour recommendation system for tourists
in Japan. Presented at the 18th International Conference on Advanced Communication Tech-
nology (ICACT), Pyeongchang, South Korea. https://doi.org/10.1109/icact.2015.7224843.
15. Jiang, D., & Shang, W. (2017). Design and implementation of recommendation system of
micro video’s topic. In IEEEACIS 16th International Conference on Computer and Information
Science ICIS 3. https://doi.org/10.1109/icis.2017.7960040.
16. Hahsler, M. (2015). Recommenderlab: A framework for developing and testing recommenda-
tion algorithms. CRAN, 40.
17. Dou, Y., Yang, H., & Deng, X. (2016, August). A survey of collaborative filtering algorithms
for social recommender systems. Presented at the 12th International Conference on Semantics,
Knowledge and Grids, Beijing, China. https://doi.org/10.1109/skg.2016.014.
18. Gemmell, J., Schimoler, T., Mobasher, B., & Burke, R. (2012). Resource recommendation
in social annotation systems: A linear-weighted hybrid approach. Journal of Computer and
System Sciences, 78, 1160–1174. https://doi.org/10.1016/j.jcss.2011.10.006.
19. Ramesh, B., & Reeba, R. (2017, April). Secure recommendation system for E-commerce web-
site. Presented at the International Conference on circuits Power and Computing Technologies
[ICCPCT], Kollam, India. https://doi.org/10.1109/iccpct.2017.8074240.
20. Paraschakis, D., Nilsson, B. J., & Hollander, J. (2015, December). Comparative evaluation of
Top-N recommenders in e-Commerce: An industrial perspective. Presented at the IEEE 14th
International Conference on Machine Learning and Applications, Miami, FL, USA. https://
doi.org/10.1109/icmla.2015.183.

[email protected]
Anomaly Detection of DDOS Attacks
Using Hadoop

Y. S. Kalai vani and P. Ranjana

Abstract A distributed denial of service is one of the major threats in the cyber
network, and it causes the computers to have been flooded with the HTTP GET
packet. As Http flood attacks used standard URL requests, it is quite challenging to
differentiate from valid traffic. In an Http flood, the Http clients such as web browser
interact with the application or server to send Http requests. The request can be either
GET or POST. The aim of attack is when to compel the server to allocate as many
resources as possible to serving the attacks, thus denying legitimate user’s access to
the server’s resource. To handle this DDOS attack, the traditional intrusion detection
system is not suitable to hold and find the huge amount of data in the network.
Hadoop is a framework that allows processing and storing huge datasets. In Hadoop
MapReduce is the programming model to process huge data stored in Hadoop. This
paper explains how to detect the HTTP flood attack using KDD CUP 99’ dataset
with improvised the algorithm in MapReduce using anomaly detection strategy.

Keywords Hadoop · HTTP GET · MapReduce

1 Introduction

The cyberattack denial-of-service is a major threat in the Internet, and it causes many
serious problems in the network. Distributed denial-of-service (DDOS) attacks come
under the category of denial-of-service attacks. A DDOS is a cyberattack which is
caused by executor using more than one IP address and flooding the request to the
victim to attack the regular response. An HTTP GET attack is an application layer
attack which consists of low and slow attacks, the target Apache, and Windows
servers. This type of attack sends a large number of request which consists of unused

Y. S. Kalai vani (B)

Department of Computer Applications, Sindhi College, Bangalore University, Bangalore
560024, India
e-mail: [email protected]
P. Ranjana
Department of CSE, Hindustan University, Padur, Chennai 600002, India

© Springer Nature Singapore Pte Ltd. 2019 543

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_45

[email protected]
544 Y. S. Kalai vani and P. Ranjana

packets to the web server to slow down the server which are in an active state and it
also shuts the server by exhausting the bandwidth limits of network.
Network intrusion detection system is a software tool which is used to detect any
malicious behavior in the network. It consists of two-part signature-based detection
which is used to detect the malicious information based on the signature, that is,
known attacks are detected in the network. Anomaly-based intrusion detection is
used to detect the anomalies in the network which is not known before, that is,
usual behavior in the network. The HTTP flooding attack is detected by the anomaly
intrusion detection system using the Hadoop environment. Hadoop is a platform
which is used to store huge amount of data in the form of zeta bytes, so this Hadoop-
based IDS is used to hold and detect huge volume of data in the network. The dataset
used here is KDD CUP 99’ [1] which is used to find the HTTP GET flooding rate using
counter-based algorithm in MapReduce, using the parameter values which consists
of some important classes such as normal, ICMP, TCP, UDP, Port scan HTTP, and
IP. There are number of parameters used to check the anomalies in network such as
protocol type, source bytes, destination bytes, HTTP, session rate, and counter IP.
Proposed system follows the improvised counter-based algorithm which analyzes
the packets and filter the packets based on the request types. If the request HTTP
GET is more than the threshold, then that attack is filtered and emitted. Detection
of anomalies in the network sues the preprocessing technique in the counter-based
algorithm.
Hadoop provides the platform to achieve the efficient way of detecting HTTP GET
flooding [2] attacks using the MapReduce scheme. It is used to calculate the number
of packets per second in the network; if it exceeds the threshold, the network traffic
flow is more. This impact causes decrease in speed of the computer. Counter-based
algorithm is used to measure the parameters in the log file. If the protocol is HTTP,
then attack is confirmed to HTTP flooding which exceeds the threshold and then the
log files ignore anomalous attack.

2 Related Work

As the vulnerabilities are increased in the network in the form of cyberattacks, use
powerful IDS to detect the anomalies in the network. Many research papers are given
the information about the cyberattacks and detection methods. This research paper
focuses on DDOS attacks and its different types. A DDOS is an attack derived from
DOS. DOS attack is a malicious attempt by a single user or group of user to cause the
victim, site, or server to deny service to its customers. A distributed denial-of-service
is an attack in which multiple computer systems attack a target such as web server
or network. Due to this attack sometimes, the web server may crash, not accepting
any request from the user side. To handle DDOS attack, the traditional IDS is not
suitable to hold the number of packets and to detect the attack. The network anomaly
intrusion detection system in the Hadoop framework is used a tool to detect the HTTP
GET flooding attack.

[email protected]
Anomaly Detection of DDOS Attacks Using Hadoop 545

2.1 Categories of DDOS Attacks

Volume-based attack

It aims to saturate the bandwidth of the affected website, and its magnitude is mea-
sured in bits per second. It consists of user datagram packet (UDP) flood, Internet
message protocol (ICMP), and spoofing of IP address. Volume-based attacks are
measured by bits per second.

Protocol Attacks
Protocol is a set of constraints used to transmit the data in the form of packets. It
comprises the attacks such as SYN FLOOD, fragmented packets, ping of death,
Smurf DDOS, etc. SYN flood attack creates the attack by sending the SYN request
which contains malicious code to attack the server. Protocols are measured in packets
per second.

Application attacks

Application attacks are low and slow attacks which target the servers like Apache,
Window, or OpenBSD vulnerabilities. This attack sends the request to server through
HTTP GET/POST methods, which contains malicious code that creates the vulner-
abilities in the server.

2.2 Methods to Detect the DDoS Attacks

This paper focuses on the application layer attack called HTTP GET flooding. This
HTTP flooding attack can be detected by several approaches are used in intrusion
detection scheme. Many algorithms are used to detect the intrusions in the network-
based; on this category among all this, research paper focuses on MapReduce algo-
rithm which is used to solve the problem of existing system drawbacks.
• Statistical method methods [1] are used to identify the detection by using entropy
and Chi-square test by comparing the values of packet traffic coming from the
clients.
• Support vector machine is a method which consists of a set of efficient data mining
algorithm that comprises the attributes taken part in the network anomalous attack.
• Soft computing method is used to find the DDOS attack by specifying the neural
network strategies to detect the HTTP Food attacks.
• Hadoop-based method is a framework which is used to detect the HTTP GET [3]
flood attack using the MapReduce algorithms.

[email protected]
546 Y. S. Kalai vani and P. Ranjana

2.3 HTTP GET Flood Attack

An HTTP GET flood attack is an application layer attack which damages the system
information in the application level. This attack sent along with HTTP GET request
which contains malicious code to damage the server. It is low and slow, but it creates
a lot of damage in the TCP connection between client and user.

2.3.1 Network Attack Layer

HTTP flood attack is an application layer attack which sends the request from client to
server after the TCP connection establishment. Attackers implementing their HTTP
flood attack on web servers and its resources so the users unable to access the
resources and server also. The following diagram explains about a single HTTP
GET request attack, and how it slows down the server as it threshold is less than the
traffic Rate.
Figure 1b explains how the flooding HTTP GET requests in a TCP connection that
is multiple HTTP GET requests are coming from the client side for one single TCP
connection. As the load is increased and it contains malicious info in the request,
server cannot handle the request in specific format so it is unable to process the user
request.

(a)
CLIENT SYN ACK GET RST SYN ACK GET RST

SERVER

SYN-ACK SYN- ACK

(b)
CLIENT SYN ACK GET GET GET

SERVER
SYN-ACK Response Response Response

Fig. 1 a A single HTTP GET request attack in TCP connection. b A multiple HTTP GET request
attack in TCP connection

[email protected]
Anomaly Detection of DDOS Attacks Using Hadoop 547

2.4 Intrusion Detection Techniques

An intrusion detection is a software [4] application which monitors a network for

malicious activity. Any type of malicious activities is collected and sent to the admin-
istrator; security issues are taken place to fix an alarm when the abnormal activities
are occurred in the network.
Intrusion detection is classified into three types.

Host-Based Intrusion Detection System:

It is an intrusion detection system which monitors the internal part of the computer
system and monitors the network packets in its network Interfaces.

Hybrid Intrusion Detection System:

Network-based intrusion detection system is a combination of host-based intrusion

detection system and network-based intrusion detection system which monitors the
computer’s internal activities and network activities.

Network-Based Intrusion Detection System:

A network intrusion detection system is an intrusion detection system that attempts

to discover unauthorized access to a computer network and detect the cyberthreats
in the network.
A network intrusion is classified into two types:
1. Misuse detection (Signature-based) and
2. anomaly detection (profile-based).
Signature-based detection: This type of mechanism requires signature-based
files, i.e., known patterns of attacks. If a pattern is matched, there is a possibility of
an attack and then an alarm will be triggered. It has low false-positive alarm rate as
matches are based on known patterns. Signature detection [1] fails to detect attacks
which are variations in the known attacks.
Anomaly-based detection: This type analyzes computer and network activities
and looks for an anomaly; if an anomaly is found, alarm is triggered. Anomaly is
abnormal behavior of network or deviation from common rule for attacks. It has high
false-positive alarm rate. The rise in use of technology has led to increase in amount
of network traffic data. The traffic data is expected to be in the range of Zettabytes
[2] and is significantly increased from past few years. This data having high Volume,
Velocity, and Variety is often termed as Big Data. In this generation of Big Data, the
IDS should be good enough to process huge volume of data in real time.

Unsupervised Anomaly detection Approach:

It is a type of anomaly detection approach which is the most flexible setup which does
not require any label; moreover, there is no distinction between a training and test
dataset. This idea behind an unsupervised anomaly detection algorithm scores the

[email protected]
548 Y. S. Kalai vani and P. Ranjana

data exclusively based on intrinsic properties of the dataset. Distances and densities
are used to give estimation which is normal and what is an outlier or anomalies.
This research paper proposed a model of anomaly detection system using unsu-
pervised data in KDD cup dataset. We have the test dataset which are not having
class labels which are the example for the unsupervised dataset.

Unlabeled Data Unsupervised Detection

Anomaly detection Algorithm Result

3 Proposed System Design and Implementation

3.1 Hadoop Frame Work

Hadoop is a framework which is used to process and store huge amount of data.
KDD CUP99’ is taken as dataset, and it is preprocessed by the Hadoop framework to
avoid the duplicate records. Here, we have taken the dataset for DDOS attack which
consists of a set of classes with its attributes. Anomaly-based intrusion detection
technique is implemented to detect the abnormal activities. HTTP flood attack is
caused due to many requests sent by the same user or from the different IP addresses
which creates a traffic in the network and slows the process of the server. To handle
the huge amount of IP address in the network, the Hadoop framework is taken as a
proposed framework.
Hadoop is a framework which allows to store huge volume of datasets and to
process the dataset. Hadoop framework is divided into two types such as processing
and storage. MapReduce is a programming model which is used to process huge
data stored in Hadoop. When the used install Hadoop in a cluster, the MapReduce
will be provided as a service. In MapReduce, the user can write the programs to
perform some computations in parallel and distributed fashion. MapReduce [3] is a
programming framework and it takes the dataset as a huge volume in a distributed
environment.
MapReduce is a programming framework that allows us to perform distributed and
parallel processing on large datasets in a distributed environment. In this proposed
system, we have taken KDD CUP 99’ dataset for DDOS with specific classes and a
set of attributes. MapReduce consists of two tasks, namely, Map and Reduce. First,
the DDOS dataset is taken for sample and it is processed by the mapper. The output of
the map work is key–value pairs of the dataset which does not contain the redundant
data. The output of map job is the input of the reducer. The reducer aggregates
the values, and applying the strategy gives the result as in the form of intermediate
format.

[email protected]
Anomaly Detection of DDOS Attacks Using Hadoop 549

Software configuration for this process is Apache Hadoop, and it is open-source

software utilities that facilitate using a network of many computers to solve the
problems of enormous amount of data and its computation. The core part of the
Apache Hadoop [4] consists of a storage part which is called as Hadoop Distributed
File System (HDFS), and processing model is considered as MapReduce.
Hadoop Modules:

The Apache Hadoop is a framework which consists of modules, namely, Hadoop

Distributed File System and Hadoop YARN [5] which is responsible for managing
computing resources in clusters and using them for scheduling user’s applications;
Hadoop MapReduce consists of the programming model for large-scale data pro-
cessing. Hadoop framework is written in Java Programming Language and some
code is written in C language.

3.2 MapReduce

MapReduce is a programming model and an associated implementation for process-

ing and generating big datasets with a parallel, distributed algorithm on a cluster.
A MapReduce program is composed of map procedure which performs filtering
and sorting [5], reduce method which performs a summary operation. MapReduce is
composed of a map procedure which performs filtering and sorting so the detection
algorithm which is used to separate the malicious activity from normal activity.

3.3 Intrusion Detection in MapReduce

HTTP GET flooding is considered as one of the serious HTTP flooding attacks. The
attackers send the massive HTTP GET request to the victim server until the server
and its resources are get exhausted. Several techniques are introduced to solve the
problem such as threshold determination, pattern analysis, traffic analyzer, etc. HTTP
flood attack is an application layer attack that causes damages in web applications and
web servers. The requests are in the form of HTTP GET or POST method flooded
[6, 7] to the target web server which is not handled by the web server. Requests
coming from different IP addresses which causes the serious problem so the efficient
algorithm is used to detect the intrusion in the network. The intrusions are in the form
of anomalous behaviors and it should be classified into supervised and unsupervised
data format.
To handle unsupervised data, the classification analysis is used and to handle
supervised data the clustering algorithm is used (Fig. 2).

[email protected]
550 Y. S. Kalai vani and P. Ranjana

Map Function Reduce Function

Capture the Packet
Request count, response count per server

Sum of Request count, response count per

Filter HTTP Packets
server

Detection Algorithm
KEY: Server IP, Client IP, Timestamp
VALUE:URI count, request count,
Response count KEY: ServerIP, Timestamp
Value: Clinet IP

Fig. 2 MapReduce for HTTP GET flood attack detection

4 Proposed System and Design for DDOS

In this research work, the anomaly detection system which consists of KDD CUP
99’ used as a dataset. Anomaly detection is used to find any abnormal activities in
the process. Hadoop is a platform which takes a data in a huge format and stores the
data in the MapReduce [8]. It is open-source software which is used to detect any
anomalies in the network. If servers are getting affected with the overloaded request,
so server is unable to process other requests in the network (Fig. 3).
Algorithm 1: Packet Analysis Algorithm using Map Function

Step 1: Read the dataset in KDD CUP 99’.

Step 2: Preprocess the dataset for further processing.
Step 3: Filter the selected feature in the KDD CUP 99’.
Step 4: Take the test dataset.
Step 5: Filter the HTTP GET dataset.
Step 6: Tabulate the result including server IP and client IP address as key and value.

KDD CUP 99 PREPROCESS THE Training Data Set Filter : HTTP GET
DATA SET DATA request

Accuracy comparison Reduce Function:

between existing system HTTP GET Attack Input HTTP GET
and proposed system detection algorithm Packets, IP address
(HAGA)

Fig. 3 Processing steps in proposed system

[email protected]
Anomaly Detection of DDOS Attacks Using Hadoop 551

Algorithm 2: HTTP GET Attack Detection Algorithm (HAGA) using Reduce

function
Input: c: KDD CUP 99’ data set,<key,value>as
< server IP, packet count >,
< client IP, request Count >,
< masked of timestamp, response count >
Output: emits the attacker’s IP Address

Step 1: procedure Packet Analyzer using Map Reduce class

Step 2: Generate <key, valye> till the end of the file
Step 3: Aggregate <key, value> as
<serverIP, packetcount>
<clinetIP,requestcount>
<maskedTimestamp, responsecount >

Step 4: calculate Detection Accuracy = TP+FN/TP+FN+FP+TN

Step 5 : If total request > threshold

Detects the attacker’s Address
Else
Continue.
End if

Result:
In this research paper, the improvised algorithm for DDOS attack consists of a KDD
CUP 99’ dataset and attack is generated to server. Different types of attacks such
as ICMP flooding, TCP flooding, UDP flooding, Smurf flooding, port scan, land
flooding, HTTP flooding, session flooding, and IP flooding are generated and log
files are created for every 30 min. To detect attack, attributes are derived from the
log files and from the derived attributes dataset is created with nine different types
of attack classes such as ICMP, TCP, UDP, and SMURF. Improvised counter-based
algorithm for DDOS attack consists of dataset, and its attribute based on the HTTP
GET attack detection is a simple method that counts the total traffic volume or the
number of web page requests [5]. The algorithm is used to detect the false alarm rate
in the network attack with the low false alarm rate. This algorithm needs three inputs
parameters: time interval, threshold, and unbalance rate. The above algorithm is
calculated based on the threshold value in the network. This algorithm calculates the
false alarm rate base on the threshold rate and request, response rate. The execution
time for the proposed algorithmic approach is around eighty seconds (80 s) which is
comparatively less than the total time of the attack scenario in which the experiment
was performed. So it can be concluded that this algorithmic approach is a near real-
time one (Table 1).

[email protected]
552 Y. S. Kalai vani and P. Ranjana

Table 1 Table of detection accuracy

False alarm Detection data
Normal data Detection data Normal data Detection data
KDD CUP 99’ 0.06 0.02 0.01 0.933

5 Conclusion

This paper focuses on the detection of intrusion detection in network in the form of
HTTP GET FLOOD attack which is detected by the Hadoop because the traditional
IDS is not suitable for the detection of huge amount of data, so the Hadoop is a
framework which processes the huge amount of data. To handle the detection in
network, the tool SNORT is used to find the network traffic in the network and
it is used to identify the flooding attack for the server and it detects the traffic in
the network. So the improved snort is used to detect the HTTP FLOOD attack in
the network in an efficient manner, and it follows the counter-based algorithm in
MapReduce which is used to process the huge amount of data and is capable of
detecting the attacks in the network.

References

1. Kaskar, J., Bhatt, R., & Shirsath, R. (2014). A system for detection of distributed denial of
service (DDoS) attacks using KDD cup data set. International Journal of Computer Science
and Information Technologies (IJCSIT), 5(3), 2014, 3551–3555. Department of Information
Technology, Dr. D. Y. Patil College of Engineering, Pune, India.
2. Department of CSE, TCE Madurai, India, [email protected], [email protected],
[email protected] 978-1-4673-0671-3/11/$26.00©2011 IEEE.
3. Navale, G. S., Kasbekar, V., Ganjepatil, V., & Bugade, S. (2014). Detecting and analyzing DDos
attack using Map Reduce in Hadoop. Sinhgad Institute of Technology and Science, Narhe, Pune
41 University of Pune, India.
4. Zhiguo Shi, J. A. (2015). An intrusion detection system based on Hadoop. IEEE.
5. Mazhar Rathore, M., & Paul, A. (2016). Hadoop based real-time intrusion detection for high-
speed networks. IEEE.
6. Basappa Kodada, S. P. (2015). Big-data analytics based security architecture to detect intrusions.
NJCIET.
7. Jangla, G. K., & Amne, D. A. (2015). Development of an intrusion detection system on big-data
for detecting unknown attacks. IJARCCE, 4.
8. Shaik Akbar, T. R. (2016). A hybrid Scheme based on big-data analytics using intrusion detection
system. IJST .

[email protected]
Pedestrian Detection and Tracking:
A Driver Assistance System

Shruti Maralappanavar, Nalini C. Iyer and Meena Maralappanavar

Abstract Detection and tracking of pedestrian is a challenging task due to variable

appearances, wide range of poses, and irregular motion of pedestrian along with
motion of tracking camera under complex outdoor environmental conditions. In this
paper, we propose an algorithm for pedestrian detection and tracking using HOG
descriptors and particle filtering technique. A robust algorithm for pedestrian detec-
tion is proposed which works under nonlinear motion and overcomes occlusions.
The performance of the above algorithm is tested for outdoor environment using
standard dataset. The particle filter has benefits of handling nonlinear motion and
occlusions, and they concentrate consecutively on the higher density regions of the
state space and it is simple to realize which provides a robust tracking environment.
Performance comparison of particle with conventional Kalman is also presented for
the above-said cases.

Keywords Histogram of oriented gradients (HOG) · Support vector machine

(SVM) · Particle filter · Kalman filter

1 Introduction

Human beings have an amazing potential of detecting and tracking objects in the
surrounding even if they are in motion or in stationary. This knowledge of human
visual system allows them to prevent colliding with other pedestrians in crowded
places. From a human beings point of view, it is a simple task to detect pedestrian
from the surrounding objects such as trees, vehicles, etc. but from the machine
vision outlook it is not an easy task as it involves dealing with many parameters

S. Maralappanavar · N. C. Iyer (B) · M. Maralappanavar

KLE Technological University, Hubballi, India
e-mail: [email protected]
S. Maralappanavar
e-mail: [email protected]
M. Maralappanavar
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2019 553
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_46

[email protected]
554 S. Maralappanavar et al.

such as illumination, cluttered backgrounds, pose of pedestrian clothing, size, and

shape. Thus, it is challenging for a machine vision system to implement pedestrian
detection and tracking functionality owing to their variable appearance: Variation
shape, size, color of clothes of the tracked pedestrian with accessories, umbrella,
child in trolley and outdoor environment: environmental conditions like rain, fog,
or complex and cluttered backgrounds with probable occlusion from surrounding
objects such as trees, vehicles, buildings, poles, other pedestrians, etc.

2 Related Work

The important attributes for both detection and tracking problem include robustness,
accuracy, and speed. Detection algorithm can be patch based or part based, and the
tracking accuracy depends on how good is the detection algorithm. Object detection
algorithm based on the formation of patches from image proposed by Prioletti et al.
[1] is cumbersome because of the formation of more number of patches from image
which determines the classification decision. Further, custom database needs to be
manually interpreted making implementation difficult task.
Later, Aghajanian et al. [2] proposed pedestrian detection algorithm to overcome
the creation of huge database, by modeling humans as flexible assemblies of parts.
These parts are represented by co-occurrences of local features which captures the
spatial layout of the part’s appearance. But the disadvantage with this method was
false detection and inaccuracy because of considering unwanted parts similar to
humans as human detection. Accuracy level is quite average.
To improve the performance of detection with respect to earlier methods [1, 2],
Dalal and Triggs [3] proposed localizing using Histograms of Oriented Gradients
(HOG) method.
On the other hand, for tracking, prediction of the most probable object in the
current frame based on inputs from the previous frame is given by Huang et al.
[4] using Kalman filter which is restricted only to neighboring area of the location
and fails for occlusion and nonlinear motion of the object. Later, pedestrian tracking
algorithm to overcome occlusion was proposed by Owczarek et al. [5] using a particle
filtering approach, where pedestrian is tracked based on the state information of
particle.
In this paper, we discuss pedestrian detection using location coordinates and its
changes in consecutive image frames for tracking by deriving HOG features for the
detection [3] and performing particle filtering for tracking [5].
The algorithm proposed is tested for its performance with occlusion and nonlinear
motion of pedestrian and is also validated with standard database.
The rest of the paper is as follows: Methodology in Sect. 3.1 which deals with
pedestrian detection using Histogram of Oriented Gradients (HOG) features and
training using Support Vector Machines (SVM) and particle filter details in Sect. 3.2.
Realization and experimental results are discussed in Sect. 4 with conclusion in
Sect. 5.

[email protected]
Pedestrian Detection and Tracking: A Driver Assistance System 555

3 Methodology

Details of pedestrian detection and tracking algorithm are given in this section. To
track an object in a video sequence, detecting of a pedestrian using HOG features [3]
is performed and location information is obtained which is then followed by tracking
using particle filtering [5]. Details of detection and tracking of a particle are briefed
in part A and part B, respectively.

3.1 Pedestrian Detection

Pedestrian detection using HOG features extraction involves the operation such as
cell formation, filtering, and block normalization as shown in Fig. 1 [3]. The details
of operations are given below.

Extraction of HOG features

Cell formation, filtering, and obtaining 1D histogram:

• Dividing original image into cells of each 8 × 8 pixel.

• Applying Sobel operator, obtain magnitude and direction of gradient the internals
of which clearly indicates the spreading of intensity gradients or edge directions.
• Creating 1D histogram of unsigned gradients obtained above, by distributing
among nine bins which indicates edge orientations over the pixels of the cell.
To overcome lighting and illumination effects, block normalization is done as
follows.

Fig. 1 Block diagram of pedestrian detection system

[email protected]
556 S. Maralappanavar et al.

Block normalization

• Divide original image into block of each 16 × 16 pixel.

• Block is normalized with respect to histogram of the cells, leading to features to
be extracted for an original image.

Testing Model

The extracted HOG features are used to obtain the model as follows:

• Original image as a test input to the trained model is given.

• Model gives the input about the presence or absence of pedestrian along with
location.
• The location information is obtained from the detection algorithm which is used
as an input to the tracking system as an initial state of input.

3.2 Particle Filter

Particle filtering is also named as the sequential Monte Carlo (SMC) [5], a simulation-
based technique. Particle filter involves the operations such as initialization, particle
prediction, assignment of weight to the particles, and resampling for an input frame
of a given video after identifying and obtaining the location of the pedestrian.
In the particle filtering approach, distribution of the system state is approximated
by a set of so-called particles and every particle is represented by a vector.
Algorithm can be summarized as follows [5]:
Initialization: Create a set of N-particles. Each particle has a state vector and an
initial weight which corresponds to the target pedestrian’s representation which is
given as
 T
dx dy
S  x, y, , (1)
dt dt
1
π0(n)  (2)
N
where n  1, 2, …, N,
where (x, y) are the coordinates of the bounding box, N is the number of particles,
and dx/dt and dy/dt stands for the velocity in x-axis and y-axis.
Particle Prediction: The transition equations help in predicting new particle state.
The state of each sample in every new frame is defined as function of driving vector,
noise introduced to the state due to the measurement error of wt−1

[email protected]
Pedestrian Detection and Tracking: A Driver Assistance System 557

st(n)  Ast−1
(n) (n)
+ wt−1 (3)

where A is transition matrix which is defined by

⎡ ⎤
1 0 dt 0 0
⎢ 0 1 0 dt 0 ⎥
⎢ ⎥
⎢ ⎥
A  ⎢0 0 1 0 0⎥
⎢ ⎥
⎣0 0 0 1 0⎦
00 0 0 1

Assignment of weight to the particles: Each measurement zt updates the weights

of the particles.

zt
πt(n)  πt−1
(n)
.p (4)
st(n)
⎛  2 ⎞
(n)
d
1 ⎜ t ⎟
πt(n) (n)
 πt−1 . √ exp⎝− 2 ⎠
(5)
σ 2π 2σ

where dt(n) represents the Hellinger distance between the hypothetical location and
tracked pedestrian.
Perform weight normalization and sum up the weights to one.
Resampling: Ignore the particles which have negligible weights and again assign
equal weights to the particles which have larger weights.
The distribution of particles after every operation is listed in Fig. 2. In the first
frame of the video, particles are spread throughout the frame. The initial position of
the pedestrian is obtained from the pedestrian detection system as shown in Fig. 2a.
Then, the particles are weighted based on the measurement update equation given
in Eqs. (3), (4), (5), and thus the particles near the pedestrian will obtain higher
weights based on conditional probability density. Discard the particles with negligible
weights as they are of no use in the tracking process as shown in Fig. 2b. Resample
the particles and assign equal weights to those particles which had higher weights in
the previous iteration as depicted in Fig. 2c.
Once the pedestrian starts moving, again the weights of the particles are updated
and particles with negligible weights are discarded. Further, resampling of particles is
done. This process is done continuously until the pedestrian is tracked continuously.

[email protected]
558 S. Maralappanavar et al.

Fig. 2 Steps in particle filter algorithm

[email protected]
Pedestrian Detection and Tracking: A Driver Assistance System 559

4 Realization and Experimental Results

The proposed algorithm of tracking of pedestrian using particle filtering is imple-

mented in MATLAB and tested using CALTECH database. Two test videos having
occlusion and nonlinear motion of pedestrian are taken as input. This is fed as input to
Kalman and particle filter, and the results of this are shown in Figs. 3 and 4. The track
speed is 50 frames/s by estimation. The results obtained for the two test conditions
are explained below.

Frame 150 frame 155 frame 170

(a) Kalman filter

Frame 150 frame 155 frame 170

(b) Particle filter

Fig. 3 Comparison of Kalman and particle filter

[email protected]
560 S. Maralappanavar et al.

Case 1: Occlusion

For the first test video, with Kalman filter, tracking failed in frame 170 due to occlu-
sion by the other pedestrian, whereas in the particle filter tracking the pedestrian who
was tracked in frame 150 is still being tracked in frame 170 successfully as shown
in Fig. 3a, b, respectively.

Case 2: Nonlinear motion

For the second test video, with Kalman filter, tracking of nonlinear motion of pedes-
trian in frame 185 failed as shown in Fig. 4a but particle filter overcomes nonlinear
motion successfully as shown in frame 185 in Fig. 4a.

Frame 170 frame 175 frame 185

(a) Kalman filter

Frame 170 frame 175 frame 185

(b) Particle filter

Fig. 4 Comparison of Kalman and particle filter

[email protected]
Pedestrian Detection and Tracking: A Driver Assistance System 561

5 Conclusion

Pedestrian detection and tracking using HOG descriptors particle filter are imple-
mented and tested for short video sequences from standard database. As compared
to Kalman filter, it is able to overcome occlusion as well as nonlinear motion on the
pedestrian. There is a tradeoff between accuracy and the processing time.

References

1. Prioletti, A. Part-based pedestrian detection and feature-based tracking for driver assistance:
Real-time algorithm filtering approach. Computer Science and Information Systems.
2. Aghajanian, J., Warrell, J., Prince, S. J. D., Li, P., Rohn, J. L., & Baum, B. Patch-based within-
object classification. In 12th International Conference on Computer Vision (ICCV).
3. Dalal, N., & Triggs, B. Histograms of oriented gradients for human detection. IEEE Computer
Society Conference on Computer Vision and Pattern Recognition.
4. Huang, S., & Hong, J. Moving object tracking system based on Camshift And Kalman Filter.
5. Owczarek, M., Barański, P., & Strumiłło, P. Pedestrian tracking in video sequences: A particle
filtering approach.
6. Fen, X., & Ming, G. (2010). Pedestrian tracking using particle filter algorithm. In International
Conference on Electrical and Control Engineering.
7. Zhang, L., & Li, L. (2013). Improved pedestrian detection based on extended histogram of
oriented gradients. In 2nd International Conference on Computer Science and Electronics
Engineering, ICCSEE.
8. Mittal, S., Prasad, T., Saurabh, S., Fan, X., & Shin, H. (2012). Pedestrian detection and tracking
using deformable part models and Kalman filtering. In ISOCC.
9. Isard, M., Blake, A. CONDENSATION—Conditional density propagation for visual tracking.
International Journal of Computer Vision, 29.
10. Nkosi, M. P., Hancke, G. P., & dos Santos, R. M. A. Autonomous pedestrian detection. IEEE.

[email protected]
Exploiting Parallelism Available in Loops
Using Abstract Syntax Tree

Anil Kumar and Hardeep Singh

Abstract Performance of a program depends on two factors: better hardware of

the executing machine and exploiting parallelism for concurrent execution. Loops
with multiple iterations provide efficient parallelism in an application, and are used
to reduce overall execution time and also to increase performance. Abstract Syntax
Tree (AST) can be used as an effective tool for exploiting parallelism at the compiler
level. This definitely saves time and automates the decomposition of a parallel job
that is to be executed in parallel framework. AST can be used to detect loops in the
source code, therefore this approach can be used to design a new parallel computing
framework where simple codes written for normal machines can be parallelized by
the framework itself.

Keywords Parallel computing · Abstract Syntax Tree · Job decomposition ·

Parallelism · Parallel framework

1 Introduction

Source codes of software are becoming huge and complex. Compiling large codes
is a time-consuming process. Parallel compilation of codes will help in reducing
the time complexity. Parsing is the phase of a compiler which requires a significant
amount of time for compilation. Many techniques have already been developed to
extract the parallelism available in a parser. AST is proposed to be an effective
way for exploiting parallelism at the compiler level via parsing and decomposing
the code in this study. Studies of parallel applications so far suggest that achieving
high performance with these applications is very significant. Not only do we find

A. Kumar (B)
Department of Computer Engineering and Technology, Guru Nanak Dev University,
Amritsar 143005, India
e-mail: [email protected]
H. Singh
Department of Computer Science, Guru Nanak Dev University, Amritsar 143005, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 563

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_47

[email protected]
564 A. Kumar and H. Singh

adequate parallelism in the program, but it is also important that we minimize the
synchronization and communication overheads in the parallelized program. In fact,
it is common to find parallel programs that are not able to run fast due to overheads
of parallel execution. It is therefore required to increase the granularity to reduce the
synchronization problems [1, 2]. We have studied two forms of parallelism based on
granularity since these exploit parallelism available in loops which, further, is one
of the techniques to improve the performance of the computer system.
To exploit loop parallelism at the instruction level, fine-grained parallel architec-
tures are used. At runtime, various forms of available dependencies between opera-
tions in a program must be checked either with the help of compiler (which would
be a static check) or with the help of hardware (to ensure that only independent
operations are issued simultaneously and this is a kind of dynamic check). To extract
maximum parallelism, the dependence-checking technique is used and not only it
examines the basic blocks of the program but also goes beyond the boundaries of
these blocks to find maximum independent operations from various iterations. On
the other hand, coarse-grained architectures exploit available parallelism in loops
by scheduling the entire set of iterations on separate processors. Parallelism can be
exploited from two types of loops, those with cycles in their dependence graphs,
and those with no cross-iteration dependencies. Both types can be executed on either
fine-grained or coarse-grained parallel architectures but as a consequence, different
performance judgments will be depicted by the two architectures. As the iterations of
a loop can be executed several times, so they provide pronounced parallelisms avail-
able in an application. To exploit this parallelism, the suitable technique depends
on the architecture of the executing parallel machine and the characteristics of each
single loop. An optimum strategy for loop and machine architecture is thus required
[3].

2 Forms of Parallelism

Figure 1 shows the various forms of parallelism which are further discussed in brief.

2.1 Loop Level Parallelism

Loop Level Parallelism is that form of parallelism, which is used to exploit paral-
lelism among iterations of the loop. A simple example to understand the basic loop
parallelism is [4]:
for (i  1; i < 500; i  i + 1)
a[i]  a[i] + a[i + 1];
Loop Level Parallelism can be exploited by unrolling loops either dynamically,
via branch prediction, or statically, via loop unrolling using compiler. The loops

[email protected]
Exploiting Parallelism Available in Loops … 565

Fig. 1 Forms of parallelism

with dependencies are executed using the do-across model based on coarse-grained
architecture and software pipelining as the scheduling strategy based on fine-grained
architectures. However, in these loops, explicit synchronization is not required [3].
The loops with no dependences are executed using do-all model based on coarse-
grained architecture and software pipelining as the strategy based on fine-grained
architecture. Due to less complexity and overheads, most of the frameworks use
do-all loops for exploiting loop parallelism.

2.2 Instruction Level Parallelism

Instruction Level Parallelism (ILP) is a technique which simultaneously executes the

sequential instructions from a program on multiple functional pipelining units [5].
For example,
1. x  y + z
2. s  a − b
3. m  x * s
Operations 1 and 2 are independent of each other, so they can be executed at
the same time. But operation 3 depends on the outcome of operations 1 and 2, thus
it can be calculated only after their completion [4]. ILP is helpful in exploiting
Loop Level Parallelism [6, 7]. It has been observed that usually the runtime of the
program gets reduced when we use the ILP approach because the execution time
of considerable instructions gets superimposed over each other using this technique
for parallelism. This approach is also very expensive. Machines with such designs
increase the complexity in control logic and thus limit performance. Hence, ILP is
not suitable for all conventional high-level language compilers [8].

[email protected]
566 A. Kumar and H. Singh

2.3 Thread Level Speculation

Thread Level Parallelism is a form of parallel computing which uses the technique
of distributing the execution of threads across different parallel processor nodes [9].
Focusing only on loops does not yield complete parallel potential of an application.
Hence, to uncover the hidden Loop Level Parallelism and to strengthen parallelism
in general purpose programs, Thread Level Speculation is required [10–12]. Thread
level speculation hardware support, along with the use of several code transforma-
tions, has made it possible to expose the hidden Loop Level Parallelism in an appli-
cation. In such architectures, the threads are extracted from sequential programs and
are run in parallel. Each iteration of the loop can turn out to be a speculative thread.
The degree of parallelism available in a loop is determined by dependences; more
the independence between iterations more is the degree of parallelism. In addition
to this, if we consider parallelism coverage, then choosing outer do-across over an
inner do-all loop becomes more beneficial for parallelism coverage. Both degrees
of parallelism, as well as parallelism coverage, are the main factors for loop specu-
lation. The Procedure Level Speculation can also be used to improve the available
parallelism for overcoming the fact that procedures are less popular than loops as
a target of parallelization. It is also studied that simultaneous multithreading can
convert the thread level parallelism to Instruction Level Parallelism because mostly
superscalar processors exploit Instruction Level Parallelism by executing pipelining.

2.4 Superword Level Parallelism

Superword Level Parallelism (SLP) is a basic block vectorization technique based on

loop unrolling. It can be a better replacement for all of the above as has been studied
that it can exploit loop parallelism both across loop iterations, as well as within the
basic blocks [13]. It is closely related to ILP. However, studies have shown that there
is a need to recover the basic blocks that have not been vectorized sufficiently by this
technique.

3 Decomposition Methods

Decomposition techniques serve as the best methods to make efficient use of paral-
lel applications. These techniques divide a computation into a local part. This part
does not require any inter-process intervention during its generation, and it mainly
involves communication between neighboring and distant processors [14, 15]. When
a problem is decomposed into a large number of tasks, it is called Fine-grained decom-
position; otherwise, it is called coarse-grained decomposition. However, the nature
of the problem decides which one is better, as the large number of tasks ensures better

[email protected]
Exploiting Parallelism Available in Loops … 567

Fig. 2 Decomposition
techniques

scalability with more concurrency and hence more utilization of resources. Lesser
number of tasks usually includes less communication overheads. Figure 2 enlists the
decomposition techniques which are as follows:
• Recursive Decomposition
• Data Decomposition
• Functional Decomposition
• Exploratory Decomposition
• Speculative Decomposition
• Primal Decomposition
• Dual Decomposition
• Gradient-Based Decomposition.

3.1 Recursive Decomposition

It is generally suited to problems that are solved using the Divide and Conquer strat-
egy. A given problem is first decomposed into a set of subproblems. These subprob-
lems are recursively decomposed further, until a desired granularity is reached [15].
The results of the subproblems are merged together when needed. Usually, smaller
tasks are independent of one another, so they can be executed in parallel. So, good
scalability can be achieved by sorting algorithms in parallel often use this approach
for decomposition. Quad-tree, Oct-trees, and K-trees are the few approaches based
on recursive decomposition for parallelization [16]. Figure 3 shows the task depen-
dency graph of Quicksort algorithm based on recursive decomposition. The problem
is solved using Divide and Conquer technique.

[email protected]
568 A. Kumar and H. Singh

Fig. 3 Quicksort task

dependency graph based on
recursive decomposition [15]

3.2 Data Decomposition

Data Decomposition technique is used when data structures carrying bulk of similar
data need to be processed. The data groups form tasks. These can be input data, output
data, or even intermediate data. The decompositions of the computation into tasks
are done by using the Owner Computes Rule [8]. All data forms can be decomposed.
In Fig. 4, the input–output matrices are first partitioned into sub-matrices then
these are further decomposed into tasks. The tasks are independent of each other and
thus, it is easily solved and scales well [15, 17].

3.3 Functional Decomposition

The technique of decomposing functions into tasks is known as Functional Decom-

position. These tasks are executed concurrently on different data by processes which
use the pipeline approach. Instead of its performance simplicity, it does not scale
well because enough functions are not available for splitting. As an example, Fig. 5
shows the flow of data in counting word problem using pipelining.

[email protected]
Exploiting Parallelism Available in Loops … 569

Fig. 4 a The input–output sub-matrices are formed. b Matrices are decomposed into tasks

Fig. 5 Pipeline data flow in Word Count [5]

3.4 Exploratory Decomposition

Exploratory Decomposition is a technique which first searches the predefined space,

and then partitions it into tasks to process them concurrently. It is a special case
decomposition which is generally not applicable [15, 18]. An example is breadth-
first search in trees as shown in Fig. 6. Traverse order are: F, B, G, A, D, I, C, E, H.

[email protected]
570 A. Kumar and H. Singh

Fig. 6 Breadth first traversal

of tree

3.5 Speculative Decomposition

Sometimes, there are many functions available in a program and only one of those
needs to be carried depending upon execution of some condition, such at switch state-
ment. All such functions are turned into tasks and carried out before the evaluation
of the condition. This is called Speculative Decomposition [15, 19]. As soon as the
condition is evaluated, the result of only one task is used, and the rest are discarded.
But this decomposition does not make good use of resources, and thus is not popular.
The different decomposition methods described so far can be combined together
to form hybrid decompositions which usually overcome the limitations of the com-
prising decompositions and often propose good solutions.

3.6 Primal Decomposition

Primal Decomposition is the optimization technique which solves the structured opti-
mization problems in mathematics. It uses the master–slave approach for decompo-
sition. The master problem manipulates the primal variables (variables comprising
linear objective function) and subproblems are solved in parallel [20].

[email protected]
Exploiting Parallelism Available in Loops … 571

Table 1 Summarized table of various decomposition methods

Decomposition Granularity Scalability Communication Time/space
methods overheads complexity
Recursive Coarse grain High Low Low
Data Fine grain High High High
Functional Fine grain Low High Low
Exploratory Coarse grain Low High High
Speculative Coarse grain Low High Low
Primal Coarse grain High High High
Dual Fine grain High High High
Gradient based Fine grain High Medium Medium

3.7 Dual Decomposition

Dual Decomposition is an optimization technique which uses sub-gradient algorithm

for the master [20]. It is obtained from Lagrangian Formation [32]. It solves the dual
subproblems in parallel and updates the dual variables (variables comprising linear
objective function) iteratively. The dual problem solution defines lower bound for
the primal solution.

3.8 Gradient-Based Decomposition

The Decomposition Optimization Algorithm, also known as Path-following

Gradient-Based Decomposition Method, decomposes the problem into smaller sub-
problems by using Dual Decomposition [21]. It is a combination of three tech-
niques, namely Dual Decomposition [20], Smoothing [22], and Lagrangian Relax-
ation Method [23].
Table 1 shows the summary of decomposition techniques and their evaluation
based on parameters such as Granularity, Scalability, Communication Overheads,
and Time/Space Complexity.

4 Code Parsing Using Ast and Decomposition

Abstract Syntax Tree is used to represent the syntactic objects used by the systems that
manipulate the programs, formulas, rules, etc. It is the outcome of the syntax analysis
phase of the compiler of the program [24, 25]. Matching, substitution, and unification

[email protected]
572 A. Kumar and H. Singh

are its main operations. The AST matching simply notifies the user about the changes
which the global variables and functions in the program undergo. Basically, it is con-
sidered that these changes prove to be very useful from the point of software evolution
because it is considered to be helpful for updating software. High order Abstract Syn-
tax Trees is the variations of the basic Abstract Syntax Trees. For formal software
development, there is need of a language generic environment; for that purpose high
order AST has been designed. From parallel computing point of view, every paral-
lelizer consists of various compiler passes. Each compiler uses AST as a data structure
to represent the structure of the source code. It is studied [3] that the parallelism in do-
all loops is useful in reducing the execution time of the program. AST can be used to
detect the loops in the source code. The concrete structure of the program is a part of
the definition of the language which is basically defined by the context-free language,
whereas the Abstract Syntax Tree is a part of the implementation of the program and
is defined by the tree structure [26, 27]. Thus, it can be said that the essential syn-
tactic information cannot be given by the concrete structures and parse trees, which
sometimes contains useless information. Thus, the Abstract Syntax Trees came up
as a solution. Further, it has been studied that high-order syntax trees were not able
to fulfill requirements such as accounts of structural index and induction, and recur-
sive equation for abstract syntax. Abstract syntax with variable binding presents the
algebraic (categorical) view [26] of syntax, which to some extent serves as one of
the satisfactory solutions to fulfill the abovementioned requirements. Abstract syn-
tax is the initial such model, with the algebra structure obtained as the solution to a
recursive equation and substitution defined by an associated structural recursion.

4.1 Methodology

AST is parsing technique which can make the relevant information available at
compiler level by parsing the source code of the program to exploit loop parallelism.
Algorithm 1 is the explanation of the flowchart shown in Fig. 1. This procedure
explains how the parser generates the AST.
Algorithm 1: Create AST

1. LOAD Parser
2. INPUT source code
3. READ number of bytes of source code
4. COMPILE source code
5. RUN Parser
6. GENERATE parsed code.

[email protected]
Exploiting Parallelism Available in Loops … 573

7. HANDLE errors.
8. CREATE variables, data structures for parsed code
9. CREATE database
10. ALLOCATE unique counter variable to each database table
11. GENERATE AST
12. INCREMENT counter variable on visiting node of AST
The detailed description of flowchart stages is shown below:
• Loading parser: Parser is imported from Java package “Package parser” and the
loader links all the required Java libraries for the parser.
• Input source code: Source code is given as an input to the parser. Since Java
is a strictly typed language so the source code should strictly follow the naming
and typing rules. If possible, different names should be used for each element in
coding.
• Reading number of bytes of source code: Parser calls the Java method that returns
the number of bytes of the source code to be read by the parser.
• Compiling code: Source code is compiled using compilation unit CU and the
variables are created.
• Run parser: Parser runs as a different thread by implementing Java’s runnable
interface.
• Generating parsed code: Source code is parsed.

Fig. 7 a Source code (a)

parsing. b Generate Abstract Package
Syntax Tree Parser

Import all required

java libraries
Create required
variable

Parser class inherits

runnable interface
Parser runs on
thread

Input source code

File is parsed

Calls java method that

returns the number of
byte to be read
Error Handling

Compilation unit CU

[email protected]
574 A. Kumar and H. Singh

(b)
Parse code

Input all necessary files Create database file

Visitor inherits visitor adapter

Create required database

Creates array list and variables tables with their associated
counter variables as check

Builds connection with SQLite database

Create Syntax table and insert values Return

counter== = ()

Create for and while loop database table False

Create
True return
Create variable database table Repeat 3 database
times table

Insert values into

return table

Dir exists () dir mkdir ()

False Close database

connection

True

File dir =new file () Abstract syntax tree is

generated

Fig. 7 (continued)

[email protected]
Exploiting Parallelism Available in Loops … 575

• Handling errors: Errors are handled with various Java handling techniques.
• Creating data items: Data structures like array lists, and variables are created for
the parsed code.
• Create database: Database is created when the parser builds connection with
SQLite database module.
• Allocating counter variable: A unique counter variable is allocated to each
database table of loops, variables, etc.
• Generating AST: Abstract Syntax Tree is generated.
• Incrementing counter variables: Each counter variable is incremented in
database for each useful visit to the node of AST. The different module in Java
is created and named as visitor which inherits Visitor Adapter and imports all
required Java libraries and packages.
The flowcharts (Fig. 7) explain how the Abstract Syntax Tree is generated. This
AST will help to detect the loops available in the source code and enables the user
to choose which among them are to be targeted to achieve parallelism. Moreover,
the parsed information can be converted back to source code using the unparsing
technique and this helps in modifying the code for parallelism.

5 Conclusion

With the support of Abstract Syntax Tree, the source code of a parallel job can be
parsed and relevant information about different structures used in the code can be
revealed at the compiler level. This information is useful for any parallel framework
to parallelize a sequential algorithm with for-loops. It is clear from the literature
survey that the loop parallelism for do-all loops seems to be the most practical and
common within the programs. So, this type of parallelism is considered as the target
to be exploited for parallelization. AST facilitates to decompose the problem at the
compiler level, which can reduce the complexity at design and coding of parallel
jobs using special parallel programing tools. This further reduces the need to know
more parallel language constructs to parallelize the given job. This approach of
using AST output at the compiler level can be used to design new parallel computing
frameworks where simple codes, written for normal machine, can be parallelized by
the framework itself.

References

1. Parhami, B. (2002). Introduction to parallelism. Introduction to parallel processing: Algorithms

and architectures (pp. 3–23).
2. https://en.wikipedia.org/wiki/Instruction-level_parallelism.
3. Lilja, D. J. (1994). Exploiting the parallelism available in loops. Computer, 27(2), 13–26.
4. https://simple.wikipedia.org/wiki/Task_parallelism.

[email protected]
576 A. Kumar and H. Singh

5. Kumar, R., & Singh, P. K. (2014). An approach for compiler optimization to exploit instruction
level parallelism. In Advanced Computing, Networking and Informatics (Vol. 2, pp. 509–516).
Cham: Springer.
6. Rau, B. R., & Fisher, J. A. (2003). Instruction-level parallelism.
7. Lo, J. L., Emer, J. S., Levy, H. M., Stamm, R. L., Tullsen, D. M., & Eggers, S. J. (1997). Convert-
ing thread-level parallelism to instruction-level parallelism via simultaneous multithreading.
ACM Transactions on Computer Systems (TOCS), 15(3), 322–354.
8. http://www2.phys.canterbury.ac.nz/dept/docs/manuals/FORTRAN.
9. https://cloud.google.com/dataflow/examples/wordcount-example.
10. Zhong, H., Mehrara, M., Lieberman, S., & Mahlke, S. (2008, February). Uncovering hidden
loop level parallelism in sequential applications. In IEEE 14th International Symposium on
High Performance Computer Architecture, 2008. HPCA 2008 (pp. 290–301). IEEE.
11. Tullsen, D. M., Eggers, S. J., & Levy, H. M. (1995, June). Simultaneous multithreading: Max-
imizing on-chip parallelism. In 22nd Annual International Symposium on Computer Architec-
ture, 1995. Proceedings (pp. 392–403). IEEE.
12. Wall, D. W. (1991). Limits of instruction-level parallelism (Vol. 19, No. 2, pp. 176–188). ACM.
13. Larsen, S., & Amarasinghe, S. (2000). Exploiting superword level parallelism with multimedia
instruction sets (Vol. 35, No. 5, pp. 145–156). ACM.
14. Quarteroni, A., & Valli, A. (1996). Domain decomposition methods for partial differential
equations.
15. Johnson, T. A., Eigenmann, R., & Vijaykumar, T. N. (2007, March). Speculative thread decom-
position through empirical optimization. In Proceedings of the 12th ACM SIGPLAN Symposium
on Principles and Practice of Parallel Programming (pp. 205–214). ACM.
16. Jackins, C. L., & Tanimoto, S. L. (1983). Quad-trees, Oct-trees, and K-trees: A generalized
approach to recursive decomposition of euclidean space. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 5, 533–539.
17. Stone, R. B., & Wood, K. L. (2000). Development of a functional basis for design. Journal of
Mechanical Design, 122(4), 359–370.
18. Arabnia, H. R. (1990). A parallel algorithm for the arbitrary rotation of digitized images using
process-and-data-decomposition approach. Journal of Parallel and Distributed Computing,
10(2), 188–192.
19. Jojic, V., Gould, S., & Koller, D. (2010, June). Accelerated dual decomposition for MAP
inference. In ICML (pp. 503–510).
20. Devaurs, D., Siméon, T., & Cortés, J. (2013). Parallelizing RRT on large-scale distributed-
memory architectures. IEEE Transactions on Robotics, 29(2), 571–579.
21. Neamtiu, I., Foster, J. S., & Hicks, M. (2005). Understanding source code evolution using
abstract syntax tree matching. ACM SIGSOFT Software Engineering Notes, 30(4), 1–5.
22. Dinh, Q. T., Necoara, I., & Diehl, M. (2014). Path-following gradient-based decomposition
algorithms for separable convex optimization. Journal of Global Optimization, 59(1), 59–80.
23. https://web.stanford.edu/class/ee364b/lectures/decomposition_slides.pdf linear algebra with
applications 7.7-8 (2000), 687–714.
24. Pfenning, F., & Elliott, C. (1988, June). Higher-order abstract syntax. In ACM SIGPLAN Notices
(Vol. 23, No. 7, pp. 199–208). ACM.
25. Fiore, M., Plotkin, G., & Turi, D. (1999). Abstract syntax and variable binding. In 14th Sym-
posium on Logic in Computer Science, 1999. Proceedings (pp. 193–202). IEEE.
26. Wile, D. S. (1997, May). Abstract syntax from concrete syntax. In Proceedings of the 19th
International Conference on Software Engineering (pp. 472–480). ACM.
27. Lim, A. W., & Lam, M. S. (1997, January). Maximizing parallelism and minimizing synchro-
nization with affine transforms. In Proceedings of the 24th ACM SIGPLAN-SIGACT Symposium
on Principles of Programming Languages (pp. 201–214). ACM.

[email protected]
A Study on Cooperation and Navigation
Planning for Multi-robot Using
Intelligent Water Drops Algorithm

D. Chandrasekhar Rao and Manas Ranjan Kabat

Abstract The shortcomings of the existing multi-robot navigation planning meth-

ods in an unknown environment are studied extensively. An efficient approach for
resolving these issues is addressed in the current study using a novel nature inspired-
Intelligent Water Drops (IWD) algorithm [1]. The robustness of the proposed method
for multi-robot navigation in an unknown environment is validated through V-Rep
simulator. The performance of the algorithm is verified through simulation outcomes
in terms of total path travelled, total path deviation, number of turns, and execution
time. Further, the efficiency of the proposed method is compared with the existing
state of the art to verify its potency. The simulation outcomes reveals that the pro-
posed method takes in an improvement of 9.93%, 35.26%, 33.33%, and 13.04% in
terms of the total path travelled, total path deviation, number of turns, and total exe-
cution time for all the robots to arrive their target, respectively, as compared to the
existing state of the art. Moreover, the current study confirms the superiority of the
proposed approach as compared to the existing state of the art in terms of generating
optimal and safe navigation path for individual robots.

Keywords Multi-robot · Navigation planning · Waypoints · Intelligent water

drops · Static environment

1 Introduction

In recent years, the application of robotics has been evolved as the most cost-effective
approach in many fields [2–8]. However, navigation planning is one of the key con-
cerns in most of the applications employed with mobile robots. In navigation planning

D. Chandrasekhar Rao (B)

Department of Information Technology, VSSUT, Burla, Sambalpur 768017, Odisha, India
e-mail: [email protected]
M. R. Kabat
Department of Computer Science & Engineering, VSSUT, Burla, Sambalpur 768017, Odisha,
India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 577

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_48

[email protected]
578 D. Chandrasekhar Rao and M. R. Kabat

problem, the main objective of the mobile robot is to find an optimal and safe path
while navigating from its source to destination. Thus, the efficiency of the mobile
robot needs to be enhanced such that it can reach the target through a shortest path with
minimal time and energy utilization. Various research findings have been addressed
in this domain for improving the autonomy and decision accuracy of the mobile robot
to resolve these issues.
The working environment of mobile robots may be regarded as static or dynamic.
In a static environment, the obstacle and target positions are stationary, however,
these may be dynamic in the dynamic environment. Further, the navigation planning
of a mobile robot in an environment can be classified as global or local. In global
planning, the mobile robot has prior knowledge about the environment and decides
the collision-free path in advance. In local planning, the mobile robot generates
collision-free path on the go with no prior knowledge about the environment. Local
planning is sometimes referred to as online planning. Global and local planning is
most suitable for static environment and dynamic environment. However, both of
these planning may be combined to improve the efficiency of the mobile robot in a
complex environment [9].
In this study, we proposed a novel approach for multi-robot navigation planning in
a static environment using nature-based IWD algorithm. The proposed algorithm is
efficient to generate a safe and an optimal path for the individual robot with minimum
time and energy utilization. The effectiveness of the proposed algorithm is further
verified with the existing state of the art [10].
The rest of this paper is structured as follows. In Sect. 2, the related work for multi-
robot navigation has been elaborated. The problem formulation, operating principle
of the basic IWD algorithm and its implementation in multi-robot navigation are
discussed in Sect. 3. Section 4 briefs the simulation setup and performance evaluation
of the proposed method. Finally, the work is concluded in Sect. 5.

2 Related Work

Navigation planning is the primary concern in most of the mobile robot applications.
The approaches employed in resolving the robot navigation problem can be broadly
categorized as traditional or heuristics [11]. The state of the art in this field is discussed
below.
The traditional approaches comprise of Cell decomposition [12], the probabilistic
Roadmap approach [13], Voronoi Diagram [14], and Artificial Potential Field [15]
are employed to solve the Mobile Robot Navigation Problem (MRNP). However,
these approaches are trapped in local minima, incompetence in the presence of large
obstacle, and computational cost is more. Recently, with the development of heuris-
tic approaches, it is embedded in many applications to resolve the issues associated
with MRNP. These approaches have a good potential for global exploration, how-
ever, each approach has its own limitation. This boosts the researcher to employ
heuristic approaches for addressing mobile robot navigation planning. Many heuris-

[email protected]
A Study on Cooperation and Navigation Planning … 579

tic approaches such as Fuzzy Logic (FL), Genetic Algorithm (GA), Particle Swarm
Optimization (PSO), Gravitational Search Algorithm (GSA), Differential Evolution
(DE), and hybridization of these approaches have employed to solve MRNP.
FL is extensively employed to resolve the navigation planning problem of mobile
robots, which utilizes the predicted behaviour of humans to deal with uncertainties.
The FL incorporates the characteristics of human perception for resolving real-world
problems. The authors in [16] proposed a Fuzzy Logic Controller (FLC) for MRNP
to deal with an environment consisting of static obstacles. The rule set for FLC is
constructed using the distance of obstacles from right, left, and front side of mobile
robots. Simulation has been carried out to prove the effectiveness of the proposed
FLC in an environment consisting of thousands of robots and obstacles. In [10],
the authors proposed Obstacle avoidance FLC (OA-FLC) and Obstacle avoidance
FLC (OA-FLC) for escaping from obstacles and maintaining the heading direction
of robots toward goal in the static as well as the dynamic environment respectively.
However, the proposed method has not guaranteed optimal path always. Though FL
is efficient for simulating the decision making capability of human being, but it is
incompetent in the selection of appropriate membership functions for finding optimal
path. GA is another widely used method based on the natural selection procedure in
the field of robot navigation. Ozkan et al. [17] proposed multi-robot navigation using
a hierarchical-oriented genetic algorithm (HOGA) to optimize the time required by
robots rather than the path travelled by robots. In [18], navigation planning based on
GA is proposed in a dynamic environment. However, GA is inefficient in controlling
population difference on a grid-like environment and may occur in an early conver-
gence. Besides GA, PSO is another population-based method, employed in MRNP.
PSO [19] and its improvement [20] are proposed for generating the optimal trajectory
of robots in an unknown environment. Further, GSA based on the gravitational law
of masses is employed to address the MRNP for generating the optimal path. GSA
[21] and its improvement [22] are proposed to find the optimal trajectory by avoiding
the static and dynamic obstacles in an unknown environment. However, the conver-
gence rate of PSO and GSA are slow in an iterative process and may be trapped at
local minima. Rao et al. [23] proposed DE algorithm for multi-robot navigation in a
clutter environment. Similar to GA, DE uses crossover, trial, and selection operators
to generate a safe path for mobile robots in an unknown environment. However, it
does not guarantee an optimal path. Furthermore, hybridization of many heuristic
approaches [24–26] has been proposed for addressing MRNP.
The above-discussed works generate collision-free path for robots in different
environments using classical and soft computing approaches. Though, the proposed
methods able to generate safe path for mobile robots but does not guarantee optimal
path always in terms of shortest path, time and energy utilization for navigating the
robot from source to destination. Furthermore, in a large problem space, most of the
algorithms trapped in local minima, oscillate in the presence of large obstacles, and
also fail to maintain a good balance between intensification and diversification. These
factors motivated us to propose a novel approach to overcome the above-discussed
shortcomings. Thus, a novel approach by utilizing the concept of IWD algorithm
is proposed for multi-robot navigation in a static environment. The mobile robots

[email protected]
580 D. Chandrasekhar Rao and M. R. Kabat

simulate the behaviour of IWD to find optimal and safe path in the world map. The
proposed IWD algorithm is simple to implement, capable for global search and needs
tuning of few control parameters in a static environment. The main highlights of this
paper are as follows:
i. The target function is formulated by considering the weighted sum of each
individual objective function.
ii. The proposed method is able to escape from static obstacles as well as teammates
(other robots considered as dynamic obstacles) present in the environment.
iii. The proposed method performance is verified through V-Rep simulator.
iv. Further, the potency of the proposed method has been validated through the
existing state of the art.

3 Proposed Multi-robot Navigation Planning

The configuration space for multi-robot navigation problem is consisting of multiple

autonomous mobile robots and static obstacles. The primary issue is to identify an
optimal, safe. and smooth path for individual mobile robots in the configuration space
from their respective start to target position.

3.1 Problem Formulation

The problem space for the multi-robot navigation is considered as a two-dimensional

world map. The environment consists of static obstacles of different shapes and two-
wheeled multiple mobile robots. Every robot in the environment has predefined start
and the goal position. The objective of the problem is to find an optimal and safe path
for the individual robot from their respective start position to the goal position. In
the initial setup, every robot’s orientation is toward the goal position. The structure
of a two-wheeled mobile robot is shown in Fig. 1.
The current position of a two-wheeled mobile robot in two-dimensional coordinate
system is denoted by (x, y) and orientation of the robot with respect to abscissa is θ .
All the robots in the environment are of identical structure with radius r . The angular
and translational motion of the robot is denoted by ω and v, respectively. In order
to navigate safely, every mobile robot generates successive waypoints by avoiding
obstacles in their way. Let { p1 , p2 , . . . , pm } be the “m” waypoints generated by the
mobile robot while navigating from source to goal. Thus, the total path travelled by
an individual robot can be expressed as follows:


m  
j j+1
Pi  dist pi , pi (1)
j0

[email protected]
A Study on Cooperation and Navigation Planning … 581

Fig. 1 Structure of a
two-wheeled mobile robot

 
j j+1
where dist pi , pi represents the Euclidian distance between two successive way-
points j and j + 1, pi0 and pim+1 is the start and the goal position of the ith robot,
respectively. Finally, the objective function for minimizing the path cost for all the
robots in the environment can be expressed as follows:
 n 

Fpath  min Pi (2)
i1

where Pi denotes the path travelled by ith robot. Further, to obtain collision-free path,
the robot needs to maintain a minimum distance from obstacles as well as with its
teammates during navigation in the environment. In order to escape from obstacles, a
linear function can be formulated for evaluating the safety factor. It can be expressed
as follows:
OBSiavoid
    
max 0 dist Pos(Ri ), Pos Obs j > λobs

dist(Pos(Ri ),Pos(Obs j )) −
1 1
1 ≤ j ≤ l, i  j λobs otherwise
(3)
where  i ) is the position of ith robot, l is the number of obstacles surrounding Ri ,
 Pos(R
Pos Obs j is the position of jth obstacles for j  {1, 2, . . . , l}, λobs is the constant
safety factor between the obstacles and the corresponding mobile robot. However,
the value of λobs can be chosen based on the influence radius of the obstacles. Here,
the teammates (other robots) are considered as the dynamic obstacle for the current

[email protected]
582 D. Chandrasekhar Rao and M. R. Kabat

robot. In order to maintain a safe path, the robot should maintain the minimum
safety distance from the nearest obstacles. Thus, the objective function for safe path
for multi-robot can be expressed as follows:

min
Fobs_avoid  OBSiavoid (4)
1≤i ≤n

Finally, the above-discussed objective functions for multi-robot navigation can be

formulated into a single objective function. Thus, the target function can be viewed
as a weighted sum of individual objective functions and can be expressed as follows:

Ffitness  α1 · Fpath + α2 · Fobs_avoid (5)

where α1 and α2 are the nonzero positive weight factor for total path travelled and
safety factor for obstacle avoidance, respectively. In the present work the value of α1
and α2 are set to 0.6 and 0.4, respectively.

3.2 IWD Algorithm

IWD [1] is a novel population-based nature inspired optimization algorithm that

simulates the behaviour of water drops in the river bed. IWD is able to provide a
promising solution to various optimization problems [17]. The water drops of a river
play a major role in forming the river path from source to destination. The water
drops face many obstacles in its way and turn intelligently to avoid the obstacles.
The observation made from the natural water drops that it forms an optimal river bed
from source to destination in presence of obstacles. The two important properties of
water drops such as the velocity and the soil accumulation govern the IWD to find
an optimal path. The selection of subsequent sub path by IWD toward its destination
depends on the soil concentration of the path. The relationship between these two
properties is the driving force for the IWD to form an optimal path between the
source and destination. The soil concentration on the path is inversely proportional
to the velocity of IWD. Also, the amount of soil carried by IWD is an inverse relation
with the time taken by IWD to traverse the sub path. The above-discussed property of
IWD reveals the similarity between the paths finding behaviour of IWD with mobile
robot navigation in the presence of obstacles. These factors motivated us to simulate
the behaviour of IWD into a mobile robot to find an optimal path.

[email protected]
A Study on Cooperation and Navigation Planning … 583

3.2.1 Implementation of IWD for Navigation Planning of Mobile

Robots

The navigation path of a mobile robot can be represented through a graph G 

{V, E}. Where, V and E represents the number of waypoints and the sub path
between the two consecutive waypoints, respectively. Here, the IWD will be treated
as the mobile robot. In the river bed, each IWD decides its next waypoint on the
basis of the amount of soil in the sub path. Similar to IWD, the mobile robot chose
its subsequent waypoint based on the obstacle distance from it. The distance of
an obstacle is inversely proportional to the soil amount present in the path. As the
concentration of soil is inversely proportional to the velocity of IWD, the mobile
robot will reduce its velocity if it is nearer to the obstacles. Thus, the proposed IWD
algorithm for multi-robot navigation can be formulated as follows.
(a) Static parameters Initialization: Initialize the number of IWD (NIWD ), maximum
number of iteration (itermax ) velocity updating parameters and soil updating
parameters.
(b) Dynamic parameters initialization: Initialize
 the set of visited waypoints by
  path
IWD Vwp  φ , initial soil in the path initsoil  100 , initial velocity of IWD
 IWD 
(initvel  10) and
 initial soil in IWD initsoil  0 . The fitness of global best
solution Ffitness IWDgbest is set to −∞.
(c) The initial position of mobile robots is set as the initial position of IWD. Update
Vwp with the initial position of IWD.
(d) For each IWD, repeat (i) to (iv)

i. Compute the next waypoint for individual IWD using the following expres-
sion:
⎧  
⎪
⎨ f soil(i, j)
   if j ∈
/ Pos(Obs)
j 
pi,IWD / (Obs) f soil(i,k)
1≤k≤NIWD , j k,k ∈Pos (6)
⎪
⎩0 otherwise

where

1
f (soil(i, j))  (7)
εsoil + g(soil(i, j))

and
⎧
⎪
⎪ min
{soil(i, l)} ≤ 0
⎨ soil(i, j) if
l∈
/ Vwp
g(soil(i, j)) 
⎪
⎪ min
⎩ soil(i, j) − {soil(i, l)} otherwise
l∈
/ Vwp
(8)

[email protected]
584 D. Chandrasekhar Rao and M. R. Kabat

where εsoil  0.01 is a small positive constant to avoid the constraint divi-
sion by zero, soil(i, j) is the concentration of soil between the two adjacent
waypoints, the functions f () and g() is used to maintain a positive value to
soil(i, j). The new set of jth waypoints will be updated into Vwp .

ii. For each IWD, the velocity is updated to move from ith waypoint to jth
waypoint using following expression:
αv
VelIWD (t + 1)  VelIWD (t) + (9)
μv + λv · soil2 (i, j)

where αv  1, μv  0.01 and λv  1 are the constant updating parameters

for IWD velocity.

iii. After updating the velocity of IWD, the amount of soil taken by the respec-
tive IWD during its journey on the sub path from ith waypoint to jth way-
point can be expressed as follows:
αs
δsoil(i, j)  (10)
μs + λs · Time(i, j; VelIWD )

and

dist(Pos(i), Pos( j))

Time(i, j; VelIWD )  (11)
max{εvel , VelIWD }

where αs  1, μs  0.01 and λs  1 are the constant soil updating

parameters for soil accumulation by IWD, Time(i, j; VelIWD ) is the time
taken by the IWD with a velocity of VelIWD to move from i to jth waypoint
and εvel  0.001 is the small positive constant to prevent the negative or
zero value of VelIWD .
iv. The soil amount in the sub path and accumulated by IWD then needs to be
updated using the following expression:

soil(i, j)  (1 − γ ) · soil(i, j) − γ · δsoil(i, j) (12)

IWDsoil  IWDsoil + δsoil(i, j) (13)

where γ is the small positive constant selected from (0, 1) and IWDsoil is
the concentration of soil in the IWD.

(e) After finding the solution for all IWD, evaluate the iteration best solution using
the following expression

IWDibest  min{Ffitness (TIWD )} (14)

[email protected]
A Study on Cooperation and Navigation Planning … 585

where IWDibest is the best solution in the current iteration based on the evaluation
of each IWD (TIWD ) using the objective function stated in Eq. (5). Further,
the sub path corresponding to the iteration best solution is updated using the
following expression
 
soil(i, j)  ηs · soil(i, j) + ηIWD · κ Vwp · IWDibest
soil ∀(i, j) ∈ IWDibest
(15)
 
 is the soil accumulated by the iteration best IWD, κ Vwp 
ibest
where
 IWD soil
1/ Vwp  − 1 , ηs and ηIWD are the constants. After updating the soil concentra-
tion in the respective sub path, the global best solution for the current iteration
can be updated using the following expression
  
IWDgbest if Ffitness IWDgbest ≥ Ffitness (IWDibest )
IWDgbest  (16)
IWDibest otherwise

(f) Repeat Step (a) to Step (e), until (iter  itermax ) and each IWD not reached the
destination.

4 Simulation Result and Performance Evaluation

The multi-robot environment with obstacles is configured through V-Rep 3.4.0 [27].
The environment consists of four two-wheeled Pioneer p3dx mobile robots and eight
different shapes of obstacle. The initial and target position of all the robot is predefined
in the environment. The detailed specification of Pioneer p3dx can be found in [28].
The simulation program is carried out in a Laptop PC with Core i3 processor and 4 GB
of RAM. The IWD algorithm is implemented using MATLAB 2016a and interfaced
with V-Rep for controlling the robots. The size of the world map for simulation is 5
× 5 m2 . The environment is consisting of four Pioneer p3dx mobile robot and seven
obstacles. The Isometric view and top view of the environment is shown in Fig. 2a, b,
respectively. The current position of mobile robots in Fig. 2b is the start positions of
the robots and the sphere with different colour code represents the respective target
position of the robots. Initially, all the robots heading direction is toward the target
position.
The maximum speed of the robot is set to 20 cm/s. The simulation for multi-robot
navigation was carried out by employing the IWD algorithm for 30 runs and the best
performance of the algorithm is presented in Fig. 3. The intermediate configuration
of robots during simulation is shown in Fig. 3a. The final configuration of robots
where all the robots arrived at their predefined target positions is shown in Fig. 3b.
The performance of the IWD algorithm is evaluated using the simulation out-
comes. The parameters such as total path travelled, path deviation, number of turns
and execution time are considered for performance evaluation. The effectiveness of

[email protected]
586 D. Chandrasekhar Rao and M. R. Kabat

(a) Isometric view (b) Top view

Fig. 2 Initial configuration of simulation environment

(a) Intermediate configuration (b) Final configuration

Fig. 3 Simulation output using IWD algorithm

IWD algorithm is further verified using the existing state of the art [10]. To verify the
effectiveness of IWD algorithm, the state of the art [10] is implemented using Matlab
and interfaced with V-rep using the same environment. The simulation is executed
for 30 times for the state of the art [10] and the simulation outcomes are noted. The
best performance out of 30 runs employing the state of the art [10] is shown in Fig. 4.
The relative performance of IWD algorithm and the existing state of the art [10] has
been verified using simulation outcomes for the above-discussed parameters. The
simulation outcome of both the algorithms for performance evaluation is noted by
taking the average of 30 runs and presented in Table 1. Further, the improvement of
the proposed IWD algorithm over the state of the art [10] for multi-robot navigation
problem is compared and presented in Table 2.

[email protected]
 Table 1 Simulation outcome of IWD algorithm and existing state of the art
Robot Position Actual path (in cm) Path travelled (in Path deviation (in No. of turns Time (in s)
cm) cm)
Initial Target IWD [10] IWD [10] IWD [10] IWD [10]
Robot-1 (198, −197) (−139, 195) 517 568 628 51 111 12 14 30.75 35.90
A Study on Cooperation and Navigation Planning …

Robot-2 (193, 197) (−204, −195) 558 827 864 269 306 12 19 45.74 52.60
Robot-3 (78, 193) 429 458 557 29 128 7 17 27.30 31.25

[email protected]
(−104, −195)
Robot-4 (−199, 200) (200, 8) 443 587 660 144 217 13 16 33.70 37.50
587
588 D. Chandrasekhar Rao and M. R. Kabat

(a) Intermediate configuration (b) Final configuration

Fig. 4 Simulation output using the state of the art [10]

Table 2 Comparative analysis between IWD algorithm and the state of the art [10]
Algorithm Total path Total path Total number Total execution
travelled (in deviation (in of turns time (in s)
cm) cm)
IWD 2440 493 44 45.74
FLC [10] 2709 763 66 52.60
% of improvement 9.93 35.26 33.33 13.04

The simulation outcomes presented in Table 1 reveals that the individual robot
performance employing IWD algorithm is superior to the existing state of the art
[10]. Further, the relative percentage of improvement of the IWD algorithm as com-
pared to the state of the art [10] in terms of total path travelled, total path deviation,
total number of turns, and total execution time is 9.93, 35.26, 33.33, and 13.04,
respectively.

5 Conclusion

In this a paper, a novel approach for solving multi-robot navigation problem using
IWD algorithm is addressed. The objective function comprises of weighted sum of
the distance to travel with respect to the target and obstacle avoidance is formulated
for the problem. The subsequent waypoints for the mobile robot from its current posi-
tion to target position are generated through the algorithm by evaluating the objective
function. The proposed algorithm effectiveness is validated through simulation. Fur-
ther, the potency of the proposed method is verified with the existing state of the art

[email protected]
A Study on Cooperation and Navigation Planning … 589

through simulation. The simulation outcomes for multi-robot navigation reveal that
the proposed algorithm outperforms the existing state of the art [10]. However, the
performance of the algorithm is not verified in the presence of dynamic obstacles.

References

1. Shah-Hosseini, H. (2008). Intelligent water drops algorithm: A new optimization method for
solving the multiple knapsack problem. International Journal of Intelligent Computing and
Cybernetics, 1(2), 193–212.
2. Lottermoser, A., Berger, C., Braunreuther, S., & Reinhart, G. (2017). Method of usability for
mobile robotics in a manufacturing environment. Procedia CIRP, 62, 594–599.
3. Stoyanov, T., Mojtahedzadeh, R., Andreasson, H., & Lilienthal, A. J. (2013). Comparative
evaluation of range sensor accuracy for indoor mobile robotics and automated logistics appli-
cations. Robotics and Autonomous Systems, 61(10), 1094–1105.
4. Stephens, K. D., Jr. (2017, February 21). Space exploration with human proxy robots. U.S.
Patent No. 9,573,276.
5. Sokolov, S., Zhilenkov, A., Nyrkov, A., & Chernyi, S. (2017). The use robotics for under-
water research complex objects. In Computational intelligence in data mining (pp. 421–427).
Singapore: Springer.
6. Bayat, B., Crasta, N., Crespi, A., Pascoal, A. M., & Ijspeert, A. (2017). Environmental moni-
toring using autonomous vehicles: A survey of recent searching techniques. Current Opinion
in Biotechnology, 45, 76–84.
7. Krishna, K. R. (2017). Push button agriculture: Robotics, drones, satellite-guided soil and crop
management. New York: CRC Press.
8. Bakhshipour, M., Ghadi, M. J., & Namdari, F. (2017). Swarm robotics search & rescue: A novel
artificial intelligence-inspired optimization approach. Applied Soft Computing, 57, 708–726.
9. Zhang, H., Butzke, J., & Likhachev, M. (2012). Combining global and local planning
with guarantees on completeness. In International Conference on Robotics and Automation
(pp. 4500–4506).
10. Zhao, R., & Lee, H. K. (2017). Fuzzy-based path planning for multiple mobile robots in
unknown dynamic environment. Journal of Electrical Engineering & Technology, 12(2),
918–925.
11. Masehian, E., & Sedighizadeh, D. (2007). Classic and heuristic approaches in robot motion
planning-a chronological review. World Academy of Science, Engineering and Technology,
29(1), 101–106.
12. Lingelbach, F. (2004). Path planning using probabilistic cell decomposition. In IEEE Interna-
tional Conference on Robotics and Automation (Vol. 1, pp. 467–472).
13. Geraerts, R., & Overmars, M. H. (2004). A comparative study of probabilistic roadmap plan-
ners. In Algorithmic foundations of robotics V (pp. 43–57).
14. Bhattacharya, P., & Gavrilova, M. L. (2008). Roadmap-based path planning-using the Voronoi
diagram for a clearance-based shortest path. IEEE Robotics and Automation Magazine, 15(2),
58–66.
15. Kim, M. H., Heo, J. H., Wei, Y., & Lee, M. C. (2011). A path planning algorithm using artificial
potential field based on probability map. In 8th International Conference on Ubiquitous Robots
and Ambient Intelligence (pp. 41–43).
16. Pradhan, S. K., Parhi, D. R., & Panda, A. K. (2009). Fuzzy logic techniques for navigation of
several mobile robots. Applied Soft Computing, 9(1), 290–304.
17. Kamkar, I., Akbarzadeh-T, M. R., & Yaghoobi, M. (2010, October). Intelligent water drops a
new optimization algorithm for solving the vehicle routing problem. In 2010 IEEE International
Conference on Systems Man and Cybernetics (SMC) (pp. 4142–4146). IEEE.

[email protected]
590 D. Chandrasekhar Rao and M. R. Kabat

18. Elhoseny, M., Shehab, A., & Yuan, X. (2017). Optimizing robot path in dynamic environments
using Genetic Algorithm and Bezier Curve. Journal of Intelligent & Fuzzy Systems, 33(4),
2305–2316. https://doi.org/10.3233/jifs-17348.
19. Dadgar, M., Jafari, S., & Hamzeh, A. (2016). A PSO-based multi-robot cooperation method
for target searching in unknown environments. Neurocomputing, 177, 62–74.
20. Ayari, A., & Bouamama, S. (2017). A new multiple robot path planning algorithm: dynamic
distributed particle swarm optimization. Robotics and Biomimetics, 4(1), 1–15.
21. Purcaru, C., Precup, R. E., Iercan, D., Fedorovici, L. O., David, R. C., & Dragan, F. (2013).
Optimal robot path planning using gravitational search algorithm. International Journal of
Artificial Intelligence, 10, 1–20.
22. Das, P. K., Behera, H. S., Jena, P. K., & Panigrahi, B. K. (2016). Multi-robot path planning in
a dynamic environment using improved gravitational search algorithm. Journal of Electrical
Systems and Information Technology., 3(2), 295–313.
23. Rao, D. C., Pani, S., Kabat, M. R., & Das, P. K. (2017, August). Cooperation of multi-robots
for obstacle avoidance in clutter environment using differential evolutionary algorithm. In
2017 International Conference on Information, Communication, Instrumentation and Control
(ICICIC) (pp. 1–6). IEEE.
24. Das, P. K., Behera, H. S., Jena, P. K., & Panigrahi, B. K. (2017). An intelligent multi-robot
path planning in a dynamic environment using improved gravitational search algorithm. Inter-
national Journal of Automation and Computing, 1–13.
25. Abaee Shoushtary, M., Hoseini Nasab, H., & Fakhrzad, M. B. (2014). Team robot motion plan-
ning in dynamics environments using a new hybrid algorithm (honey bee mating optimization-
tabu list). Chinese Journal of Engineering, 2014.
26. Rakshit, P., Konar, A., Bhowmik, P., Goswami, I., Das, S., Jain, L. C., et al. (2013). Realization
of an adaptive memetic algorithm using differential evolution and q-learning: A case study
in multirobot path planning. IEEE Transactions on Systems, Man, and Cybernetics: Systems,
43(4), 814–831.
27. Rohmer, E., Singh, S. P., & Freese, M. (2013, November). V-REP: A versatile and scalable
robot simulation framework. In 2013 IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS) (pp. 1321–1326). IEEE.
28. Adept Technology, Inc Mobile robots, “Pioneer p3-dx specifications.” www.mobilerobots.com/
Libraries/Downloads/Pioneer3DXP3DX-RevA.sflb.ashx. Accessed May 28, 2018.

[email protected]
Customer’s Activity Recognition
in Smart Retail Environment Using
AltBeacon

M. Lakshmi, Alolika Panja, Naini and Shakti Mishra

Abstract SRMS (Smart Retail Management System) is a project based on IoT

(Internet of Things), which is an upcoming technology that deserves the attention of
the industry. IoT provides unique identifiers to objects and people and transfers data
over a network without any interaction of human to human or human to computer,
for example tracking your activity level at real-time basis. In this project, we will be
empowering retail stores with the power of IoT. In this project, we will be sending
tailored offer schemes to the customer of the store whenever he/she is in the radius of
a beacon of a particular shelf in store. This process, in turn, will help the customers
to get specific offers for happier customer experience and will aid managers of the
stores to better analyse the trends of the customers and create appropriate deals to
attain higher profits with proper management of inventory.

Keywords Virtual beacons (AltBeacon) · Smart Retail Management System ·

Consumer behaviour analytics · Location tracking · Push notification · BLE
(Bluetooth Low Energy) devices

1 Introduction

India is progressing towards a Digital India with the regularly increasing technically
sound customers, there should be an effort from retail industry to move along with
their customers and come with solutions that not only save their time but also make
their shopping experience quite user-friendly.

M. Lakshmi (B) · A. Panja · Naini · S. Mishra

Department of Information Science and Engineering, Nitte Meenakshi Institute of Technology,
Bengaluru 560064, India
e-mail: [email protected]
A. Panja
e-mail: [email protected]
Naini
e-mail: [email protected]
S. Mishra
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 591

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_49

[email protected]
592 M. Lakshmi et al.

There are some serious challenges being faced by customers in the retail industry
[1], some of which are as follows:
1. Struggling along with the economy: With the increase in the cost of commodities,
it is difficult for the retails to communicate the high cost of the products to the
customers.
2. To overcome this challenge the retailers should come up with more deals and
offers to showcase for their customer’s benefit.
3. Staging stores: With the increasing demand of online shopping brick and mortar
stores need to make their place more attractive so that customers would want
to visit the stores and spend more time just as they would in destinations like
Starbucks, CCD, etc.
With the SRMS, we tried to overcome this challenge by taking their feedback into
consideration and adding more features to our app that would help the customers
to have an enjoyable experience while they shop.
4. Analysing data: Retailers do not have much information about their customers
which makes analysing their behaviour a challenging task.
In the SRMS, the preferences of every individual customer are taken into account
during the registration procedure, which will help them to find more deals and
offers which is specifically relevant to them only. The system also does real-time
analysis of their location to provide the offers present in that particular section
of the store.
5. Figuring out the potential of mobile devices: Shoppers browse similar products
online while they are shopping at the stores to compare the prices of similar
products. Hence the retailers must find a way to determine how to best take the
advantage of this technology to attract shoppers.
There is a dedicated smartphone which belongs to each and every individual
which they can use to enhance their shopping experience using the SRMS to get
personalized notifications.
6. Standing in long queues to complete the payment procedure: Shoppers usually
need to stand in long queues to complete their payment procedure.
With the SRMS the customers can directly scan the products before adding it
to the cart and complete the payment procedure digitally using their debit/credit
cards.
Our main aim was to create an IOT-based retail environment for an improved user
retailer interaction for enhancing the marketing techniques and customers experience.
SRMS is a refreshing wave that can come as a saviour for this industry by not only
enhancing the user experience of the customers but also in attaining high gains for
the industry by proper analysis of its data. SRMS uses beacon technology which is a
low-energy device. SRMS is based on this process as it uses this connection between
the customer’s mobile device and beacons located on different shelves of the store.
The data collected in this process, in turn, will help the managers to get an idea of
customer trends, popularity of products, and other store dynamics. Implementation
of this system can reduce the losses to stores, mismanagement in stores, and better

[email protected]
Customer’s Activity Recognition in Smart Retail Environment … 593

handling of inventory of stores. This can be a big step toward creating a Digital India
and also boosting other initiatives like Make in India, etc.
In this paper, we have discussed the design, working and function of Smart Retail
Management System and Conclusion. The rest of this paper is outlined as follows:
Sect. 2 gives the literature survey, Sect. 3 presents the proposed methodology and
the algorithm used to construct the SRMS, Sect. 4 explains the implementation and
the data flow of the system with code snippets and Sect. 5 gives the conclusion and
outlines of future work.

2 Related Work

2.1 Customer Behaviour Analysis

Customer behaviour analysis includes that it identifies the customer and their buying
behaviour patterns.
In the retail industry, the analysis of this information is very important and useful
as it helps to find solutions to a lot of marketing, problems and also makes the
customer’s shopping experience enjoyable. India has made a late entry into organized
retail management techniques. There has been a lot of change over the past few years
after adopting new ideas to track consumer behaviour.
In a recent survey carried out by Accenture, it was recorded that 72% of 258 North
American business leaders said they should spend more on analytics. Amy Lin, a
mother of two in suburban New York City said that it was easy to shop online rather
than physically visiting the stores. The supermarket needs to come up with more
attractive ideas to make people want to be there [1].
In 2014, Infinite Dial conducted a study which estimated that 61% among 160
million Americans, aged 12 or more, owned a smartphone [2]. It was predicted by
Business Insider intelligence, which is a research service that an annual growth of
287% with 45 lakhs of beacons by 2019 out of which 35 lakhs of beacons is used
in the retail environment. SWIRL is a mobile marketing firm. It reported that when
67% of the shoppers in retail industry received an in store alert; 81% of them opened
the alert message and 79% made a purchase related to the message [3] (Fig. 1).

2.2 Beacon Technology

1. Bluetooth Low-Energy (BLE)

BLE or Bluetooth Low Energy is a wireless network technology which is used in
many applications such as retail, health care, home entertainment, and security-
based industries. When compared to Bluetooth technology, BLE provides a sim-
ilar communication range with lower power consumption and minimal cost.

[email protected]
594 M. Lakshmi et al.

Fig. 1 Case study carried out by SWIRL

Table 1 Comparison between Low energy and Bluetooth

S. No. Technical specification Bluetooth technology Bluetooth Low-Energy
technology
1 Distance/range 100 m >100 m
2 Active slave 7 Not defined,
implementation
dependent
3 Minimum total time to 100 ms 3 ms
send data
4 Power consumption 1 W as the reference 0.01–0.50 W (depending
on use case)
5 Application throughout 0.7–2.1 Mbit/s 0.27 Mbit/s
6 Network topology Point-to-point/scatternet Point-to-point/star
7. Primary use cases Mobile phones, PCs, Mobile phones, gaming,
headsets, automotive fitness, medical,
electronics and
automotive

Table 1 depicts a comparison between Bluetooth Low Energy and Bluetooth

devices [4].

2. BLE devices and Beacon

A BLE device communicates via its service and characteristics. A service speci-
fied for BLE device may have one or more characteristics. Each service and char-
acteristics are represented by a Universal Unique ID (UUID). A Beacon device
is broadcasting BLE device. It is a new technology with an indoor positioning
system or local positioning system. This technology is based on Bluetooth low
energy (BLE). At the hardware level, BLE devices are broadcasting data and on
the other side, at the software level, beacons are messages sent by broadcasting
devices that are then detected and processed by receiver devices. UUID serves
as beacon identifier, while transmitting the data from beacon, which is one-way
discovery mechanism, i.e., beacons devices are not aware of phones. Beacon has

[email protected]
Customer’s Activity Recognition in Smart Retail Environment … 595

the benefit to consume less power and is low cost when compared to other indoor
positioning systems [5].
Hardware perspective of beacon
Beacons are small devices which are as small as a palm-sized object attached
to the walls. These devices have tiny Bluetooth radio transmitter which continu-
ously broadcasts radio signals. The signals are transmitted at regular intervals is a
combination of letters and numbers.
Beacons mainly consist of three components: CPU, Radio, and Batteries. These
devices use small lithium chip batteries or can be connected via powers like USB
plugs. Beacons are available in different shapes and colours and may include
additional features like temperature sensors, accelerometers, etc.

Software prospective of beacon

Beacons transmit a unique id number that communicates listening devices like
smartphones to which beacon it is adjacent to. The concept of beacons is often
misconceived. They are used for broadcasting a signal and not for tracking the
objects.
Beacon software specifications [5] are as follows:
• Durability—Most beacon has a battery life of more or less 1.5–2 years. Some
might range from 6 to 7 months. Beacon with can last over 60 months if they
have energy-saving capabilities since Bluetooth is incredibly energy efficient.
• Interval—It depends on a specific scenario. It can be set according to ones
need (ms-millisecond).
• TxPower—The transmission power of a beacon can range from 13 ft 1.4
inches to 164–295 ft depending on the specific use.
• Price—Beacons can cost as cheap as $5 but the average cost of beacon may
range from $15 to $25.
• Devices are not connected to beacons, they just receive radio waves broad-
casted by a beacon and calculate distance (by RSSI). So, there is no limit.
• The maximum range of a beacon is around 70 m.
3. Different types of beacons
There are various kind of becons available in the market:
• iBeacons (Apple)
• Eddystone (Google)
• AltBeacons (Radius Network).

The comparative study [6] is given in Table 2.

[email protected]
596 M. Lakshmi et al.

Table 2 Comparative study of different beacon devices

S. No. Parameters iBeacon Eddystone AltBeacons
1 Technology In December 2013, In July 2015 In July 2014
Apple announced Google announced Radius Networks
iBeacon. Built into Eddystone. announced
Apple’s iOS 7 and Google’s AltBeacon. It is an
later versions of Eddystone is a open source
mobile operating beacon protocol designed to
system that allows for open source overcome the issue
iPhones and iPads beacons, which of protocols
to constantly scan can be favouring one
for Bluetooth manufactured by vendor over the
devices nearby. any business at other
Beacons use BLE affordable costs
which is part of
Bluetooth 4.0
specification
2 Compatibility Android and iOS It is cross platform Compatible with
compatible but and thus is other mobile
native only to iOS compatible with operating
any platform that platforms
supports BLE
beacons
3 Profile It is proprietary It is an open It is an open
software thus the source. The source. The
specification is specification is specification is
controlled by published openly published openly
Apple on GitHub on GitHub
4 Ease of use It is simple to It is flexible but It is more flexible
implement requires more with customized
complicated source codes and
coding when it has the ability to
comes to have different
integration Manufacturer IDs
and different
Beacon codes
5 Broadcasted Each beacon Eddystone Each beacons
packets broadcasts broadcasts three broadcast
information which different information with
is identified as a packets—(i) a four parameter-
packet which has a unique id number s—UUID, major,
unique id number (Eddystone uid), minor and Tx
comprising of (ii) a URL address power
three (Eddystone URL),
parts––UUID, (iii) sensor teleme-
major and minor try(Eddystone
TLM)

[email protected]
Customer’s Activity Recognition in Smart Retail Environment … 597

2.3 Benefits of the Retailers

Using beacon devices in the retail shop enhances the experience of the user. The
benefits of retailers are as follows [7]:
• Beacon technology can pinpoint location more accurately than other top con-
tenders GPS, WIFI and NFC.
• It helps the retailers in blending the digital and physical realms.
• It also helps the retailers in directing customers to their purchase locations.
• It helps in maintaining consumer attention within the retail environment for a long
time than usual.
• It enables retailers to create more tailored experiences for customers, thereby
developing a good relationship with customers.
• It acts as an advertising tool which will enhance their marketing techniques.

2.4 AltBeacons

AltBeacon is a protocol specification that defines a message format for the beacon
in the same vicinity. The messages are transmitted by devices for the purpose of
signalling their proximity to nearby receivers. The contents of the transmitted mes-
sage contain information that the receiving device can use to identify the beacon and
to compute its relative distance to the beacon. The receiving device may use this
distance as a condition to execute services and implement actions that are relevant
to being in the vicinity to the transmitting beacon [8].
• AltBeacon is easier to use as it has greater transparency than other beacons to what
a beacon transmits and how the data obtained can be used by Windows, Android,
and other operating systems.
• It helps the customers as well as retailers to shift towards a shared platform for
non-iOS devices.
• In the future, AltBeacons might empower developers to take true advantage of
what a BLE beacon can do.
• AltBeacons can confidently compete against other beacons to increase the use of
beacon adoption for both iOS as well as Android devices.

3 Proposed Parallel Computing Frameworks

3.1 Methodology

Figure 2 depicts how communication is established between the different components

of SRMS. Each beacon represents different sections of the stores and has a unique

[email protected]
598 M. Lakshmi et al.

Fig. 2 Working of SRMS

minor number which helps us to identify a particular section. The first step in the
(Fig. 2) depicts when the smartphone with the installed application enters the range
the smartphone picks up the signal transmitted by the different beacons and finds
the beacon closest to the smartphone. In the second step, the smartphone sends the
beacon ID of the closest beacon to the database. In the third step, the database returns
the deals and offers associated with that particular beacon ID. The deals and offers
then appear on the smartphone as push notifications.

[email protected]
Customer’s Activity Recognition in Smart Retail Environment … 599

Fig. 3 Distance calculation

3.2 Algorithm

The proposed algorithm for implementing this methodology has been represented in
Fig. 3. The algorithm works as follows:
Step 1: Initialize the system by considering becond id  0.
Step 2: Compute the distance between all the beacons.
Step 3: The minimum distance is set to highest.
Step 4: Identify the minimal distance among all the distances. This beacon serves as
a current beacon.
Step 5: If the current beacon has not been used prior, activate the current beacon.
Start pushing the notification.
Step 6: If the beacon has already served in the past, enable all the push notification
sent earlier.

3.3 Distance Calculation

Figure 3 displays the following four distance groups: Unknown, Far, Near and Imme-
diate. These states define the distance of user from the beacon. In the figure, Unknown
state signifies that distance cannot be calculated as the user is more than 30 m away
from the beacon. Far state defines a range between 2 and 30 m. The near state ranges
from 2 to 0.5 m. The distance of the object from the beacon can be easily calculated
by this method. The last state is immediate, which is within the range of 0.5 m.
The distance calculation uses the strength of the signal from the beacon (Received
Signal Strength Indication, or RSSI) and it gives the distance in metres. As the signal
strength increased, the accuracy also increases (Algorithm 1).

[email protected]
600 M. Lakshmi et al.

Step1: Start.
Step2: Initialize System.
BID(Beacon_id)=‘0’;
Step3: Scan the distance of all the beacons using getDistance().
AltBeacon.getDistance();
Step4: Sets minimum distance to the highest.
minDist=Double.MAX_VALUE;
Step5: Calculates the minimum distance of all distances.
for (allBeaconsInRange)
if(current_beacon.getDistance()<minDist)
minDis=current_beacon.getDistance

Step6: Checks if the current_beacon_id is not equal to the previous beacon_id.

if(current_b_id !=BID)
Step7: If the condition is true it will activate the current beacon .
isActive=false;
activate(current_b_id)
Step8:If the condition is false it will show notification and sets isActive as true.
showNotification();
isActive=true;
Step9: Stop.

Algorithm 1 Beacon communication and push notification

4 Experimental Setup and Performance Analysis

4.1 Data Flow

A data flow diagram gives the diagrammatic representation of the flow of the data
through the system. It is the first step in creating an overview of the system. The flow
diagram (Fig. 4) shows the data flow of the application and the interaction between
the different interfaces of the application.
In order to avail the benefits of the application, the user first needs to register and
then login successfully that will bring them to the homepage. The homepage consists
of a navigation drawer with the following options:
1. myAccount in which the user’s details will be stored.
2. Scan and Buy which can be used by the user to scan the product directly via their
smartphone. This will add the products into the cart where they can proceed with
the online payment procedure.
3. Deals and Offers will showcase all the products available in the stores with
discounts.
4. Logout option to log out of the application.

4.2 Code Snippets

Here are some of the code snippets, which explain certain functionalities of the
application:
1. Scanning and finding beacon with minimum distance

[email protected]
Customer’s Activity Recognition in Smart Retail Environment … 601

Fig. 4 Data flow of the application

[email protected]
602 M. Lakshmi et al.

2. Send Notification

3. Scan and Buy

4. Result

The graph in Fig. 5 shows the power consumption of the Beacons with time it
spends near the smartphone which contains the application. The more time it spends
near the smartphone the more power it consumes and vice versa.

[email protected]
Customer’s Activity Recognition in Smart Retail Environment … 603

Beacons

Fig. 5 Power consumption versus time graph

5 Conclusion

In this paper, we presented the design, working and the features of our Smart Retail
Management System. Besides, this project successfully demonstrates a new role of
Bluetooth low-energy device and the use of AltBeacons in the Retail Environment.
SRMS will help the customers make the shopping experience more pleasurable and
efficient by providing them with personalized notifications and an option to avoid
long queues in the supermarket by enabling them to scan the product directly and
make their instant digital payments.
Future Work
In the future, if all the studies are favourable we would like to implement the idea
in a real-world scenario and add-on more features to attract more customers to the
store, like keeping track of purchase history of individual customers which will give
the retailers an idea about the likes and dislikes of the customers and accordingly
provide personalized customer service. In future, we would also like to provide the
customers with an indoor mapping system, which will guide them in and around the
store. We will also try to implement this application on a hybrid platform to support
smartphones with other operating systems.

References

1. Di, M. (2012, February 16). Challenges facing today’s retail industry. Amecian Express Com-
pany.
2. Tim, H., & Tom, W. (2014, October 20). The mobile commerce revolution and the current state
of mobile. Pearson Articles.

[email protected]
604 M. Lakshmi et al.

3. Moody, M. (2015, Spring). Analysis of promising beacon technology for consumers. Elon
Journal of Undergraduate Research in Communications, 6(1).
4. Bluetooth Low Energy (Online). Available https://en.wikipedia.org/wiki/Bluetooth_Low_
Energy#Hardware. Accessed March 21, 2018.
5. What is a beacon (Online). Available https://kontakt.io/beacon-basics/what-is-a-beacon/.
Accessed March 22, 2018.
6. Mallik, N. The key differences between iBeacon, Eddystone & AltBeacon (Online). Avail-
able https://www.quora.com/What-are-the-key-differences-between-iBeacon-Eddystone-
AltBeacon. Accessed March 20, 2018.
7. William, J. (2014, October). What’s the biggest influencer in consumer purchase decisions?
Small Business Trends.
8. Altbeacon Protocol Specification 1.0 (Online). Available https://github.com/AltBeacon/spec.
Accessed April 9, 2018.
9. Zaim, D., & Bellafkih, M. (2016). Bluetooth Low Energy (BLE) based geomarketing system.
In 2016 IEEE, INPT , Madinat al irfane, Rabat, Morocco.
10. Kohne, M., & Sieck, J. (2014). Location-based services with iBeacon technology. In 2014
Second International Conference on Artificial Intelligence, Modelling and Simulation.
11. Alt-Beacon (Online). Available www.altbeacon.org. Accessed March 21, 2018.

[email protected]
An Active Mixer Design For Down
Conversion in 180 nm CMOS Technology
for RFIC Applications

B. H. Shraddha and Nalini C. Iyer

Abstract This paper introduces a component of the radio frequency transceiver

called the mixer. Mixers are found in almost all the communication systems at the
front end. Radio frequency mixing is a key process within the RF technology and
RF design. It is a nonlinear process that involves one signal level affecting the other
signal levels at the output side instantaneously. The mixer design has the following
design parameters Conversion gain, Linearity, Noise Figure, and port isolation. It is
important to have better isolation between the ports as it is the measure of leakage
or feedthrough from one port to another. Poor isolation leads to mixing of unwanted
dripped signal with desired output signal creating inter-modulation products and
adding distortion. The proposed Gilbert mixer is intended to produce IF frequency
range of 1 MHz in UMC180 nm CMOS technology with a conversion gain of 8 dB,
Noise figure of >10 dB, RF frequency range 5.001 GHz, reverse isolation >15 dB,
and a stability factor of 1 at a low operating voltage of 1.8 V using a double-balanced
topology. The mixer being designed provides a better isolation factor between the
ports with less power dissipation of <10 mW.

Keywords Gilbert mixer · Conversion gain · Noise figure · Port isolation · UMC
technology

1 Introduction

Radio frequency circuitry is gaining importance in recent years. Radio frequency

communication has taken great advances from mobile phones to base stations; com-
munication industry has been revolutionizing the way the entire globe transmits
and receives information with growing demand. With this increase in the need and

B. H. Shraddha · N. C. Iyer (B)

Department of ENC, BVBCET College of Engineering and Technology,
Hubballi 580031, India
e-mail: [email protected]
B. H. Shraddha
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 605

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_50

[email protected]
606 B. H. Shraddha and N. C. Iyer

demand of communication industry, there is ample scope of expanding and creating

more reliable and efficient components. In order to realize the goals, there is a need
for increasing the frequency range software and hardware design [1].
Mixers are an important component in any of the RFIC process applications. The
processes in communication technology operate at least at ultra high frequencies
and higher than that. Amplifying a high-frequency signal is not preferable as they
induce a variety of parasitic into the design. The main purpose of making mixer an
important component in the receiver front end is that it translates a signal from one
frequency to another where the signal can be amplified or processed more effectively.
Figure 1 shows the basic architecture of the mixer that is fundamentally a multiplier.
The mixing unit has two input ports, i.e., radio frequency signal (RF) and local
oscillator signal (LO) and one output port, i.e., intermediate frequency (IF) signal,
respectively. The intermediate frequency signal contains two frequency components,
one component depicts the down conversion of original signal and the other is up
conversion compared to the frequency of the original signal. This work is intended
to implement a down conversion mixer. Figure 1 makes use of low-pass filter which
passes the down converted signal as an output.
The heterodyning process in the mixer is discussed in the following case, if both
the input signals are:

a  A cos(ω1)t (1)

b  B cos(ω2)t (2)

The mixer output will yield

a · b  A cos(ω1)t · B cos(ω2)t
AB AB
 cos(ω1 − ω2)t + cos(ω1 + ω2)t (3)
2 2
Equation 3 shows that the mixer output consists of the difference and summation
of both input frequencies. Equations 1 and 2 are used to demonstrate the heterodyning
process using alternating signals. The unwanted function can be filtered out. In the
mixer design, there are four parameters that act as mixer specifications. They are
conversion gain (CG), Linearity, Noise Figure (NF), and port isolation. The CG is

Fig. 1 Basic block of mixer

[email protected]
An Active Mixer Design For Down Conversion in 180 nm … 607

a ratio between the output signal and the input signal usually in the measure in
decibels (dB) or milli-decibels (dBm). The linearity of the mixer is defined as how
well the mixer reacts to the mixing of frequencies and ideal law of superposition in
the ideal case explained in the above text. NF is a ratio of the signal-to-noise ratio
(SNR) at the IF output and the SNR at the RF input port. Finally, the port isolation
parameter shows how much leakage of signal occurs between two ports. Besides,
the frequency of interest, there also exists the image frequency, interferers known
as spurious response that has to be eliminated using filters during the system design
[2].
Mixers are broadly classified into three types based on the number of differential
ports. They are:
• Unbalanced mixer
• Single-balanced mixer
• Double-balanced mixer.
The design and implementation of unbalanced mixer cell include all three ports
in a single-ended configuration. Single-balanced mixer cell has one port with dif-
ferential configuration, whereas the other two are still in single-ended configuration
[3]. The double-balanced mixer cells are implemented with all three ports in differ-
ential configuration. Table 1 shows the output components of three different types
of mixers.
It is evident from Table 1 that the double-balanced mixer cell is the better choice
as it does not provide any leakages from the input ports to output ports. There are
multiple mixer designs and topologies that are well suited to meet the requirements of
the system. The most used and popular mixer design is the Gilbert mixer cell design.
Unlike the mixing circuits which make use of nonlinear components such as diodes,
the Gilbert mixer cell uses linear time-invariant circuits to achieve the multiplication
of time domain signals, and hence obtaining frequency shift in frequency domain.
Gilbert cell mixer design provides balanced operation and there is a clearer signal
expected at the output [4]. The work proposed by the author Yang et al. [5] on a low-
voltage low-power and highly linear down conversion mixer operating at 900 MHz,
though having a high linearity as an advantage this design provides a very low
conversion gain of −6.5 dB. The down conversion mixer proposed by Faitah et al.
[6] works with improved conversion gain but exhibits a higher noise figure. A down

Table 1 Output components of three types of mixers

Unbalanced Single balanced Double balanced
Desired products 1
A cos(w1)t. A cos(w1)t. A cos(w1)t.
2
sgn(cos(w2)t) sgn(cos(w2)t) sgn(cos(w2)t)
LO feedthrough 1 –
2 V r f sgn(cos(w2)t) V r f sgn(cos(w2)t)
RF feedthrough 1 – –
2 (V r f +
Asgn(cos(w2)t)

[email protected]
608 B. H. Shraddha and N. C. Iyer

conversion mixer proposed by Caverly et al. [7] uses Gilbert cell configuration for
wireless applications in 900 MHz band which has a very low conversion gain of
2.7 dB at 10 MHz intermediate frequency. The author Klumperink et al. [8] have
proposed up conversion mixer using 0.18 µm technology with conversion gain of
10 dB but it has a drawback of providing a large noise figure of 24 dB. Darabi
et al. (2005) proposed an up conversion mixer with low noise figure of 11.8 dB but
the drawback is reduced conversion gain of 1 dB which is very low. The design
of a direct conversion mixer included an improvement in the conversion gain of
6.6 dB but had a drawback of increased noise figure of 21 dB [9]. Hence, this paper
promotes a design of balanced Gilbert mixer cell having a better performance in
terms of conversion gain, noise figure, port-to-port isolation, linearity, and power
consumption for various wireless applications. The brief of the paper is as follows,
Sect. 2 discusses the proposed mixer design, comparison between the conventional
Gilbert mixer cell and designed topology, Sect. 3 discusses the performance analysis
and simulation results of the implemented mixer cell, and lastly, the conclusion is
discussed in Sect. 4.

2 Proposed Mixer Design

This section discusses designed mixer cell and its comparison with the conventional
topology of Gilbert cell.

2.1 Conventional Double-Balanced Gilbert Cell

It consists of four unturned devices interconnected by multiple hybrids, transformers,

or baluns. Figure 2 shows the usual topology of double-balanced Gilbert cell mixer
[10, 11].
The mixer consists of two stages one stage is the switching stage and the other
is the transconductance stage. In the double-balanced architecture, there are two
transistors at the bottom referring to the transconductance stage and on the other
hand, the switching stage has two pairs of transistors. The LO signal is preferably a
square wave which switches the transistors M3, M4, M5, and M6 on or off depending
on the magnitude of the LO signal represented as LO+ and LO− shown in Fig. 2.
Each time when switching occurs, the current through the load moves through the
switching transistors and multiplies with the LO signal and converts the current to
the voltage. The output is differentially measured and all mixer parameters like gain
are tabulated.

[email protected]
An Active Mixer Design For Down Conversion in 180 nm … 609

Fig. 2 Double-balanced
Gilbert mixer cell

2.2 Mixer Design Parameters

Referring to Fig. 2 and solving for the output voltage, we get

4
VIF  IIF ∗ R  gmVRF R[cos(wRF − wLO )t], (4)
π
Hence, the conversion gain is given by

2
CG  ∗ gm ∗ R (5)
π
From Eq. 5, it is clear that the intermediate frequency output signal of a double-
balanced Gilbert mixer cell contains frequency components other than LO frequency
signal. Hence, this topology offers better isolation between the ports. The linearity
in terms of IIP3 is given by

32 ISS
IIP3  (6)
3 Cox WL11

The noise figure is given by

 
π2 2Υ 2
NF  1+ + (7)
4 gm . Rs . Rf gm . Rf . Rs . Rload

[email protected]
610 B. H. Shraddha and N. C. Iyer

Port-to-port isolations are given by

LOIF isolation  dBm(mix(HB . Vout2 , (0, 1))) − LOPower (8)

RFIF isolation = dBm(mix(HB . Vout2 , (1, 0))) − RFPower (9)

RFLO isolation = dBm(mix(HB . lo.0, (1, 0))) − RFPower (10)

Equation 4 is used to derive the gain of the mixer cell. Equations 8, 9, and 10
compute the dBm of the mixed outputs at desired frequencies and they are being
subtracted from the respective input ports. Based on the above equations, the aspect
ratios of the MOSFET’s are designed. The mixer cell designed is aimed to meet the
following specifications as shown in Table 2.
The tool used to design and implement the proposed mixer cell is CADENCE
tool. This tool consists of managing libraries, schematic editing, symbol creation,
test bench verification, analysis of simulation results, complete physical design and
comparison of pre- and post-simulation results. Figure 3 shows the circuit topology
of a conventional double-balanced Gilbert mixer cell. The circuit topology when
simulated for the different parameters of the mixer cell, the values obtained were
not able to meet up the set specifications. The circuit topology shown in Fig. 3 were
modified to meet up the required specifications and hence degeneration resistors
are added to the transconductance stage. The enhanced circuit topology of double-
balanced Gilbert mixer cell is shown in Fig. 4.

Table 2 Mixer cell S. No. Specifications Value

specifications
1 Conversion gain 0 dB
2 Noise figure Adequate
3 IP1 dB Optimum
4 IIP3 Optimum
5 RF frequency range 5.0–5.001 GHz
6 IF frequency 1 MHz
7 Reverse isolation >15 dB
8 Stability factor 1
9 Power consumption <10 mW
10 Technology UMC180
11 Supply voltage 1.8 V

[email protected]
An Active Mixer Design For Down Conversion in 180 nm … 611

Fig. 3 Circuit topology

Fig. 4 Enhanced circuit topology

3 Performance Analysis and Simulation Results

The test bench for simulating different parameters is shown in Fig. 5. The proposed
mixer design is simulated using UMC180 nm technology PDK files and results are
validated on Cadence Virtuoso.
The gain of the mixer evaluates the mixer’s frequency conversion action. Hence
it is termed as conversion gain. It is defined as the ratio of power delivered to the
load to the power available at the input port. The conversion gain of the mixer was
simulated by sweeping the signal amplitude of LO input and conducting the PAC and
PSS analysis. The plot in Fig. 6 reveals that the maximum conversion gain offered
by the mixer is 8 dB at the LO power of 8 all other parameters such as port isolation,

[email protected]
612 B. H. Shraddha and N. C. Iyer

Fig. 5 Test bench of the enhanced double-balanced Gilbert mixer cell

Gain in dB

Signal Power

Fig. 6 Voltage conversion gain versus LO signal power

noise figure, and power consumption are not considered. The stability of the gain
over the variation in the RF frequency is been simulated using PAC analysis. As seen
in Fig. 7 though there is a variation in the intermediate frequency from 0 to 10 MHz,
the gain is almost constant with a small variation of 8.05 dB at the higher side to
8.048 dB at the lower side. The voltage conversion gain versus RF frequency using
PXF analysis has been simulated to check the variation in the gain of the mixer.
Figure 8 reveals that though there is variation in RF frequency from 0 to 10 MHz the
gain is constant with 8.054 dB at the higher side to 8.051 dB at the lower side.
As there are three different ports in the mixer cell, there exists feedthrough leak-
ages between the ports. The different feedthroughs are measured in order to look into
the isolation of the mixer. Figures 9, 10, 11, and 12 show the RF to LO, RF to IF, LO to

[email protected]
An Active Mixer Design For Down Conversion in 180 nm … 613

Gain in dB

Frequency

Fig. 7 Mixer gain versus RF frequency

Gain in dB

Frequency

Fig. 8 Mixer gain versus RF frequency (PXF analysis)

Table 3 Simulated values for Different port feedthrough

port isolations
RF to LO feedthrough −6.4 kdB
RF to IF feedthrough −81.7 dB
LO to IF feedthrough −235.3 dB
LO to RF feedthrough −81.7 dB

IF, and LO to RF feedthrough, respectively. The PAC and PXF analysis is combined
together to simulate port isolations. Table 3 reveals the simulated results obtained
for different feedthroughs. The conversion gain is kept constant while performing
the isolation parameters.
Under small signal conditions the output signal power increases linearly with
input signal power. When the circuit operates with large signals it is affected, and
there are possibilities that the circuit operates in nonlinear region of operation. 1 dB
point is a measure for this nonlinearity. To simulate and calculate 1 dB compression

[email protected]
614 B. H. Shraddha and N. C. Iyer

Gain in dB

Frequency

Fig. 9 RF to LO feedthrough

Gain in dB

Frequency

Fig. 10 RF to IF feedthrough

Gain in dB

Frequency

Fig. 11 LO to IF feedthrough

[email protected]
An Active Mixer Design For Down Conversion in 180 nm … 615

Gain in dB

Frequency

Fig. 12 LO to RF feedthrough

Power in dBm

Power in dBm

Fig. 13 Output referred IP3

point and IIP3 of the mixer QPAC and QPSS analysis is performed. Figures 12 and
13 show the results for 1 dB compression point and IIP3 calculated at first-order
frequency, i.e., at 1 MHz and at third-order frequency, i.e., at 900 MHz. The input
referred IP3 is 6.24 dBm and output referred is −1.34522 dBm (Fig. 14).
For the input signal frequencies of 5.001 and 5 GHz, an intermediate frequency
of 1 MHz is been generated using transient analysis. Figures 15 and 16 show the IF
output signal and its frequency spectrum.
From Fig. 16, it is evident that the first harmonic, i.e., the fundamental frequency is
located at 1 MHz. After the fundamental frequency there are even and odd harmonics
of the fundamental frequency which should be located at least 30 dB away from the
fundamental frequency for better optimum functioning. In the implemented mixer
the difference between first and second harmonic signal is 66 dB and the difference
between first and third harmonic is 41.2 dB. The proposed mixer is successfully able

[email protected]
616 B. H. Shraddha and N. C. Iyer

Power in dBm

Power in dBm

Fig. 14 Input referred IP3

output

me

Fig. 15 Transient response of the IF signal

[email protected]
An Active Mixer Design For Down Conversion in 180 nm … 617

Gain

Frequency

Fig. 16 IF spectrum frequency domain

Fig. 17 IF output signal after filtering

to convert a 5.001 GHz RF frequency to a 1 MHz intermediate frequency. From

Fig. 15, it is evident that the IF output signal contains a lot of harmonics, hence, a
low-pass filter (LPF) is used at the output of the mixer designed for a cutoff frequency
of 1 MHz using VCVS. The output of the LPF is clean 1 MHz sinusoidal waveform
shown in Fig. 17, Transient analysis.
The implemented architecture of the mixer is tested for the variations in the process
corners and variation in temperature. There are 4 process corners in VLSI design
most widely used are typical typical (TT), slow-fast (SF), and fast-slow (FS). Table 4
shows the parameter analysis for process corner variation and temperature variation.
The temperature is varied from −40 °C to +125 °C.

[email protected]
618 B. H. Shraddha and N. C. Iyer

Table 4 Parameter analysis for process corners and temperature variations

Parameters TT FS SF Temperature variation
(TT)
RF frequency 5.001 GHz 5.001 GHz 5.001 GHz 5.001 GHz
LO frequency 5 GHz 5 GHz 5 GHz 5 GHz
IF frequency 1 MHz 1 MHz 0.98 MHz 1.3 MHz
CMOS process 180 nm 180 nm 180 nm 180 nm
Supply voltage 1.8 V 1.8 V 1.8 V 1.8 V
Power consumption 8.53 mW 8.5 mW 8.52 mW 8.53 mW
LO power (dBm) 8 7.68 8.04 8
Conversion gain (dB) 8 7.89 8 8.24
1-dB –IIP3 (dBm) −1.34522 −1.28 −1.31 −1.44522
Third order –IIP3 6.24365 6.64 5.94 6.64
(dBm)
RF-LO isolation (dB) −6.4 k −5.8 k −5.4 k −6.4 k
LO-RF isolation (dB) −81.6 −81.5 −81.05 −81.6
LO-IF isolation (dB) −235.35 −223.15 −219.3 −235.35
RF-IF isolation (dB) −81.6 −80.6 −80.72 −81.6

4 Conclusion

An active mixer with double-balanced topology is been designed with specifications

shown in Table 2 and simulated using UMC 180 nm technology. The variations
of the parameters for different process corners and temperature variation is shown
in Table 4. The proposed Gilbert mixer exhibits better performance in terms of
parameters like adequate gain, higher isolation between the input and output ports,
and less power consumption.

References

1. Su, Z., Dai, F. F., Wilamowski, B. M., Hamilton, M., Zhou, W., Wang,Y., & Fu, J. A 0.8
x2013;3 GHz wideband folded down-converter with noise cancellation in 0.18um SiGe tech-
nology. In 2015 IEEE Bipolar/BiCMOS Circuits and Technology Meeting—BCTM.
2. Sharma, U. K., Chaturvedi, A., & Kumar, M. (2016). A high gain down-conversion mixer
in 0.18 um CMOS technology for ultra wideband applications. In 2016 3rd International
Conference on Signal Processing and Integrated Networks (SPIN).
3. Bi Pham. Research report on “1.9 GHz Gilbert Mixer in 0.18 nm CMOS for a cable tuner”.
4. Ghayvat, H., Bandil, L., Mukhopadhyay, S. C., & Gupta, R. (2015). A 2.4 GHz CMOS Gilbert
Mixer in 180 nm technology. In 2015 Fifth International Conference on Communication Sys-
tems and Network Technologies.

[email protected]
An Active Mixer Design For Down Conversion in 180 nm … 619

5. Kim, H. -R. et al. (2003). A 900 MHz low voltage low power highly linear mixer for direct-
conversion receivers. In 2003 10th IEEE International Conference on Electronics, Circuits and
systems (vol. 3).
6. El Oualkadi, A. et al. (2008). CMOS RF down-conversion mixer design for low-power wireless
communications. ACM Ubiquity, 9(24).
7. Caverly, R. H., Smith, S., & Hu, J. (2000). RF CMOS cells for wireless applications. Analog
Integrated Circuits and Signal Processing, 25, 5–15.
8. Klumperink, E., et al. (2004). Upconversion mixer using 0.18 μm technology.
9. Darabi, H., & Chiu, J. (2005). A noise cancellation technique in active RF-CMOS mixers.
IEEE Journal of Solid-State Circuits, 40(12), 2628–2632.
10. Gibson, A., Jr. (2011). Design and simulation of CMOS active mixers. Research Thesis in the
College of Engineering and computer science, Central Florida.
11. Shraddha, B. H., & Iyer, N. C. (2017). Study of double balanced Gilbert mixer cell. International
Journal of Research and Advanced Development (IJRAD), 1.

[email protected]
Analysis of PAPR for Performance QPSK
and BPSK Modulation Techniques

K. Bhagyashree, S. Ramakrishna and Priyatam Kumar

Abstract In both wireless and wired communication environments, there is demand

for larger data rates, which is continuously increasing day by day. Hence, OFDM
systems have been developed for digital systems. These systems have many advan-
tages over single-carrier transmission systems like resistance to selective fading. But
these systems are characterized by large value of PAPR. Many methods have been
discussed in the literature for the reduction in PAPR value. Although these methods
provide the reduction, they affect the transmission power, data transmission rates,
error rates, and complexity in the computational model. One of the simplest ways
of measuring PAPR is CCDF. In this paper, CCDF curves are used to measure the
amount of PAPR in OFDM systems and are analyzed.

Keywords CCDF––Complementary cumulative distributing function ·

OFDM––Orthogonal frequency division multiplexing · PAPR––Peak-to-average
power ratio · Clipping · Subcarriers · MIMO––Multiple input multiple output

1 Introduction

The demand for 4G wireless communication systems has increased exceptionally in

the field of multimedia transmission. With the upsurge in the number of users and due
to the limited bandwidth, there is a necessity for more advanced modern communica-
tion techniques which give good spectral and bandwidth efficiency. A system which
is immune to multipath fading is coined as multicarrier system. This kind of multi-
carrier transmission systems is highly reliable and facilitates high data transmission
rates for large numbers of users. In multicarrier systems, the main highlight is band-

K. Bhagyashree · S. Ramakrishna · P. Kumar (B)

Department of ECE, BVB College of Engineering, Hubli 580031, India
e-mail: [email protected]
K. Bhagyashree
e-mail: [email protected]
S. Ramakrishna
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2019 621
N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_51

[email protected]
622 K. Bhagyashree et al.

width distributed among many subcarriers. On the contrary, single-carrier systems

have single carrier occupying the whole bandwidth. Now, these humungous features
of multicarrier systems together with OFDM features provide better performance
efficiency.
Audio–video broadcasting and mobile multimedia communications are some of
the applications which incorporate features of OFDM. Techniques like modulation
and multiplexing both are encompassed in OFDM. So, it has many benefits which
make OFDM a potential option for 4G communication systems. The OFDM is char-
acterized by subdivision of the available bandwidth into many data streams of low
data rate, and then transmitting on individual subcarriers. Using fast Fourier t trans-
forms (FFT/IFFT), modulation and demodulation are done. The modulation is done
on each subcarrier after the symbol generation. Each subcarrier has a unique central
frequency which is orthogonal to the frequency of other subcarriers. Here, orthog-
onality is maintained using IFFT. Guard bands eliminate interference between the
symbols. A wideband signal is converted to a number of narrow band signals.
Although there are many advantages, OFDM systems suffer from many problems
such as PAPR, synchronization, and Bit error rate. PAPR is the major among the
problems presented in OFDM systems. Many techniques for the reduction in PAPR
have already been presented.

2 Literature Review on PAPR Reduction

The literature on peak-to-average power ratio has been reviewed and presented in
the following section.
Foomoolijareon and Fernando proposed a solution to the problem caused by peak-
to-average power ratio in Orthogonal Frequency Division Multiplexing system in the
year 2002. Two methods were suggested [1]. The first method, wherein a list of input
vectors are maintained and an appropriate input are selected from the list. The method
is done by initially choosing the subcarriers and reducing the number in the input side
and the passing this to the Inverse Fourier Transform. Both the methods are supported
by simulation outputs showing considerable reduction in the peak-to-average ratio.
One important observation here is this was done for restricted channel numbers.
Another work on the same lines was done by Xiadong et al. The work was pro-
posed in the year 1998 which incorporated different techniques to reduce the power
ratio in the communication system [2]. Methods like filtering and clipping were
employed and spectral power density, bit error rate, etc., were the performance indi-
cators considered. Simulations indicate better results when compared to traditional
communication systems.
A reduction algorithm was proposed in the year 2011 by Wei Xeufeng. The new
algorithm proposes a new method by incorporating advantages of the conventional
method. This is a probability-based algorithm. The simulations show that the peak-
to-average power ratio reduces considerably.

[email protected]
Analysis of PAPR for Performance QPSK and BPSK Modulation … 623

2.1 Need for Orthogonal Frequency Division Multiplexing

In many of the wired and wireless transmission systems, today, Orthogonal Frequency
Division Multiplexing is like a boon. Figure 1 depicts the block level description of a
typical OFDM transmission system with the Rx and Tx blocks. Symbol interference
is caused because of high data transmission rates [3]. A guard band interleaving can
be done in two ways, one by padding zeros, which is done by including zeros in
between the symbols. The other is cyclic prefix; here, the last section of the symbol
is copied at the first section of the next symbol. The size of the guard band should be
such that it takes into account the response time of the channel to avoid interference
[2, 4]. In case of Fast Fourier Transform, cyclic prefix method is favorable than
padding zeros because this method periodizes the signal.

2.2 Effect of Peak-to-Average Power Ratio in Case

of Orthogonal Frequency Division Multiplexing Systems

Basically in Orthogonal Frequency Division multiplexing, a data stream of high data

rate is partitioned into data streams of lower rate [5, 6]. These lower data rate streams
are then transmitted at once using many subcarriers, which may overlap with one
another. At lower data rates the symbol duration increases causing dispersion in time.
The main characteristic feature of OFDM systems is large number of subcarriers.
And, high PAPR which offers problem in real-time transmission on the optical fiber

Fig. 1 Basic architecture diagram of the communication system

[email protected]
624 K. Bhagyashree et al.

cables which lead to distortions in the communication bands. This effect elevates
with the use of data converters in the system design. This leads to reduction in
the performance of the circuit. The main cause for PAPR in the transmission is the
presence of the large number of subcarrier which is not in phase. When these signals,
which are not in phase, shoot up to maximum value simultaneously the output also
raises this causes peak in the output value. In orthogonal communication systems
due to many subcarriers large peak is present as against the average value. These
demands for sophisticated transmitters which further increase the cost of the entire
system.

3 PAPR

OFDM systems are characterized by the presence of many subcarriers in the system.
Hence, the peak value is very large than the average value. This ratio is called peak-
to-average power ratio.

maximum|y(t)|2
PAPR    (1)
E |y(t)|2

Here, numerator gives the peak signal power and denominator gives average power
of signal. Let us consider the total number of subcarriers are N, and if all are in the
same phase, then the N is the PAPR. Figure 2 shows envelope peaks for the OFDM
information signal for N equals 16 which in turn leads to very large peak power [4].

Fig. 2 Relation between PDF, CDF, and CCDF

[email protected]
Analysis of PAPR for Performance QPSK and BPSK Modulation … 625

3.1 PAPR Measurement

To measure and evaluate the performance of PAPR reduction scheme, complementary

cumulative disturbing function (CCDF) is a standard candidate. If CDF and CCDF
are compared, then it is revealed that what value of the information signal is below a
required level is indicated by CDF and what amount of signal is at a given level and
above is depicted by CCDF. The probability is defined by the percentage of time the
data signal remains at a level or above [7].
The real and imaginary sections of signal in OFDM with many number of subcar-
riers is considered to be Gaussian. Under such situations, the signal envelope in the
case of OFDM has Rayleigh distribution and power is exponential. If one assumes
unity power for OFDM signal, then normalized Rayleigh distribution is PDF, that is,
Probability Distribution Function [8].

4 Proposed Methodolgy

In the proposed method, a lower order modulation scheme has been incorporated
like QPSK as it provides robustness against noise. But it is noted that it provides
lower data rates. Here, MATLAB is used for simulation. The following are the steps
followed in the method:
i. Select the input size, subcarrier number and symbol number
ii. Generate modulated signals
iii. Assign the symbols modulated to the subcarriers
iv. Compute inverse FFT
v. Calculate PAPR
vi. Plot PAPR versus CCDF graph.

5 Results

In the results, the graphs of PAPR values for different subcarrier numbers are shown.
The modulation techniques used here are QPSK and BPSK.
The peak value of the envelope would be directly affected as the subcarrier number
is increased, and this is due to large PAPR value. Figure 3, shows the simulated results
of CCDF curves in case of an OFDM system with different number of subcarriers.
Here. the 16-QAM modulation is used. As there an incremental change in the number
of subcarriers, the PAPR value also raises. Major values are in the range 7–10 dB at
CCDF 10-2.
Figure 4 depicts for BPSK the values of PAPR are lower than the other modulation
schemes. For any value of M in M—Quadrature amplitude modulation, the value of
PAPR is almost the same.

[email protected]
626 K. Bhagyashree et al.

Fig. 3 Plot of CCDF for different values of N

Fig. 4 Plot of CCDF for modulation using BPSK, QPSK in OFDM (M  2, 4, 6 and N  512)

[email protected]
Analysis of PAPR for Performance QPSK and BPSK Modulation … 627

6 Conclusion

In this paper, a method for obtaining the PAPR in transmission systems has been
proposed. There are many ways to estimate the PAPR; one of them is based on the
probabilistic approach. The reduction of PAPR is the primary issue because of the
cost and complexity of different components in OFDM Tx and Rx is affected. The
most basic parameter which evaluates the performance of PAPR reduction scheme
is CCDF. The PAPR values for a 16-QAM with varying subcarriers range 7–10 dB
at CCDF 10-2.

References

1. Wang, Z. P., Xiao, J. N., Li, F., & Chen, L. (2011). Hadamard precoding for PAPR reduction in
optical direct detection OFDM systems. Optoelectronics Letters, 7(5), 363–366.
2. Lee, B. M., Figueiredo, R. J. P., & Kim, Y. (2012). A computationally efficient tree–PTS
technique for PAPR reduction of OFDM signals. Wireless Personal Communications, 62(2),
431–442.
3. Wu, X., Wang, J., Mao, Z., & Zhang, J. (2010). Conjugate interleaved partitioning PTS Scheme
for PAPR reduction of OFDM signals. Circuits; Systems and Signal Processing, 29(3), 499–514.
4. Hassan, E. S., Xu, Z., Khamy, S. E., Dessouky, M. I., El-Dolil, S. A., & Abd El-Samie, F.
E. (2012). Peak to average power ratio reduction using selective mapping with unequal power
distribution. Journal of Central South University, 19(7), 1902–1908.
5. Renze, L., Longjiang, J., Lang, L., Jie, L., & Weile, Z. (2006). Reducing the peak to average
power ratio of OFDM system with low complexity. Journal of Electronics (China), 23(1), 26–28.
6. Hasan, M. M. (2014). PAPR reduction in OFDM systems based on autoregressive filtering.
Circuits, Systems and Signal Processing, 33(5), 1637–1654.
7. Lee, B. M., Kim, Y., & Figueiredo, R. J. P. (2012). Performance analysis of the clipping scheme
with SLM technique for PAPR reduction of OFDM signals in fading channels. Wireless Personal
Communications, 63(2), 331–344.
8. Baig, I., & Jeoti, V. (2012). A new DCT matrix precoding based RI-OFDMA uplink system for
PAPR reduction. In 4th International Conference on Intelligent and Advanced Systems (p. 680).

[email protected]
Implementation of Modified Array
Multiplier for WiMAX Deinterleaver
Address Generation

Patil Nikita, Arun Kakhandki, S. Ramakrishna and Priyatam Kumar

Abstract One of the simpler techniques, which involve the implementation of gen-
erating the address of two-dimensional deinterleaver used in the WiMAX transmitter
and receiver block is proposed using the Xilinx FPGA. The Arithmetic and Logic Unit
performs various mathematical operations such as addition, subtraction, division and
many other logical operations. Apart from these operations, Multiplication is one of
the most fundamental operations to be carried out by this unit. The implementation
of multipliers is required for the address generation of the channel interleaver. The
multipliers need to be designed in such a way that they require high speed, low power,
less area, and less delay, which is of significant interest in the research area. Many
attempts have been carried out to reduce the generation of number of partial products
in the process of multiplication. Array multiplier is one such multiplier. Most of the
arithmetic operations will be performed using multipliers which consume the major-
ity amount of the power in digital circuits. The process of multiplication involves
the shift and add operations. The performance of multiplier can be improved by
optimizing the adder circuit. The objective of this paper is to build the algorithm for
generating the address of channel deinterleaver by using modified array multiplier.
The algorithm is built using the Hardware Description Language Verilog and the
functioning of the system can be verified through simulation and implemented using
the Spartan-6. The simulation and implementation results have been obtained for the
three different modulation techniques such as QPSK, 16-QAM, and 64-QAM for
some of the information rates which proves to be a very good technique against the
conventional methods.

Keywords Field-programmable gate array (FPGAs) · Multiplication · Array

multiplier

P. Nikita · S. Ramakrishna · P. Kumar (B)

Department of EC, BVB College of Engineering, Hubli 580031, India
e-mail: [email protected]
S. Ramakrishna
e-mail: [email protected]
A. Kakhandki
Department of EC, VDRIT College of Engineering, Haliyal 581329, India

© Springer Nature Singapore Pte Ltd. 2019 629

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_52

[email protected]
630 P. Nikita et al.

1 Introduction

Wireless Communication refers to the type of communication in which the data can
be transmitted, as well as received through wirelessly. Accessing the data through
wireless communication with large rate of bandwidth is one of the more competing
tasks compared to the wired communication technology. There are certain standards
given by IEEE for carrying out the wireless mobile communication which is widely
known as mobile WiMAX [1]. The channel interleaver/deinterleaver included in
the WiMAX Tx and Rx block plays a very important role in reducing the effect of
contiguous sequence of symbols that result in errors. The address generator circuit is
designed and implemented for the channel deinterleaver which is used in the WiMAX
Tx and Rx in such a way that requires high speed and less resources. The circuit is
designed with low complexity, as well as it invalidates the necessity of generating
floor function. There is a very small amount of work performed on the hardware
implementation part of the interleaver/deinterleaver used in a WiMAX system as
discussed in the literature review [2, 3].
In [4], the authors have described and demonstrated about grouping the incoming
data streams in the form of blocks to minimize the occurrence of accessing the
memory in a deinterleaver using a traditional LUT-based approach. The work is
realized by using Complementary Metal–Oxide–Semiconductor address generator
for WiMAX. In [5], the authors have carried out the work by implementing the
address generator for IEEE 802.16e channel interleaver with only a ½ code rate by
using a hardware description language. In [6], the authors have demonstrated with
the help of FSM-based address generator of the same interleaver for the different
information rates and modulation techniques. The authors of [5, 6] have implemented
on the FPGA platform. The authors of [7] have performed the two-dimensional
transformation of the functions which are used in the WiMAX channel deinterleaver
for using the hardware architecture efficiently. But the design issues have not been
discussed clearly for 64-QAM [8].
Implementation of the floor function in the hardware circuit is very complicated
since it requires more amount of resources [7]. One of the traditional techniques,
which were used previously, is the look up table-based approach. This technique
proved to be less efficient since it performed the operations slowly, utilization of
large amount of logic resources, and so on. The proposed design provides the impro-
visation with respect to the resources, operations to be carried out by the system com-
pared to the conventional look up table approach. Complex and lengthy expressions
were used for the modulation techniques, due to the two-dimensional transformation
of the functions used in the WiMax channel deinterleaver. The proposed system uses
the minimized mathematical expressions and the resulting algorithm is presented.
The mathematical equations have been proved using [7]. The proposed algorithm
can be realized with the help of hardware circuit, which results in the architecture for
address generation having less complexity, compared with the previous technique.
The design is optimized in such a way that it requires a common hardware circuit
among the modules for all the three modulation techniques [9]. The three modulation

[email protected]
Implementation of Modified Array Multiplier for WiMAX … 631

techniques that have been used are QPSK, 16-QAM, and 64-QAM. The proposed
algorithm consists of implementation of modified array multiplier, which helps in
the reduction of power along with less area and improving the delay of circuit. This
architecture is developed using hardware description language Verilog and imple-
mented on the Spartan-6 FPGA. The proposed algorithm is intended to produce the
address generated by the channel interleaver and the functioning of the hardware
circuit is verified using Isim.

2 Design Methodology

This section describes about the overview of WiMAX transreceiver and proposed
algorithm.

2.1 WiMAX System

Due to the advancement in the field of wireless communication technology, WiMAX

is one of the most emerging technologies that provide accessing for fixed and mobile
users. The overview of WiMAX transmitter and receiver block is shown in Fig. 1.
As seen in the below figure, the most important blocks of WiMAX are transmitter
and receiver in order to transmit as well as receive the data wirelessly.
The description of the transmitter section is as follows. Source is one of the most
important parts of the transmitter. It is used to transmit the information or data through
the channel to one or more receivers. The data received from the source is sent to
the randomizer. This operation is usually carried out for a certain set of data that can
be transmitted or received in a cyclic manner. This operation is performed on each

Fig. 1 Overview of WiMAX transreceiver

[email protected]
632 P. Nikita et al.

burst on each allocation in order to avoid lengthy sequence of continuous logical

ones and zeros. Data is further encoded by using both the coding techniques. The
coding techniques used are RS and CC. The encoded data is further applied to the
channel interleaver [10, 11]. The channel interleaver performs the permutation on
the encoded data to minimize the consequence of continuous sequence of symbols
which results in errors. The signals are further modulated and composed using the
other two consecutive blocks, such as mapper and inverse fast Fourier transformation
technique as shown in Fig. 1. The receiver block consists of the arrangement of the
blocks in the bottom-up approach so that it enables the data to be restored in the
elementary sequence at the output.
The functional representation of interleaver/deinterleaver is shown in Fig. 2. It pro-
vides the internal structure of the channel interleaver/deinterleaver in the WiMax sys-
tem. The channel interleaver/deinterleaver plays a very important role in the WiMAX
system. It consists of two memory blocks, multiplexers, and an address generator.
The read and write operations are performed by the memory blocks depending upon
the select lines. The data is written into the first memory block (M-1) when select
line is selected as one. The read operation is performed on the interleaved data from
the second memory block (M-2). Thus, read and write operations are performed
simultaneously on the memory blocks depending on the status of select lines. Once
the data is read/written up to the specified location as mentioned by the depth of
interleaver, the position of the select signal will be switched to alter the read as well
as write operations.
The interleaver/deinterleaver block accommodates distinct depth rates to include
different information rates and modulation schemes for IEEE 802.16e. The data

Fig. 2 Functional representation of interleaver/deinterleaver

[email protected]
Implementation of Modified Array Multiplier for WiMAX … 633

which is encoded with the help of coding techniques is processed by using the
Eqs. (1) and (2) which are noted below.
   
X cbps k
nk  · (k%m) + (1)
m m
    
nk m · nk
pk  r · + nk + X cbps − %r (2)
r X cbps

As seen in the above equations, m represents the number of columns. Here, it is

considered as (m  16/12 for WiMAX); the values of nk and pk, which are considered
as the result obtained as soon as encoding of data is performed. The value of k as seen
in the above equation can range from 0 to X cbps − 1; r is given by r  X cpc /2; basically
fixed for different techniques. It can be considered as two, four, or six for QPSK,
16-QAM, or 64-QAM, respectively. The inverse operation which is performed by
the deinterleaver is also defined by the equations which are mentioned below. The
values of np and kp represent the permutations obtained for the deinterleaver, where
p is considered as the index value of the bits received within a block of X cbps bits.
 p  
m.p

np  r . + p+ %r (3)
r X cbps
 
 m . np
kp  m . np − X cbps − 1 . (4)
X cbps

Table 1 discusses about the depth of interleaver/deinterleaver for various infor-

mation rates with modulation techniques.

Table 1 Addresses determined for channel interleaver

Row No. (p) Column No. (i) 0 1 2 3 4
0 X cbps  96 bits, ½ code rate, 0 16 32 48 64
1 QPSK 1 17 33 49 65
2 2 18 34 50 66
3 3 19 35 51 67
0 X cbps  192 bits, ½ code rate, 0 16 32 48 64
1 16-QAM 17 1 49 33 81
2 2 18 34 50 66
3 19 3 51 35 83
0 X cbps  192 bits, ½ code rate, 0 16 32 48 64
1 64-QAM 17 33 1 65 81
2 34 2 18 82 50
3 3 19 35 51 67

[email protected]
634 P. Nikita et al.

2.2 Proposed Algorithm

The mathematical analysis for generating the address using address generator for
the WiMAX deinterleaver has been shown in the above equations. The value of m is
chosen as 16. The number of rows (p) is predetermined which are equal to m for all
the various rates of interleaver depth which are in bits. The number of columns (i) is
obtained by dividing the interleaver depth by m. The equations used for generating
the addresses for all the three modulation techniques are as follows.

kn , Q P S K  m ∗ i + p for all the values of p and i (5)

⎫
m ∗i + p for p%2  0 and for i ⎬
kn , 16 − QAM  m ∗ (i + 1) + p for p%2  1 and for i%2  0 (6)
⎭
m ∗ (i − 1) + p for p%2  1 and for i%2  1
⎫
m ∗i + p for p%3  0 and for i ⎪
⎪
⎪
m ∗ (i − 2) + p for p%3  1 and for i%3  2 ⎪⎪
⎬
kn , 64 − QAM  m ∗ (i + 1) + p for p%3  1 and for i%3  2 (7)
⎪
m ∗ (i + 2) + p for p%3  2 and for i%3  0 ⎪⎪
⎪
⎪
m ∗ (i − 1) + p for p%3  2 and for i%3  0 ⎭

As seen in the Eqs. (5), (6), and (7), p considers the values from 0, 1, …, m
− 1 and i considers the values from 0, 1, …, (X cbps /m) − 1 represents the row
and column numbers, respectively, and k n represents the addresses of deinterleaver.
Using the above Eqs. (5), (6), and (7), the addresses have been determined for all the
three modulation schemes. Table 2 shows the addresses determined for all the three
modulation techniques with code rates.
The algorithms can be further transformed into the circuits as shown in Figs. 3
and 4, respectively. The hardware circuit for the Quadrature Phase Shift Keying is
shown in Fig. 3. It consists of various blocks such as row counter, comparator, column
counter, and multiplexer. The row counter initiates with generating the row numbers
between 0 and m − 1. The column counter which is built along with the Mux M0

Table 2 Depth of interleaver/deinterleaver for various information rates with modulation tech-
niques
Modulation scheme QPSK (s  1) 16-QAM (s  64-QAM (s  3)
2)
Code rate 1/2 3/4 1/2 3/4 1/2 2/3 3/4
Interleaver depth, X cbps in bits 96 144 192 288 288 384 432
192 288 384 576 576 – –
288 432 576 – – – –
384 576 – – – – –

[email protected]
Implementation of Modified Array Multiplier for WiMAX … 635

Fig. 3 Hardware structure of QPSK

Fig. 4 Hardware structure of 16-QAM and 64-QAM

and comparator C0 is used for generating the column numbers to implement the
allowable coded bits per second termed as X cbps . The multiplier and an adder unit
are used to perform the required operations to produce the value of k n (address of the
deinterleaver). The hardware circuits for the 16-QAM and 64-QAM are similar to
the circuit of the QPSK. The additional modules required for the design of 16-QAM
and 64-QAM are an incrementer, which increases the values by 1 and provides them
to the mux, an decrementer, which decreases the values by 1 and transmits them
to the mux, modulo blocks as shown in Figs. 4 and 5. The operation of modulo 2
is performed by the hardware circuit of 16-QAM and the operation of modulo 3 is
performed by the hardware circuit of 64-QAM as mentioned in the above part of the
Eqs. (6) and (7).

[email protected]
636 P. Nikita et al.

Fig. 5 Schematic structure

of deinterleaver address
generator

The Schematic structure of the complete deinterleaver address generator is shown

in Fig. 5. It consists of row counter (j), column counter (i), QPSK block, 16-QAM
block, 64-QAM block, and multiplexer. The multiplexer is used to select the mod-
ulation techniques depending upon the combination given to the select lines. The
combination of multiplier and adder units is used to generate the address for the
deinterleaver.

3 Simulation and Implementation Results

The functioning of the hardware circuit of the deinterleaver address generator is ver-
ified by programming using the Verilog with the help of Xilinx tool. The simulation
results have been obtained for all the three different modulation techniques using
Isim.
The simulation results obtained for the 16-bit modified array multiplier is shown
in Fig. 6. The results are generated for the six different combinations of data. This 16-
bit modified array multiplier is utilized in the hardware circuit of schematic structure
of deinterleaver address generator. The usage of modified array multiplier in this
circuit significantly reduces the power.
The simulation results obtained for the QPSK is shown in Fig. 7. The deinterleaver
address has been generated for the QPSK as shown in Table 2. The results have been
verified using Table 2.
The simulation results obtained for the 16-bit QAM is shown in Fig. 8. The
deinterleaver address has been generated for the 16-bit QAM as shown in Table 2.
The results have been verified using Table 2.
The simulation results obtained for the 64-bit QAM is shown in Fig. 9. The
deinterleaver address has been generated for the 64-QAM as shown in Table 2. The
results have been verified using Table 2.

[email protected]
Implementation of Modified Array Multiplier for WiMAX … 637

Fig. 6 Simulation results of 16-bit modified array multiplier

Fig. 7 Simulation results of QPSK

Fig. 8 Simulation results of 16-QAM

Fig. 9 Simulation results of 64-QAM

[email protected]
638 P. Nikita et al.

Table 3 Synthesis report of the proposed algorithm

Logic circuits Quantity
Adders/Subtractors 19
Counters 5
Registers 13
Comparators 21
Multiplexers 581

Table 4 Performance of the proposed algorithm with the existing algorithm

Different No. of registers No. of LUTs No. of bonded Max frequency
approaches IOs (MHz)
LUT-based 62 714 26 210.402
approach
16-bit modified 47 220 30 204.505
array multiplier

Fig. 10 Power analysis of the algorithm

The implementation details are discussed in Tables 3 and 4. Table 3 specifies

about the logic circuits required for the proposed algorithm and Table 4 specifies
about the performance of the proposed algorithm with the existing method. It shows
the comparison of parameters between the two different approaches. As per the com-
parison carried out between the two approaches, in terms of resources, the modified
array multiplier accommodates with less number of resources. The power analysis of
the algorithm is shown in Fig. 10. The dynamic power consumed is nearly 0.005 W
and the quiescent power consumed is nearly 0.020 W. The total amount of power
consumed is 0.025 W.

4 Conclusion

This paper represents the three different modulation techniques of address generation
using the 16-bit modified array multiplier. The proposed algorithm is transformed into
the hardware circuit for all three different modulation techniques. The results have

[email protected]
Implementation of Modified Array Multiplier for WiMAX … 639

been obtained using the Xilinx tool and implementation is performed using Spartan-6.
Thus, it is concluded that the address values for the channel interleaver/deinterleaver
are generated by using 16-bit modified array multiplier. The implemented results
discuss about the utilization of resources for the system. The comparison between the
two approaches such as LUT-based approach and by using modified array multiplier
is carried out in terms of resources and frequency. The power analysis is also carried
out for the system.

References

1. Konhauser, W. (2006). Broadband wireless access solutions-Progressive challenges and poten-

tial value of next generation. Wireless Personal Communications, 37(3/4), 243–259.
2. Li, B., Qin, Y., Low, C. P., & Gwee, C. L. (2007). A survey on mobile WiMAX. IEEE Com-
munications Magazine, 45(12), 70–75.
3. Andrews, J. G., Ghosh, A., & Muhamed, R. (2007). Fundamentals of WiMAX: Understanding
broadband wireless networking. Upper Saddle River, NJ, USA: Prentice-Hall.
4. Chang, Y. N., & Ding, Y. C. (2007). A low-cost dual mode de-interleaver design. In Proceedings
of International Conference on Consumer Electronics (pp. 1–2).
5. Khater, A. A., Khairy, M. M.,& Habib, S. E. D. (2009). Efficient FPGA implementation for
the IEEE 802.16e interleaver. In Proceedings of International Conference on Microelectronics
(pp. 181–184). Marrakech, Morocco.
6. Upadhyaya, B. K., Misra, I. S., & Sanyal, S. K. (2010). Novel design of address generator for
WiMAX multimode interleaver using FPGA based finite state machine. In Proceedings of 13th
International Conference on Computer and Information Technology (pp. 153–158). Dhaka,
Bangladesh.
7. Asghar, R., & Liu, D. (2009). 2D realization of WiMAX channel interleaver for efficient
hardware implementation. In Proceedings of World Academy of Science, Engineering and
Technology (vol. 51, pp. 25–29), Hong Kong.
8. IEEE Standard for Local and Metropolitan Are Networks-Part 16: Air Interface for Fixed
Broadband Wireless Access Systems-Amendment 2. (2005). IEEE Std. 802.16e-2005.
9. Khan, M. N., & Ghauri, S. (2008). The WiMAX 802.16e physical layer model. In Proceedings
of IET Conference on Wireless, Mobile and Multimedia Networks (pp. 117–120). Mumbai,
India.
10. Wallace, C. A suggestion for a fast multiplier. IEEE Transaction on Electronic Computers.
11. Nikita, P., & Kakhandki, A. (2017). Study of low power array multiplier for real time applica-
tions. International Journal of Research and Advanced Development (IJRAD), 1(3).
12. Srikanth, S., et al. (2016). Low power array multiplier using modified full adder. In 2016 IEEE
International Conference on Engineering and Technology (ICETECH). IEEE.

[email protected]
Link Quality-Based Mobile-Controlled
Handoff Analysis Using Stochastic
Models

S. Akshitha and N. G. Goudru

Abstract With limited available frequency spectrum and increasing demand for
cellular communication services, the problem of handoff becomes increasingly
important. Poorly designed handoff schemes tend to generate very heavy signal-
ing traffic and thereby, a drastic decrease in the quality of service (QoS). In urban
mobile cellular radio system, the cell size is small. The handoff procedure has a
significant impact on the system performance. Blocking probability of originating
calls and forced termination probability of ongoing calls reflects the performance of
the system. In this paper, we use mobile-controlled handoff process. MS is moving
from one cell to another cell. Relative signal strength and relative residual energy are
measured. Important channel quality parameters, for the channels with line of sight
(α) and without line of sight (β) are considered. Using these parameters, weights of
the channels are determined in three neighboring cells and the servicing cell. The
channel weights of four cells are compared with the chosen signal strength threshold
value to make a decision of handoff process or to continue service with the same base
station (BS). Prioritized hard handoff is preferred in the work. Handoff performance
metrics such as arrival rate of originating calls, blocking probability of originating
calls, arrival rate of handoff calls, blocking probability of handoff calls, and net-
work parameters such as sender window size, queue length at the servicing link are
measured. Performance analysis parameters such as packet loss due to congestion,
round-trip delay, and throughput are discussed. It helps in controlling the oscillatory
behavior of the switch or router normally used for traffic control at the BS. Using
the results and applying an explicit feedback, the link congestion and packet losses
can be minimized. It helps for improving the quality of service (QoS) in 3G, 4G, and
next-generation mobile communication. Stochastic models are used to analyze the
system performance. Using MATLAB programme, result analysis is made through
graphs and statistical data.

S. Akshitha (B) · N. G. Goudru

Department of ISE, Nitte Meenakshi Institute of Technology (Affiliated to Visvesvaraya
Technological University), Bangalore 560064, India
e-mail: [email protected]
N. G. Goudru
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 641

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_53

[email protected]
642 S. Akshitha and N.G. Goudru

Keywords Cell · Congestion · Hard handoff · Mobility · Probability · Stochastic

1 Introduction

Mobility is one of the most important features of a cellular communication system.

In mobile wireless communication system, continuous service is achieved using a
procedure called handoff. During handoff process, the channel associated with the
current servicing connection is changing while a call is in progress. It is usually
initiated either by crossing a cell boundary or by deterioration in the quality of signal
in the current channel. In the handoff literature, two types of handoff procedures are
dominant, namely hard handoff and soft handoff. Hard handoff is characterized by
“break before make” and soft handoff is characterized by “make before break”. In
hard handoff, the current resources are released before new resources are used. In
soft handoff, both existing and new resources are used during the process. Hence,
we are interested in hard handoff. In hard handoff, multiple access techniques such
as frequency-division multiple access (FDMA) and time-division multiple access
(TDMA) are used. In cellular structure, different frequency ranges are used in adja-
cent channels in order to minimize the channel interference.
Figure 1a shows that MS is moving from base station (BS1) to adjacent base
station (BS2). Mobile station (MS) is serviced by BS1 and moving toward BS2 and
no handoff is taken place. Figure 1b shows that MS has entered into the handoff
region and handoff take place.

Fig. 1 a Before handoff.

b After handoff

BS1 MS
BS2

(a) Before handoff

BS1 MS
BS2

(b) After handoff

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 643

C34
C33 C35
C32 C23 C36

C22 C24 C37

C31
C25 R=3
C21 C12
C11 R=2
C13
MT R=1

Fig. 2 Cell configuration of MT for r  3

A most widely used cell structure of the coverage of mobile towers (MT) is shown
in Fig. 2. The actual design segment taken is at 120° orientation of the coverage area.
The mean signal strength of BS1 decreases as MS moves away and mean signal
strength increase as MS approach BS2 .
In prioritized hard handoff model, we assume the cell radial distance r is small
and mobile callers are equally distributed in the coverage area. Mobile station is
moving with a constant velocity in any direction from one radial distance to another.
Probability of direction of the movement either away from the current base station
or along the same radial level from current BS toward new adjacent BS is equal.
Movement of MSs in the cells is shown in Fig. 3. The movements are random and
no restriction is imposed but handoff is possible only when an MS either crosses
cell boundary or relative signal strength (RSS) is lower than the conveniently chosen
threshold value.

1.1 Relative Signal Strength Indicator (RSSI) and Threshold

Value

RSSI is the relative received signal strength in a wireless environment. RSSI is an

indication of the power level being received by the MS after the antenna. Therefore,
the higher the RSSI number, the stronger the signal. The BS controller (BSC) is in
charge of RSSI management. BSC takes care of release of the radio channel and
handoff management. The performance of RSSI scheme between two adjacent BSs
is shown in Fig. 4. When threshold value is higher than RSS value, handoff occurs at
position A. When threshold value is lower than RSS value, MS would delay handoff
until the current signal level crosses the threshold value at position B. When the
threshold value is greater than RSS value, handoff would occur at position C. When

[email protected]
644 S. Akshitha and N.G. Goudru

the delay is more so that MS drift too far into the new cell, reduces the quality of
link from BS1 and may result in connection drop [1].

a1

a2 a6

A
g1 c2 c1
g6
a5 c6
g2
a3 a4
G C
b1 b6
g3 b2 c5
g5 c3
g4 c4
B
d1
f1 b5
f6
d6
b3
f2 b4 d2
F D
f5 e1
f3 e6 d5
d3
f4 e2 d4
E

e3 e5

e4

Fig. 3 Movement of the MSs

Fig. 4 Signal strength between two BSs for potential handoff

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 645

1.2 Handoff Prediction Technique

In mobile-controlled handoff process, each MS is completely in control of the handoff

process. This type of handoff has a short reaction time of 0.1 s. MS continuously
measures the signal strengths from the neighboring BSs. A handoff will be initiated
if the signal strength of the serving BS is lower than threshold value. MS checks for
the best candidate BS based on the weight of the channel. Figure 5 illustrates the link
quality maintenance process.

2 Related Work

In digital cellular system, a BS includes an antenna, a controller and a number

of receivers. Mobile telecommunication switching office (MTSO) connects cells
between mobile units. Two types of channels are available between MS and BS.
Control channels is used to exchange information for connection setting up and main-
taining. Traffic channels which carry voice or data between users. With the growing
demand for mobile services, many researchers are working on for achieving faster
and guaranteed quality of mobile services. The researchers B.B. Madan and others
discuss mobile-assisted handoff (MAHO) where mobile terminal (MT) assists to the
serving base station system (BSS) and mobile switching center (MSC) in making
handoff decisions. To avoid the chances of handoff drops due to non-availability of
channels, MAHO is combined with channels reserved for handoff. This makes sys-
tem implementation at BSS more complicated [2]. The researchers Amine Boubekri
and others use both link strength between mobile and potential static parent nodes
and residual energy of the parent nodes to make handoff decisions. They ignore the
node energy state, and the results are discussed through simulation using simula-
tors [3]. The authors Kishore Trivedi and others present a closed-form solutions to
the probabilities of blocking and dropping in wireless cellular networks. A fixed
point iteration method is used to estimate handoff arrival to BS. Allowing failures
and repairs of channels tried to give performance analysis for cellular system [4].
The authors Ing-Ray Chen and others apply algorithmic-based admission control
has been used by the authors to analyze a class of partitioning and threshold. They
give “change by time” scheme for optimal resource allocation without sacrificing
QoS requirements [5]. The researchers Buyurman Baykal and others preset expected
visitor list (EVL) method to achieve minimum handoff blocking probability keep-
ing QoS levels and also minimize the handoff latency. The call admission control
(CAC) run against each EVL entry. Without resource reservation or resource alloca-
tion, the analysis is made for varying network conditions [6]. The authors Biswajit
Bhowmik and others discuss on modeling hard handoff scheme and implementing
traffic model to study the handoff behavior for wireless mobile networks. They tried
to calculate originating calls, probability of handoff blocking. [7]. The authors Tao
Liu and others suggest a link estimation method to predict link quality status. It is a

[email protected]
646 S. Akshitha and N.G. Goudru

Fig. 5 MS link quality maintaining process

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 647

three-stage implementation, involving data collection, offline modeling, and online

prediction. Logistic regression model is regarded as a best model for prediction [8].
The researchers Qiuming Liu and others study the influence of network topology on
delay analysis in large-scale wireless ad hoc networks. The packet transmission delay
is used as an important parameter in the network design. Applying routing algorithm
deduced the service curve [9]. The research reviews reveal that very less work has
been discussed on link quality-based mobile-controlled hard handoff analysis using
stochastic models.

3 System Model for Mobile Wireless Networks

With reference to [3], the link quality metric is given by

rssin (t)
f n (t)  (1)
maxrssin (t)

where rssin is the radio signal strength indicator value measured after time t and
maxrssin is the best signal strength indicator value measured. The residual energy of
the BS is given by

En (t) − Ecn (t)

Ern (t)  (2)
En (t)

where En (t) is the total energy available at the BSs initially and Ecn (t) is the consumed
energy. The link weight is estimated by using the relation

1
Wn  a · + β · Ern (t) (3)
f n (t)

where α and β are parameters associated with the channel quality. The parameter α
is associated with line of sight and the parameter β is associated for receiver without
line of sight. The parameter α and β are estimated by using the relations
 2
d0
α  Pr (d0 ) (4)
d
 
d
β  10n logn (5)
d0

With reference to [10], the sender window dynamics is given by

∂w 1 w(t)w(t − R(t)) w(t − R(t))

 − p(t − R(t)) + β (6)
∂t R(t) 2R(t − R(t)) R(t − R(t))

[email protected]
648 S. Akshitha and N.G. Goudru

The sender TCP window operates on the principle of additive increase and mul-
tiplicative decrease (AIMD) congestion control strategy. The parameter α is the rate
of decrease in source window and normally fixed as 0.5. L a (t) is arrival rate of loss
due to congestion at any time t and L a (t)  w(t−R(t))
R(t−R(t))
p(t − R(t − R(t))). The loss
is proportional to packet sending rate. The first term on right-hand side (RHS) of
the Eq. (2.1) represent linear growth of cwnd, until congestion occurs in the bottle-
neck link, second term deals with congestion control scheme using feedback from
RED system, and third term L t (t) refer to bit error loss and immediate-recovery in
wireless. But transmission loss is proportional to the sending rate, therefore L t (t) is
proportional to w(t−R(t))
R(t−R(t))
. The queue dynamics is given by

∂q(t) N w(t)
 − Cd (7)
∂t R(t)

where w(t)/R(t) give an increase in queue size because of packets arrival from
N—TCP flows. C d  q(t)/R(t) represent decreasing queue because of departure
of packets after servicing from the router. In the proposed scheme, service time is
variable. The most important advantages of using instantaneous queue length over
average queue length is faster detection of congestion at the BS.
The buffer size at the BS is given by

Cd
Wmax  R(t) + M (8)
S
The round-trip time is given by

q(t)
R(t)  + Tp (9)
C
The wireless loss predictor using normal delay gradient is given by
  
RTTi − RTTi−1 Wi + Wi−1
fNDG  (10)
RTTi + RTTi−1 Wi − Wi−1

The mathematical version of congestion feedback scheme is given by

⎧ 
⎪
⎪ 0, q(t) ∈ 0, tmin
⎨

P(t)  q(t) − tmin Pmax , q(t) ∈ tmin , Wmax (11)
⎪
⎪ Wmax −t
min
⎩
1, q(t) ≥ Wmax

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 649

4 Simulation and Performance Analysis

A number of experiments using simulation were conducted to study the performance

of mobile assisted handoff. The proposed network model with one BS is illustrated
as shown in Fig. 6. Similarly, in mobility, we use four base stations BS1, BS2, BS3,
and BS4. Computers 1, 2, and 3 are the sources, R1 and R2 are the routers, and D1,
D2 , and D3 are the mobile stations which are destinations. In the experiment, the
link between R1 and R2 is the bottleneck link. Packetsize is 1000 bytes, Pmax 
0.01 s, queue buffer at the router has a minimum threshold value, t min  300 packets,
maximum threshold value, W max  500 packets, initial RTT  10 ms, C d  10 Mbps,
Cu  500Kbps, N  10, β  0.1, t p (propagation delay)  10 ms. S  20, Sc  10,
+Sr  10.
The graphs in Figs. 7, 8, 9, and 10 shows the variation of relative signal strength as
MS moves away from BSs. The reason behind considering four BSs is that the range
of 120° has coverage of three adjacent BSs and the other one is current serving BS.
The BSs 1, 2, 3, and 4 are assigned with −30 to −60, −61 to −90, −91 to −120, and
−121 to −150 MHz frequencies. The RSSI has maximum value nearer to BSs and
decrease as MS moves away from the BSs. The variation of RSSI corresponding to
BSs 1, 2, 3, and 4 varies over the range [0.9917, 0.5167], [0.9944, 0.6778], [0.9958,
0.7583], and [0.9967, 0.8067], respectively.
The graphs of Figs. 11, 12, 13, and 14 depict the variation of the residual energy
as MS moves away from BSs. The residual energy level decrease as MS moves away
from BSs. The variation of energy levels corresponding to BSs 1, 2, 3, and 4 varies
over the range [0.4790, 0.0159], [0.3184, 0.0081], [0.2417, 0.0055], and [0.1906,
0.0041], respectively.
The graphs of Figs. 15, 16, 17, and 18 shows the variation of the link weight as
MS moves away from BSs. The link weight levels decrease slowly as MS moves

Fig. 6 Network model

[email protected]
650 S. Akshitha and N.G. Goudru

0
10
Link Quality of BS-1, -60 MHz to -30 MHz

Relative RSSI

30 35 40 45 50 55 60
Radio frequency

Fig. 7 RSSI variation of BS1

0
10
Link Quality of BS-2
Relative RSSI

60 65 70 75 80 85 90
Radio frequency

Fig. 8 RSSI variation of BS2

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 651

0
10
Link Quality of BS-3

Relative RSSI

90 95 100 105 110 115 120

Radio frequency

Fig. 9 RSSI variation of BS3

0
10
Link Quality of BS-4
Relative RSSI

120 125 130 135 140 145 150

Radio frequency

Fig. 10 RSSI variation of BS4

[email protected]
652 S. Akshitha and N.G. Goudru

0
10
Residual Energy wrt BS-1

Residual Energy

-1
10

-2
10
30 35 40 45 50 55 60
RF

Fig. 11 Residual energy at BS1

0
10
Residual Energy wrt BS-2

-1
10
Residual Energy

-2
10

-3
10
60 65 70 75 80 85 90
RF

Fig. 12 Residual energy at BS2

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 653

0
10
Residual Energy wrt BS-3

-1
10
Resisual Energy

-2
10

-3
10
90 95 100 105 110 115 120
RF

Fig. 13 Residual energy at BS3

0
10
Residual Energy wrt BS-4

-1
10
Residual Energy

-2
10

-3
10
120 125 130 135 140 145 150
RF

Fig. 14 Residual energy at BS4

[email protected]
654 S. Akshitha and N.G. Goudru

0
10
Link Weight wrt BS-1

weight

-1
10

-2
10
30 35 40 45 50 55 60
RF

Fig. 15 Variation of link weight versus RF (BS1)

0
10
Link Weight wrt BS-2
weight

-1
10

-2
10
50 55 60 65 70 75 80 85 90
RF

Fig. 16 Variation of link weight versus RF (BS2)

away from the BSs. The link weight levels corresponding to BSs 1, 2, 3 and 4 varies
over the range [0.5140, 0.0858], [0.3417, 0.0858], [0.2559, 0.0858], and [0.2046,
0.0858] respectively.

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 655

0
10
Link Weight wrt BS-3

weight

-1
10

-2
10
60 70 80 90 100 110 120
RF

Fig. 17 Variation of link weight versus RF (BS3)

0
10
Link Weight wrt BS-4
weight

-1
10

-2
10
80 90 100 110 120 130 140 150
RF

Fig. 18 Variation of link weight versus RF (BS4)

[email protected]
656 S. Akshitha and N.G. Goudru

4
Arrival rate of handoff calls
3.8

3.6

3.4

3.2
Handoff calls

2.8

2.6

2.4

2.2

2
0 2 4 6 8 10 12 14
radius

Fig. 19 Arrival rate of handoff calls versus radius

In cell structure, the adjacent cell is considered to be just upper or lower level
where MS is moving toward or away from the BS. The handoff request arrival rate
in radius order is demonstrated by a graph in Fig. 19. The statistical data illustrates
that λH varies over the range of [3.3333, 3.9310].
The graph presented in Fig. 20 demonstrates the growth rate of blocking proba-
bility of originating calls in different radial distances. The value of BH varies over
the range [0.001, 0.026].
The graph presented in Fig. 21 demonstrates the growth rate of blocking proba-
bility of handoff request calls in different radial distances. The value of BH varies
over the range [0, 1].
The variation of TCP source window size is an important factor for the evaluation
of performance of throughput. Graph of Fig. 22 show the source window dynamics.
It varies over the range of [7, 67] packets.
Figure 23 describes the behavior of queue in the bottleneck link due to handoff
blocking. Queue varies over the range of [14, 500] packets. During simulation, it
is observed that queue size is larger, round-trip time is higher causing reduction in
sender window size. By maintaining smaller queue length, performance of through-
put can be improved.
The graph of Fig. 24 illustrates round-trip time delay distributed over the radius.
The rtt-delay changes in the interval [0.032 s, 0.3832 s]. Analysis of the graph leads
to a conclusion that smaller the rtt-delay, faster the end-to-end transportation and
higher the sending rate. In reality, if rtt value of latest packet is smaller than previous
packet, the network ignores that value for some time. Thus, an efficient rtt value can
be selected to use as metric for balance load of handoff request calls (λH ).

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 657

0.03
Blocking prob.of originating calls

0.025

0.02
Handoff

0.015

0.01

0.005

0
0 2 4 6 8 10 12 14
radius

Fig. 20 Blocking probability of originating calls versus radius

-8
x 10
2.5
Blocking prob.of Handoff request calls

1.5
Handoff

0.5

0
0 2 4 6 8 10 12 14
radius

Fig. 21 Blocking probability of handoff calls versus radius

[email protected]
658 S. Akshitha and N.G. Goudru

70
sender window dynamics

60

50

40
packets

30

20

10

0
0 2 4 6 8 10 12 14
radius

Fig. 22 Dynamics of sender window versus radius

600
queuelength at receiver link

500

400

300
packets

200

100

-100
0 2 4 6 8 10 12 14
radius

Fig. 23 Queue length at receiver MS link versus radius

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 659

0.4
round trip time
0.35

0.3

0.25
rtt(s)

0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14
radius

Fig. 24 Variation of round-trip delay versus radius

300
Packet loss due to Congestion

250

200
packets

150

100

50

0
0 2 4 6 8 10 12 14
radius

Fig. 25 Packet loss due to congestion versus radius

The packet loss because of congestion is illustrated in the graph of Fig. 25. The
main cause of congestion occurrence is due to blocking of handoff calls. The graph
of Fig. 26 demonstrates the loss of data packets due to handoff blocking.

[email protected]
660 S. Akshitha and N.G. Goudru

250
Packet loss due to handoff block

200

150
packets

100

50

0
0 2 4 6 8 10 12 14
radius

Fig. 26 Packet loss due to handoff blocking versus radius

300
Variation of throughput
280

260

240

220
packets

200

180

160

140

120
0 2 4 6 8 10 12 14
radius

Fig. 27 Variation in throughput versus radius

The performance of average throughput is presented in the graph of Fig. 27. As MS

move longer distance away from the BS, the performance of throughput decreases
by small amount and fluctuates over the rang of [140, 170] packets.

[email protected]
Link Quality-Based Mobile-Controlled Handoff Analysis … 661

5 Conclusion

The paper presents a novel research work on handoff management. An important

factor that defines packet loss is handoff call blocking. Packet loss during the hard
handoff is responsible for deterioration in the quality of service. Data packets that
are transported from the old base station till the time MS gets connected to the
new BS are lost. The packet loss is also proportional to the handoff loop time. The
results discussed faithfully convincing that model-based analysis is one of the best
approaches for improving the QoS is mobile wireless communications. Through this
analysis, it is possible to minimize many unnecessary handoff request calls, even
when the signal of the current base station is still at an acceptable level. This helps
in energy saving.

References

1. Zeng, Q. A., & Agrawal, D. P. (2002). Handoff in wireless mobile networks. In I Stojmenović
(Ed.) Handbook of wireless networks and mobile computing. Wiley & Sons, Inc., ISBN 0-471-
41902-8 ©2002.
2. Madan, B. B., Dharmaraja, S., & Trivedi, K. S. (2006). Combined guard channel and mobile
assisted handoff for cellular networks. IEEE Transactions on Vehicular Technology, 1–9.
3. Boubekri, A., Ajib, W., & Boukadoum, M. (2017). EAM: Energy aware mobility over wire-
less sensor networks. In 2017 IEEE 30th Canadian Conference on Electrical and Computer
Engineering (CCECE). 978-1-5090-5538-8/17/$31.00 ©2017 IEEE.
4. Trivedi, K. S., Dharmaraja, S., & Ma, X. (2002). Analytic modeling of handoffs in wireless
cellular networks. Information Sciences, 148(1–4), 155–166. ELSEVIER.
5. Chen, I.-R., Yilmaz, O., & Yen, I.-L. (2006). Admission control algorithms for revenue opti-
mization with QoS guarantees in mobile wireless networks. Wireless Personal Communica-
tions, Springer, 38, 357–362.
6. Akan, Ö. B., & Baykal, B. (2005). Handoff performance improvement with latency reduction
in next generation wireless networks. Wireless Networks, Springer Science, 11, 319–332.
7. Bhowmik, B., Pooja, Sarkar, P., & Thakur, N. (2012). Modeling prioritised hard handoff man-
agement scheme for wireless mobile networks. International Journal of Networks and Infor-
mation Security, 4(8), 21–32.
8. Liu, T., & Cerpa, A. E. (2014). Data-driven link quality prediction using link features. ACM
Transactions on Sensor Networks, 10(2), 37.
9. Liu, Q., & Jiang, X. (2017). Per-flow end-to-end delay bounds in heterogeneous wireless
networks. In 3rd IEEE International Conference on Computer and Communications. 978-1-
5090-6352-9/17/$31.00 ©2017 IEEE.
10. Goudru, N. G., & Vijaya Kumar, B. P. (2016). Enhancement of performance of TCP using
normalised throughput gradient in wireless networks. Journal of Information and Computing
Science (JIC), 11(3), 214–234. (Published by world Academic Press, England, UK).

[email protected]
A Comparative Analysis of Lightweight
Cryptographic Protocols for Smart Home

Rupali Syal

Abstract Smart home is an application of Internet of Things. A smart home is a

network of smart resource-constraint devices which require secure and confidential
transmission of information. Lightweight cryptographic protocols are suitable for
IOT applications like smart home. Lightweight cryptographic protocols are charac-
terized by small key size, small block size, and less number of iterations. This paper
presents a comparative analysis of lightweight cryptographic protocols in terms of
the encryption type, advantages, and resistance to attacks.

Keywords Internet of things · Lightweight cryptography · Smart home

1 Introduction

Internet of Things commonly called as IOT is the network of smart devices which
communicate with each other. Within the network, many devices are resource-
constrained devices with limited battery, processing power, and memory. These
points become weak points in the IOT architecture which are susceptible to attacks.
For maintaining the integrity of information, the need is to transfer crucial infor-
mation from these smart devices in encrypted format. Lightweight cryptographic
protocols encrypt data and are computationally less intensive.
Lightweight cryptographic protocols can be broadly classified as block ciphers,
hash functions, message authentication codes, and stream ciphers [1]. Lightweight
block ciphers are characterized by small block length, small key size, and less num-
ber of iterations. Some lightweight block ciphers are PRESENT, SIMON, SPECK,
RC5, TEA, and XTEA. Lightweight hash functions have small message size, small
internal state, and output size contrary to conventional hash functions, which have
large internal state size and high power consumption. The examples of lightweight
hash functions are PHOTON, Quark, and SPONGENT. Message authentication code
generates tags based on message and key which serve as authentication token which

R. Syal (B)
Computer Science and Engineering, Punjab Engineering College, Chandigarh, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 663

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_54

[email protected]
664 R. Syal

can be verified by the recipient [1]. Stream ciphers are state ciphers that operate
on plaintext bits and keystream to generate ciphertext stream. The stream ciphers
are categorized based on the use of linear/nonlinear feedback shift register. Grain,
Trivium, Mickey, and Fruit are the examples of stream ciphers [2].
Lightweight cryptography can be implemented both in hardware and software.
The target hardware devices for implementation are RFID and sensors [1]. IBM has
launched the world’s smallest computer, which has computing power of ×86 chip. It
has several hundred thousand transistors. It can perform functions like monitor, ana-
lyze, communicate, and act on data [3]. The future smart devices will have such tiny
computer embedded, which can perform software encryption. For hardware imple-
mentation of lightweight ciphers, the NIST report specifies the resource requirement
in terms of gate equivalents and logic blocks for ASIC and FPGAs, respectively. For
software implementations, the resources used are in terms of registers, RAM, and
ROM.
This paper presents a brief survey of various lightweight cryptographic proto-
cols in Sect. 2. Section 3 presents the various levels of security required in a smart
home scenario. This section also discusses the various ways for providing security
at a particular level. Since IOT architecture of smart home has various resource-
constraint devices, therefore, lightweight cryptographic protocols are suitable for
hardware/software implementation in IOT. Section 4 presents comparative analysis
of lightweight cryptographic protocols in terms of encryption type, advantages, and
resistance to attacks. Section 5 gives the conclusion.

2 Related Work

NIST has presented a report on standardization of lightweight cryptographic algo-

rithms. The authors have surveyed various block ciphers, hash functions, and stream
ciphers for implementation on resource-constraint devices. The metrics for hardware
and software are discussed. A questionnaire is presented at the end which considers
various aspects of lightweight cryptographic protocols [1].
Lightweight cryptography is required for secure and efficiency in end-to-end
communication in applications of IOT. The complexity and overhead associated
with lightweight cryptography is less compared to conventional cryptographic tech-
niques. The authors [4] have concluded that block ciphers have a practical use in the
field of IOT. At present, an ultra-lightweight block cipher is suitable for extremely
constrained environments like RFID tags, smart cards and sensor networks [5]. It is
based on substitution and permutation. PICO [6] is an ultra-lightweight block cipher
based on substitution and permutation network that operates on 64-bit plaintext and
has key length of 128 bits. It has compact key scheduling like SPECK.
Block ciphers SIMON and SPECK are lightweight ciphers with flexible block
and key sizes. They support block sizes 32, 48, 64, 96, and 128 bits and three key
sizes with each block size [7]. Simon is a Feistel-based block cipher which requires
three XOR operations in each round. LiCi, an ultra-lightweight block cipher is based

[email protected]
A Comparative Analysis of Lightweight Cryptographic Protocols … 665

on Feistel network. It has 31 rounds and operates on 64-bit plaintext and 128-bit key.
The data complexity of LiCi is 264 . It has 52 and 48 active S boxes in linear and
differential trails, respectively [8]. The design of S box is robust as single bit change
in input does not result in single bit change in output.
Feistel based block ciphers apply round function to only one half of input block,
whereas substitution permutation-based block ciphers apply slightly complex round
function. The latter can be implemented with fewer rounds as compared to more
executions of round function in former to maintain security [9].
A tiny encryption algorithm, TEA based on Feistel iteration and large number
of rounds was presented in [10] where the authors presented the C code for TEA.
Encryption and decryption in TEA use mixed algebraic group operations. A fast TEA
algorithm for embedded and mobile devices which minimize the memory require-
ments were presented in [11].
Stream ciphers are based on Vernam cipher and use a random keystream based
on their internal state. They are categorized into synchronous and self-synchronous
ciphers based on the updated operation of internal state. Based on feedback opera-
tion, the stream ciphers are categorized into linear-feedback shift registers (LFSR)
and feedback with carry shift registers. The authors [12] have discussed alternative
ciphers that use simple operations like add, rotate and XOR operations which result
in fast, efficient, and timing attack-resistant implementations.
Lightweight ciphers were implemented on 8-, 16-, and 32-bit processors to eval-
uate the execution time, RAM footprint, binary code size, and cycle count [13].
Lightweight Encryption Scheme (LES) [14] employs the technique of identity based
and stateful encryption. The authors state that the scheme provides flexibility in
encryption key management and efficiency in encryption process.
Security is required at all layers of IOT architecture [15]. At physical layer, it
includes RFID and wireless sensor security. At communication layer, it requires
secure transmission over network. At application layer, the software must be pro-
tected from malwares. The authors [15] have presented the advantages and disad-
vantages of various lightweight ciphers.

3 Smart Home: An Application of IOT

A smart home is a home with many smart devices like smart refrigerators, smart
washing machines, smart light control systems, smart security mechanisms, and
other electronic devices, which communicate with each other. These devices can be
controlled by members of the family from within the home or through a remote loca-
tion with a smart phone or a laptop [16]. The smart devices can be controlled either
through voice, remote control, tablet or smartphone. The communication between
smart devices within a home also raises the issue of confidentiality, integrity, and
authenticity.
IOT can provide an efficient and smarter security mechanism for the smart home
[17, 18]. The enhancement to various levels of security is as follows:

[email protected]
666 R. Syal

(a) Physical security

Smart door locks are physically accessible to anyone. The inbuilt sensors in door
locks can capture the fingerprint of invader and transmit to central monitoring
home system. Based on the data obtained from fingerprints, an alarm can be
sent to home members about tampering of door lock. Physical access to other
home devices can be controlled either by fingerprint matching or through voice
control.
(b) Logical Security/ Access to hardware or data
The smart resource-constraint devices must be protected with a firewall or gate-
way which has strong security mechanism. The authentic users are allowed to
control smart devices remotely. User behavior must be analyzed and predictive
mechanisms must be implemented in gateway/firewall.
(c) Message confidentiality
The message sent from smart device must be encrypted for secure transmission.
Since the smart devices are resource constrained, therefore, lightweight ciphers
will be more suitable for more efficiency.

4 Comparative Analysis of Lightweight Algorithms

This section presents the comparative analysis of lightweight cryptographic protocols

(LWCP). The LWCP can be implemented on hardware and software. Lightweight
ciphers based on identity-based encryption, SP network, Feistel structure, and stream
cipher are compared based on encryption type, advantages, and resistance to attacks
in Table 1.

4.1 Ciphers

4.1.1 Identity-Based Encryption

The identity of smart devices is represented by string which acts as a public key. A
private key is generated for the public key. Though the scheme does not require a
certificate but requires pairing and is computationally more expensive.

4.1.2 Substitution Permutation Network

PRESENT is a block cipher with block size 64 bits and key size 80 bit or 128 bit. It
has 31 rounds and each consist of XOR operation for round key, SBoxLayer, and p
Layer.

[email protected]
A Comparative Analysis of Lightweight Cryptographic Protocols … 667

Table 1 Comparative analysis of lightweight ciphers

References Lightweight encryption Advantages Attacks
type
Lightweight Combines identity Flexibility in Indistinguishable
Encryption based and stateful encryption key encryption against
Scheme [14] Diffie–Hellman management chosen plaintext attack
encryption Efficiency gains for
encryption process
Does not require
complex certificate
handling
PRESENT [5, 64-bit block- 80 and One of the first Susceptible to side
9, 19] 128-bit keys lightweight ciphers channel attacks,
31 rounds Efficient on hardware Related key attack,
Substitution platform as it uses fully Equivalent key attack,
Permutation Network wired diffusion layer Differential attack
4-bit S boxes
PICO [6] Ultra-Lightweight Large number of Active Thwart Linear and
Block Cipher S boxes in fewer rounds Differential attacks
Substitution Robustness and Good resistance against
Permutation Network Avalanche effect due to Biclique attacks
Operates on 64-bit very strong substitution
plaintext supports key layer
length 128 bits Strong and Compact
32 rounds Key Scheduling
Good Performance on
Hardware and Software
LiCi [8] Ultra-Lightweight More number of active Resist Linear and
Block cipher S boxes in minimum Differential attacks
Feistel-Based Network number of rounds Good resistance against
Operates on 64-bit Good Performance on Biclique and Zero
plaintext with 128-bit Hardware and Software correlation
key length and
generates 64-bit
ciphertext
31 rounds
TEA [9, 10, 11] Feistel structure Simplicity Susceptible to
Takes 32-bit words. Ease of implementation equivalent key attack
64-bit input divided and related key attack
into two parts
128-bit key divided into
four parts
64 rounds
Does not use S boxes
FRUIT [2] Synchronous Stream Efficiency in terms of Relatively secure to
Cipher hardware constraints Related key attacks,
Use Linear and Cube attack, algebraic
nonlinear Feedback attack and Weak key IV
shift register (initial vector)

[email protected]
668 R. Syal

4.1.3 Feistel Network

TEA, a tiny encryption algorithm assumes 32-bit words. It is based on Feistel struc-
ture and has large number of rounds. The key is divided into 4 vectors and data is
divided into 2 vectors each 32-bit words. It takes a key schedule constant and the
cipher runs for 32 rounds.

4.1.4 Stream Cipher

FRUIT an ultra-lightweight cipher works on 80-bit register with 43-bit LFSR and
37-bit NFSR, 80-bit key and 70-bit public initial vector (IV) [2].

5 Conclusions

A smart home requires a lightweight, robust, and secure cryptographic protocol for
confidential and integral communication between smart resource-constraint devices.
A trade-off exists between simplicity (in terms of block/key size, number of itera-
tions) and resistance to attacks. It is imperative for the lightweight protocol to thwart
attacks so that the smart home devices work as per the directions given by the family
members. This paper presents a comparative analysis of lightweight ciphers in terms
of cipher type, their advantages, and resistance to attacks. PRESENT was one of the
first lightweight ciphers. TEA is simple but susceptible to related and equivalent key
attack. Block ciphers like LiCi, PICO, and stream cipher FRUIT have good resistance
against attacks.

References

1. McKay, K., Bassham, L., Turan, M. S., & Mouha, N. (2017). Report on lightweight cryptog-
raphy. NISTIR 8114.
2. Bhasin, A., & Mishra, G. (2016). Recent advances in lightweight stream ciphers. Special issue
REDSET 2016 of CSIT 2016, pp. 1–4.
3. https://mashable.com/2018/03/19/ibm-worlds-smallest-computer/#om8X3Zzhugqp.
Accessed on March 24, 2018.
4. Katagi, M., & Moriai, S. (2008). Lightweight cryptography for the internet of things. Sony
Corporation.
5. Bogdanov, A., Knudsen, L. R., Leander, G., Paar, C., et al. (2007). PRESENT: An ultra-
lightweight block cipher. In Cryptographic Hardware and Embedded Systems, CHES2007
(Vol., 4727, pp. 450–466) Springer.
6. Bansod, G., Pisharoty, N., & Patil, A. (2016). PICO: An ultra lightweight and low power
encryption design for ubiquitous computing. Defence Science Journal, 66(3), 259–265.
7. Nithya, R., & Kumar, D. S. (2016). Where AES is for internet, SIMON could be for IOT.
Procedia Technology, 25, 302–309.

[email protected]
A Comparative Analysis of Lightweight Cryptographic Protocols … 669

8. Patil, J., et al. (2017). LiCi: A new ultra-lightweight block cipher. In ICEI (pp. 40–45).
9. Hatzivasilis, G., Fysarakis, K., Papaefstathiou, I., & Manifavas, C. (2017). A review of
lightweight block ciphers. Journal of Cryptographic Engineering, 1–44.
10. Wheeler, D. J., & Needham, R. M. (1994). TEA a tiny encryption algorithm. In Fast Software
Encryption (LNCS, 1008, pp. 363–366). Springer.
11. Hunn, S. A. Y., et al. (2012). The development of tiny encryption algorithm (TEA) crypto-core
for mobile systems. In ICEDSA (pp. 45–49).
12. Manifavas, C., Hatzivasilis, G., Fysarakis, K., & Papaefstathiou, Y. (2016). A survey of
lightweight stream cipher for embedded systems. Security Communications Network (Wiley),
1226–1246.
13. Dinu, D., et al. (2015). Triathlon of lightweight block ciphers for the internet of things. Cryp-
tology ePrint Archive Report.
14. Al Salami, S., Baek, J., Salah, K., & Damiani, E. (2016). Lightweight encryption for smart
home. In 11th International Conference on Availability, Reliability and Security (pp. 382–388).
15. Naru, E. R., et al. (2017). A recent review on lightweight cryptography in IOT. In International
Conference on IT in Social, Mobile, Analytics and Cloud (pp. 887–890).
16. https://www.smarthomeusa.com/smarthome/. Accessed on 28 March 2018.
17. Madakam, S., & Date, H. (2016). Security mechanisms for connectivity of smart devices in the
internet of things. In Connectivity frameworks for smart devices, computer communications
and networks (pp. 23–41). Springer (Ch 2).
18. https://www.iotforall.com/iot-physical-security-technology/. Accessed on March 29, 2018.
19. Singh, S., et al. (2017). Advanced lightweight encryption algorithms for IOT devices: Sur-
vey, challenges and solutions. Journal of Ambient Intelligence and Humanized Computing
(Springer).

[email protected]
Algorithm Study and Simulation Analysis
by Different Techniques on MANET

Nithya Rekha Sivakumar and Abeer Al Garni

Abstract This paper provides a protocol in ROUGH method which governs

APBMAN method and FLOODING method to manage the route request packets
on Fisheye State Routing in Grid. ROUGH method finds the runner node set by the
discovery of the similarity relation between the one-hop and two-hop neighbors. In
this paper, comparison is made with the results of three techniques such as ROUGH
method, FLOODING method, and APBMAN method in Grid FSR protocol. It is
found that ROUGH method has shown improved performance in several important
parameters like Throughput, energy consumption, Packet Delivery Ratio, Delay,
Overhead, and Normalized Overhead with respect to Pause time. Certainly, ROUGH
method is good with effect to Speed in Average Consumed Energy, Total Consumed
energy, Packet Delivery Ratio, and Throughput. But, there is a certain increase in
Delay as there is also an increase in Speed. Delay is well decreased in FLOODING
method.

Keywords Weighted rough set (ROUGH) · Propagation neighborhood

information algorithm (FLOODING) and probabilistic broadcasting algorithm
(APBMAN) · Speed · Pause time

1 Introduction

The well-known Fisheye State Routing (FSR) protocol [1, 2] determines a route when
no route exists or route breaks. To reduce flooding, it is necessary to create a new
path from source to destination. This paper provides a protocol in Weighted Rough
Set model (ROUGH method), which regulates APBMAN method and FLOODING

N. R. Sivakumar (B)
Department of Computer Sciences, College of Computer and Information Sciences,
Princess Nourah Bint Abdulrahman University, Riyadh 11671, Kingdom of Saudi Arabia
e-mail: [email protected]
A. Al Garni
Department of Information Technology, College of Computer and Information Sciences,
Princess Nourah Bint Abdulrahman University, Riyadh 11671, Kingdom of Saudi Arabia

© Springer Nature Singapore Pte Ltd. 2019 671

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_55

[email protected]
672 N. R. Sivakumar and A. Al Garni

method to regulate the route request packets in Grid FSR. ROUGH method discovers
the runner node set by finding the similarity relation among the one-hop and two-hop
neighbors [3]. ROUGH method deliberates the object (node) importance also and this
method gives better results compared with FSR in Grid. In this research, comparison
is made with the results of ROUGH method in Grid Fisheye State Routing (FSR)
protocol with APBMAN method and FLOODING method. By determination, it is
found that ROUGH method has improved concert in several important parameters
like Throughput, Energy Consumption, Packet Delivery Ratio, Delay, Overhead, and
Normalized Overhead with respect to Pause time. Also, another comparison is made
with the same parameters with respect to Speed.

2 Related Work

2.1 Probabilistic Flooding with FSR Protocol in Grid

The concepts highlighted in this paper [3–5] with the great need for a new broadcast-
ing strategy that can dynamically adjust the broadcast probability to take into account
the current state of the node in two hops in order to ensure a certain level of control
over rebroadcasting, and thus help to improve reachability and saved rebroadcasts.
The easy way to implement by guaranteed message dissemination is the straight-
forward broadcasting. In this scheme, a source broadcast messages to every neighbor
who, in turn, rebroadcasts received messages to its neighbors [6]. This procedure
lasts, until all reachable nodes have received and rebroadcast the message once. In
this scheme, the nodes broadcast the information with a probability P, by getting a
broadcast information for the first time, so that all the node has the same probability
to rebroadcast the message, regardless of its number of neighbors [7]. In condensed
networks, multiple nodes share similar transmission range. Hence, the probability
governs the frequency of rebroadcasts and thus could save linkage resources without
moving delivery ratios. It should be noticed that in scarce networks, there is much
less shared coverage, thus some nodes will not receive all the broadcast messages,
unless the probability parameter is high. Another possible area for improving includes
investigating the effect of nodes transmission ranges on the rebroadcast probability
[5, 8].

2.2 Propagation Neighborhood Information with FSR

in Grid

When there is demand, the nodes obtain the routes by using the flooding method to
attain the routes. It is reduced evenly also when the flooding is done in the initial
stage. The cache and the time out period are used for every route. Here, the time

[email protected]
Algorithm Study and Simulation Analysis by Different … 673

out value is chosen and the nodes are determined. The nodes are well-found with a
small cache to save the routes. However, the variation in the approach comes from
the fact that the expiry of the time out period does not trigger an update. After the
time out period, the routes are removed from the cache. The primary emphasis of the
protocol is on distribution information about the neighborhood. In other words, the
neighborhood reflects the entries of routes in the cache. It also checks the information
from other neighbour nodes.

2.3 Weighted Rough Set FSR in Grid

The proposed protocol zeroes in reduction of the redundant flooding in Route Request
Phase (RREQ) of FSR [3, 9]. In this technique, when there is change in topology, a
special hello messages are announced. It carries not only the existing status of the
neighbor node but also sends neighbor node attributes. In this to provide the quick
response to create new routes, the routing tables are structured within the neighbor
nodes. The primary objective in the existing FSR algorithm is more effectively uti-
lized in the present work as follows. The broadcast of discovery packets take place
only when needed.
The algorithm of the Improved FSR in Grid RREQ Period and the algorithm for
Relative Weight Calculation is explained in my previous research which follows this
technique for the below comparison results.

3 Design of Implementation Algorithm Study

3.1 Algorithm 1

Step 1 All nodes in the network send hello message periodically.

Step 2 Neighbor nodes receive hello message and update Neighbor table with expire
time.
Step 3 If neighbor timer is expired, nodes checks for Neighbor expire time with
current clock time and if current clock time is greater than Neighbor expire
time then Neighbor entry is purged from Neighbor Table.
Step 4 Nodes periodically calculates own probability value using the formula,
Probability  (Maximum number of Neighbors − current number of Neigh-
bors)/Maximum number of Neighbors.
Step 5 Nodes send broadcast message with unique broadcast message id and prob-
ability value.
Step 6 Neighbor nodes receive broadcast message and check message is already
received by using message id and sender id.
Or if it is already received, then simply discard the message.

[email protected]
674 N. R. Sivakumar and A. Al Garni

Step 7 If it is not received, then it compares the node probability value with broad-
casting probability.
Step 8 If the value is less than broadcasting probability, then it broadcast the mes-
sage else, discard the message.

3.2 Algorithm 2: Improved Algorithm for ROUGH Model

1. Battery Power, Speed, and Pause time are maintained with all nodes which are
inbound and outbound flow of data.
2. The Complete neighbor nodes exchange “hello” messages periodically.
3. The neighbor Table which is with node information receive the above messages
also with traffic, Speed and Pause time.
4. In neighbor table, there with one hop information from source node which stores
information in neighbor nodes.
5. Now, it initiates path discovery with RREQ to the selective nodes whenever
source node likes to transmit data packets.
6. Then, Source node will send RREQ straight without broadcasting if destination
to packet is in first hop.
7. Else it selects neighbor node from neighbor table if destination is not found and
then will transmit packet.
8. Then, finally by applying approximation with ROUGH method, it now forms
defined rules by selecting neighbor nodes.

4 Results by Simulation

Here, the Network Simulator 2 is used to investigate. The simulation was done with
45 nodes around 15 min by placing the nodes randomly over 2000 × 2000 m2 with
20 s for Pause time.

4.1 Trials on Effect to Speed

Using the following metrics, the three different algorithms are performed.

4.1.1 Average Consumed Energy (ACE) Versus Speed

ACE is the energy exists within nodes. In Fig. 1, ACE is less in ROUGH method
with respect to Speed. ACE in FLOODING method is high than APBMAN method
and ROUGH by Speed.

[email protected]
Algorithm Study and Simulation Analysis by Different … 675

Fig. 1 Speed versus average energy

Fig. 2 Speed versus total energy

[email protected]
676 N. R. Sivakumar and A. Al Garni

Fig. 3 Speed versus packet delivery ratio

Fig. 4 Speed versus throughput

[email protected]
Algorithm Study and Simulation Analysis by Different … 677

Fig. 5 Speed versus delay

4.1.2 Total Consumed Energy (TCE) Versus Speed

TCE is the energy consumed through the whole network. Figure 2 elucidates that
the energy consumed is less in ROUGH method than APBMAN and FLOODING
method by Speed.

4.1.3 Packet Delivery Ratio Versus Speed

Packet Delivery Ratio has substantial decrease in APBMAN and FLOODING

method as Speed increases. But Fig. 3 elucidates that packet delivery is more effec-
tive to certain period in ROUGH method as between 80 and 90% than APBMAN and
FLOODING method. Figure 3 shows that a Packet Delivery Ratio result is varied
for different mean Node Speeds.

4.1.4 Throughput Versus Speed

This can be said as the average rate of effective delivery of messages on the whole
network by all channels. It is shown in Fig. 4 with respect to Speed. Figure 4 shows

[email protected]
678 N. R. Sivakumar and A. Al Garni

Fig. 6 Pause time versus average consumed energy

that a Throughput result is varied for different mean Node Speeds as proportionally
equal to Packet Delivery Ratio in Fig. 3.

4.1.5 Delay Versus Speed

Delay is well decreased in FLOODING in Fig. 5. But it has increased in ROUGH

(WRS Model) and APBMAN technique. For each simulation, the Delay has been
varied from 1, 2 to 5 m/s.

4.2 Trails on Effect to Pause Time

Using the following metrics, the three different algorithms are performed.

[email protected]
Algorithm Study and Simulation Analysis by Different … 679

Fig. 7 Pause time versus total consumed energy

4.2.1 Average Consumed Energy (ACE) Versus Pause Time

In Fig. 6, the energy consumed is less in ROUGH method by Pause time. ACE in
ROUGH method is reduced than APBMAN and FLOODING method by 20 s Pause
time in simulation.

4.2.2 Total Consumed Energy (TCE) Versus Pause Time

TCE is the energy consumed through the whole network. Figure 7 elucidates that
the energy consumed is less in ROUGH method than APBMAN and FLOODING
method by Pause time.

4.2.3 Packet Delivery Ratio Versus Pause Time

Figure 8 elucidates that packet delivery is more effective in ROUGH method by

giving 90% efficiency than APBMAN and FLOODING method.

[email protected]
680 N. R. Sivakumar and A. Al Garni

Fig. 8 Pause time versus packet delivery ratio

4.2.4 Throughput Versus Pause Time

Figure 9 elucidates that Throughput with ROUGH method transfers data packets are
greater rate than APBMAN and FLOODING method to Pause time.

4.2.5 Overhead Versus Pause Time

Efficient result is produced in ROUGH method than other techniques which have
produced the efficient performance with Overhead in Fig. 10. The whole efficacy of
MANET routing scheme is determined with the network overhead.

4.2.6 Normalized Overhead Versus Pause Time

Figure 11 clearly elucidates that ROUGH method have been moderately reduced
than APBMAN and FLOODING method by Pause time.

[email protected]
Algorithm Study and Simulation Analysis by Different … 681

Fig. 9 Pause time versus throughput

Fig. 10 Pause time versus overhead

[email protected]
682 N. R. Sivakumar and A. Al Garni

Fig. 11 Pause time versus normalized overhead

5 Conclusion

This paper explains the three different algorithms which are compared and discussed
to reduce flooding with different metrics. ROUGH method, here, reduces unwanted
route request packets as it has an efficient path through the entire network through
the determination from above results. The virtual study between APBMAN method,
FLOODING method, and ROUGH method was considered with FSR Protocol in
Grid. Efficient growth in ROUGH method with effect to Pause time has been deter-
mined using MANET outcomes. Certainly, ROUGH method is good with effect to
Speed in Average consumed energy, Total Consumed energy, Packet Delivery Ratio,
and Throughput. But there is a certain increase in Delay as there is also an increase in
speed. Delay is well decreased in Propagating Neighborhood (FLOODING) method.

[email protected]
Algorithm Study and Simulation Analysis by Different … 683

References

1. Pei, G., Gerla, M., & Chen, T. W. (2000). Fisheye state routing: a routing scheme for ad hoc
wireless networks. In Proceedings of the IEEE International Conference on Communications,
(ICC’00) (Vol. 1, pp. 70–74). New Orleans.
2. Chu,T. H., & Hwang, S. I. (2006). Efficient fisheye state routing protocol using virtual grid in
high density adhoc networks. In Proceedings of the 8th International Conference on Advanced
Communication Technology (Vol. 3, pp. 1475–1478).
3. Nithya Rekha, S., & Chandrasekar, C. (2012). A strategy to reduce flooding in grid fisheye state
routing (GFSR) protocol with weighted rough set method in MANET. International Journal
of Mobile Network Design & Innovation—Inderscience Journal, 4(4), 192–200.
4. Tseng, Y. C., & Yao, N. S. (2002). The broadcast storm problem in a mobile ad hoc network.
In Wireless Networks (Vol. 8, pp. 153–167).
5. Colagrosso, M. D. (2007). Intelligent broadcasting in mobile ad hoc networks: Three classes
of adaptive protocols. EURASIP Journal on Wireless Communications and Networking.
6. Nithya Rekha, S., & Chandrasekar, C. (2012). Performance analysis of probabilistic rebroad-
casting in grid FSR for MANET. International Journal of Computer Science Issues (IJCSI),
9(2), 293–300 (Indexed by DBLP & Elsevier, EBSCO, Scirus) International Impact factor:
0.242.
7. Nithya Rekha, S., & Chandrasekar, C. (2012). An improved approach in flooding with packet
reachability in FSR (fisheye state routing) protocol using MANET. Journal of Theoretical
and Applied Information Technology (JATIT), 40(1), 98–104. (E-ISSN 1817-3195/ISSN 1992-
8645) (Indexed by Scopus).
8. Nithya Rekha, S., & Chandrasekar, C. (2012). A comparative analysis of probabilistic broad-
casting to reduce flooding with FSR (Fisheye State Routing) protocol and grid FSR using
MANET. In Proceedings of the American Institute of Physics (AIP)—The Sixth Global Con-
ference on Power, Control and Optimization. Las Vegas, USA (Indexed by Thomson Reuters
and SCOPUS, and promoted by EBSCO, Springer09 and Springer10).
9. Nithya Rekha, S., & Chandrasekar, C. (2014). An energy efficient routing to reduce flooding in
weighted rough set method using MANET. International Journal of Communication Networks
and Distributed Systems—Inderscience Journal 13(1), 83–105.
10. Rodoplu, V., & Meng, T. H. (1999). Minimum energy mobile wireless networks. IEEE Journal
on Selected Areas in Communications, 17(8), 1333–1344.

[email protected]
Cloud-Based Agricultural Framework
for Soil Classification and Crop Yield
Prediction as a Service

K. Aditya Shastry and H. A. Sanjay

Abstract Agriculture is one of the important occupations in India. Digitization in

the field of Indian agriculture is in the initial stage. Indian farmers are suffering from
various issues such as ignorance about soil parameters and inability to predict the
yield of crops. Also, various agriculture-related information from the government
agencies is not communicated to the farmers. To address the above-said issues, we
have built a cloud-based agricultural framework which enables the Indian farmers,
agricultural departments, and agro industries to extract useful agricultural informa-
tion. The designed agricultural cloud framework is providing two services, i.e., soil
classification as a service and crop yield prediction as a service. For soil classifi-
cation, hybrid support vector machine (M-SVM) and for wheat yield prediction,
customized artificial neural network (M-ANN) was developed. To store the agricul-
tural data, we are using Amazon S3 and for deployment of the services, we have
used Heroku cloud. The performance improvements in the range of 2–43%, 4–35%,
and 1–11% were observed for M-SVM with respect to k-Nearest Neighbor (k-NN),
Naïve Bayes (NB), and standard SVM classifiers, respectively. M-ANN performed
with an improvement of 2% over standard artificial neural network (ANN) and 5%
over multiple linear regression (MLR) models. We also observed that our agricultural
cloud framework is able to provide reliable and accurate agricultural services.

Keywords Agriculture · Classification · Prediction · Cloud · Framework

1 Introduction

Agriculture is the primary occupation in India. Yet, in many developing countrie,s

digitization in agriculture is still evolving [1]. Cloud computing services provide
a collective pool of resources, computing power, anywhere at anytime [2]. Using
cloud computing in agriculture has several benefits [3]. First, agricultural data will
be available at anytime from any location. Second, agriculture functions such as crop

K. Aditya Shastry (B) · H. A. Sanjay

Department of ISE, NMIT, Bangalore 560064, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2019 685

N. R. Shetty et al. (eds.), Emerging Research in Computing, Information, Communication
and Applications, Advances in Intelligent Systems and Computing 882,
https://doi.org/10.1007/978-981-13-5953-8_56

[email protected]
686 K. Aditya Shastry and H. A. Sanjay

yields may be provided to stakeholders. Further, data security is also provided [3].
Data mining is another important area which involves extracting useful information
hidden in data [4]. Incorporating data mining analytics in agriculture is beneficial
for providing improved crop yields through crop yield prediction [5]. Keeping these
points in mind, we have developed a cloud-based agricultural framework for pro-
viding soil classification as a service and crop yield prediction as a service. The end
users of the framework will be agro industries, government organizations, farmers
(landlords), or any other agricultural-related agencies.
Proper soil classification is very much required for crop growth. Classification of
soils based on its characteristics is critical to assess the soil quality [6]. Soil classifi-
cation is performed based on its chemical properties and its texture [7]. Proper soil
classification based on chemical properties aids farmers to choose the magnitude of
fertilizer and farmyard manure to be applied at various stages of the development
cycle of the crop [8]. The soil texture is another important property for agriculture soil
classification, which impacts the soil fertility, ability of soil to hold water, exposure
of soil to air, tillage of soil, and soil strength [9]. In India, traditional soil classi-
fication involving tables, flowcharts are followed. They make use of the standard
statistical analysis techniques which consume a lot of time and cost more. In view of
this, we have developed an efficient soil classifier based on M-SVM for performing
soil classification centered on chemical properties and texture [10]. The end users of
soil classifier will be the soil experts who may enter values or read a file. In case of
chemical properties, the parameters considered for soil classification are power of
hydrogen (pH), electrical conductivity (EC), organic carbon (OC), available phos-
phorous (AvP), available Boron (AvB), and available potassium (AvK). This data
was obtained from ICRISAT [11]. Sand, silt, and clay are considered as parameters
for classifying soils based on texture with data obtained from NBSS [12]. This input
soil data will then be classified by our proposed soil classifier using M-SVM based
on texture and chemical properties of soil.
Crop yield prediction plays a very significant role in agriculture. Accurate predic-
tion of crop yields leads to better harvests [13]. Currently, the primary method of crop
yield prediction in Karnataka state is performed using regression models which are
not able to capture nonlinear relationships between attributes [14]. Artificial neural
network (ANN) is a significantly better model than regression models for performing
predictions on continuous data [15]. In this regard, we have developed an M-ANN
model in which number of hidden layers; number of neurons in hidden layer and
learning rate (LR) are varied [16]. Parameters considered for crop yield prediction
are weather parameters such as rain, soil parameters such as soil evaporation, tran-
spiration, and extractable soil water (ESW) and fertilizer parameters such as nitrogen
(N). The end users of crop yield prediction software will be agro industries who will
have various data available on which prediction can be done. They can pass on this
expert information to the farmers. The crop yield prediction has been performed on
real-world datasets obtained from [10, 11].
Some of the benefits of deploying the application to the cloud include easy scaling
of application resources, cheaper storage, easy disaster recovery, all time availability
of application to end users at all times, etc. [17]. To make soil classification and

[email protected]
Cloud-Based Agricultural Framework for Soil Classification … 687

crop yield prediction available to end users at all times, the soil classifier and crop
yield predictor software’s have been deployed on Heroku cloud. The datasets of
soil, weather, and fertilizer have been stored in Amazon S3. The main advantages of
deploying to the cloud are that, the software’s are available to all users at anytime,
and also secure with proper backup.
The remainder of the paper is structured as follows; Sect. 2 discusses some of
the recent and significant works done in the development of cloud frameworks for
agriculture. Sections 3 and 4 present the soil classifier and crop yield predictor.
Section 5 describes the deployment of the soil classifier and crop yield predictor
applications on Heroku cloud. Section 6 describes the results obtained from soil
classifier, crop yield predictor and cloud framework followed by conclusion and
references.

2 Related Work

In this section, we review some of the significant works done in the development of
cloud-based agricultural frameworks. In [18], the authors have built a cloud-based
decision support system for orchards. The system provides cloud-based automation
from sensor inputs. No data mining has been incorporated. In [19], the authors pro-
pose a cloud-based data analytics framework for crop management in India. The
authors plan to use machine learning library. It is just a proposal and has not been
implemented. Further, they do not discuss in detail which data mining techniques to
use and what type of services are provided to end users. In [20], the authors have
implemented Internet of Things (IoT) architecture for precision agriculture. Sen-
sors collect data, for reading air temperature, humidity, moisture, leaf wetness, wind
speed, and rain volume. The authors have shown how IoT is used in agriculture for
sensing data without doing any analysis. In [21], the authors have developed an agri-
cultural autonomic system in cloud which provides information on fertilizer, crop,
pesticide weather, equipment, etc. No analysis is incorporated in cloud. In [22], the
authors developed a cloud-based system to manage cultivation. It is record man-
agement software without any intelligent analysis involved. In most of the works,
sensors collect agricultural data and reproduce the same data without any intelligent
analysis. That is, no data mining is incorporated in cloud framework for agriculture.
Also, most of the cloud agricultural frameworks are proposals without implemen-
tations. As per our survey, no cloud framework with data mining algorithms exists
in Indian agricultural scenario to analyze the data. In our work, relevant agricultural
services such as soil classification and crop yield prediction are provided through
data mining algorithms in cloud environment. Hence, our work has combined the
benefits of two powerful technologies, i.e., data mining and cloud computing.

[email protected]
688 K. Aditya Shastry and H. A. Sanjay

3 Cloud-Based Agricultural Framework

Here, we have built soil classification and crop yield prediction systems in the cloud
environment for providing soil classification and crop yield prediction as a service.
For this purpose, we made use of open-source platform called Heroku cloud platform
for deploying cloud applications. The entire coding of soil classification and crop
yield prediction was done using MATLAB. This MATLAB code was converted
into Java code using MATLAB coder. This conversion was done, as most cloud
deployment services are Java friendly. The soil classification application is useful
for geological engineers for soil survey as well as agriculturists who are interested
in soil types. The crop yield prediction application can be used directly by agro
industries to give valuable information to farmers who may be illiterate and not be
able to use the software directly. Figure 1 depicts the deployment process followed
for deploying soil classification and crop yield prediction on Heroku cloud which
contains three phases.

AMAZON S3

Fertilizer
Soil data Crop data Weather data

Crop yield predictor

Soil classifier

Build phase

HEROKU CLOUD
Build pack

Execution phase

Slug

Release phase

Soil classification and Crop yield prediction as a service

Fig. 1 Agricultural cloud framework

[email protected]
Cloud-Based Agricultural Framework for Soil Classification … 689

In the first phase, Amazon S3 was used to store the soil data, crop data, weather
data, and fertilizer data in the form of buckets. Soil samples containing data about
pH, EC, OC, AvP, AvB, AvP, sand, silt, and clay were collected from NBSS [12] and
ICRISAT [11]. Crop data contained the data about area sown, production in tones,
and historic crop yields of wheat crop from various districts of Karnataka. These
were collected from IndiaStat [23]. The weather data contained rain, temperature,
and precipitation data collected from India Water Portal [24]. The fertilizer data
contained data about nitrogen, phosphorous, and potassium collected from Indiastat
[23].
In the second phase, soil classifier and crop yield predictors were developed.
The soil classifier user interface contained text fields for the user to enter pH, EC,
OC, AvP, AvB, AvP, sand, silt, and clay values. It also contained import file option
in which user imports the soil data file in excel or text format. In case of the soil
classifier, M-SVM developed in our earlier work [10] was used. It is briefly described
in Sect. 4.
In case of crop yield predictor, the user interface contained text fields for user to
enter fertilizer information (nitrogen, phosphorous, and potassium), weather infor-
mation (rain, temperature, and precipitation), soil information (pH, EC, OC, AvP,
AvZn, AvB, AvK, and AvS), and crop information (area sown and crop production).
Before entering the values, the user has to choose the type of crop. Currently, the
system has been implemented to predict crop yields of wheat crop for Karnataka
state, India. The front end contains an import file option where the user chooses text
file containing fertilizer, soil, and crop information. The prediction was done using
the hybrid artificial neural network (M-ANN) developed in our earlier work [16]. It
is briefly discussed in Sect. 5.
In the third phase, the actual deployment to Heroku cloud is done. In the build
phase, the applications (soil classifier and crop yield predictor) containing source
code, dependencies were built using build pack. In the execution phase, a slug is
created. Slugs represent compact replicas of applications enhanced for delivery to the
dyno manager [25]. This slug was run-on dynos which are isolated Unix containers
that offer an environment to run the app [25]. The release phase binds the config,
slug, and add-ons.

4 Soil Classifier (M-SVM)

In our earlier work [10], we had developed a hybrid kernel SVM classifier which
is briefly described here. The agricultural data is first divided into training and test
sets. The tuning of SVM parameters is done for the training set. Here, the kernel
parameters C and σ are first selected using genetic algorithm (GA). In this work,
using linear, quadratic, RBF, MLP, and polynomial kernels, ten hybrid kernels were
developed. The hybrid SVM classifier is then run-on the test set samples. The model
is then evaluated using the performance metrics Accuracy, Sensitivity, Specificity,
Precision, and F-Score. Gradient descent method is used to select the SVM param-

[email protected]
690 K. Aditya Shastry and H. A. Sanjay

eters, if performance is poor. Each kernel is then applied for classification. Superior
performance was achieved on the real world and benchmark datasets for the hybrid
kernel QRK (Quadratic and RBF kernel) for GA selected SVM parameters. Algo-
rithm 1 describes the hybrid kernel SVM for multiclass agricultural datasets.

Algorithm-1 Hybrid Kernel SVM

1: procedure M-SVM(TrngSet TstSet)

2: [TrngInput,TrngCls]=SplitData(TrngSet);
3: [TstInput,TstCls]=SplitData(TstSet);
4: Compute[C1,γ1]=GD(TrngInput,TrngCls);
5: Compute[C2, γ2]=GA(TrngInput,TrngCls);
6: for k←1 to numClasses do
7: H- Kernel=[Linear+(Quad/Poly/RBF/MLP)|Quad+(Poly/RBF/MLP)|Poly
+(RBF/MLP)|RBF+MLP];
8: models(k) TrainSVM(TrngSet,C1|C2,γ1|γ2,H-Kernel);
9: end for
10: for j←1 to size(TstSet) do
11: for k←1 to numClasses do
12: BinSVMClassify(models(k),TstSet)
13: end for
14: end for
15: [Accuracy,Sensitivity,Specificity,Precision,FS] ←Predict(TstCls,Pred-Class)
16: end procedure

The performance of ensemble classifiers is found to be comparatively better than

single classifiers in many instances. Similarly, it is found that multiple kernels are
more useful when compared to single kernels. This concept called multiple kernel
learning (MKL) has two advantages: First, single kernels suffer from bias; while mul-
tiple kernels associate the advantages of different kernels. Second, different kernels
utilize inputs from different representations having various similarity measures [26].
Data from multiple sources are collated when kernels are combined [26]. Structures
which are not apprehended by single kernels are captured by hybrid kernels [26]. In
this work, hybrid combinations of SVM kernels are developed and run-on actual and
standard agricultural datasets. Experiments indicated that the QRK kernel exhibited
superior performance than other hybrid kernels. This kernel is given by Eq. (1)
 2  
QRK(a, b)  αa T b + m + exp −γ a − b2 (1)

where a, b are inner products and m is an optional constant, T is the transpose.

QRK kernel exploits the benefits of quadratic and RBF kernels which are the
extraction of global and local features, respectively.

[email protected]
Cloud-Based Agricultural Framework for Soil Classification … 691

5 Crop Yield Predictor

In our earlier work [16], we had developed a customized artificial neural network
model for wheat yield prediction which is described here. M-ANN is developed
by changing N h (amount of hidden layers from 1 to 2), N n (quantity of neurons
in every hidden layer from 10 to 1000), and learning rate (LR). Best performance
was achieved for 2 hidden layers comprising of 60 neurons in the first layer and 30
neurons in the second layer with LR  0.20. The transfer functions used in hidden and
output layers were logsig and purelin, respectively. Wheat dataset containing 1674
records was obtained from [23]. After the elimination of redundant, inconsistent and
missing values, the dataset contained 1645 records. M-ANN model was trained on
70% of the total records (1151 records). 15% of total records (246 records) were
considered as validation set in order to generalize the M-ANN model. Remaining
15% of records (246 records) were utilized as test set to ascertain the accuracy of
M-ANN model in terms of R2 and PPE. M-ANN model achieved a high R2 and low
PPE values when compared to MLR and ANN models which are described in detail
in the results section. The Algorithm-2 describes the M-ANN for predicting wheat
yield [16].

Algorithm-2 Customized-ANN

1: procedure M-ANN (TrngSet TstSet)

2: [TrngInput,TrngCls]=SplitData(TrngSet);
3: [TstInput,TstCls]=SplitData(TstSet);
4: Use Levenberg Marquardt algorithm for training;
5: Use logsig transfer function for hidden layers and purelin transfer function for
the output layer;
6: Customize feed forward back-propagation network by varying the following
parameters
i) Number of hidden layers (1 to 2)
ii) number of neurons in hidden layers (20 to 100)
iii) learning rates (0.25, 0.5)
iv) network weights (random)
7: [Accuracy, Percentage Prediction Error] ←Perf(TestClass,Pred-Class)
8: end procedure

6 Experimental Setup and Results

Experimental setup for M-SVM classifier comprised of Windows 7 64-bit operating

system possessing 6 GB RAM, with 500 GB hard disk. M-SVM was programmed
in MATLAB R2016a. Several trials were performed on four standard agricultural

[email protected]
692 K. Aditya Shastry and H. A. Sanjay

Table 1 Experimental results of M-SVM soil classifier

Data Classifier Accuracy Sensitivity Specificity Precision F-Score
Benchmark datasets
Urban land cover k-NN 35.7 53.93 31.82 14.41 22.75
NB 62.52 84.27 56.94% 29.41 43.6
SVM 68.05 86.33 66.99 32.02 44.52
M-SVM 79.09 88.76 77.03 45.14 59.85
Real world datasets
Soil texture NB 51.56 81.25 41.67 31.71 45.61
k-NN 52.63 85.66 45.23 33.33 47.21
SVM 59.38 93.75 47.92 37.5 53.57
M-SVM 68.75 95.75 60.42 44.12 60
pH based soil k-NN 58.49 91.3 26.67 51.11 67.65
classification NB 81.13 100 73.3 72.4 80.77
SVM 96.23 100 93.33 92 95.83
M-SVM 98.11 100 96.67 95.83 95.87
OC based soil k-NN 75.47 92.31 59.26 68.57 78.69
classification NB 81.13 92.31 66.67 73.53 83.33
SVM 94.34 96.15 96.3 96 94.12
M-SVM 96.23 96.15 96.3 96.15 96.15
AvB based soil k-NN 73.58 75.86 70.83 75.86 75.86
classification NB 90.57 86.21 95.83 96.15 90.91
SVM 94.34 93.1 95.83 96.43 94.74
M-SVM 96.23 96.5 95.83 96.55 96.55

datasets from UCI [27] and four real-world datasets from NBSS [12] and ICRISAT
[11]. M-SVM classifier was compared with NB, k-NN and default SVM classifiers.
NB classifier was run using different distributions (Gaussian, multivariate, and multi-
nomial). The k value in k-NN classifier was varied from 1 to 1000. Experiments using
standard SVM were performed using linear, quadratic, polynomial, and RBF ker-
nels. The proposed M-SVM was run using ten hybrid kernels with C and σ values
obtained from genetic algorithm. Table 1 depicts the comparison of results obtained
from the classifiers.
As can be observed from Table 1, the proposed M-SVM performed superior
when compared to NB, SVM, and k-NN classifiers with respect to the performance
metrics. For wheat yield prediction using M-ANN, experiments were carried out
using MATLAB R2016a on Windows 7 environment, Intel® Pentium® CPU P6200
@2.13 GHz processor with 6 GB RAM and 500 GB Hard Disk. The accuracy of
the three models (MLR, D-ANN, and M-ANN) using R2 statistic and the percentage
error (PPE) was measured. Table 2 depicts the prediction results based on R2 value.

[email protected]
Cloud-Based Agricultural Framework for Soil Classification … 693

Table 2 Experimental results of M-ANN based on R2

R2 PPE
Prediction Training Validation Test set Training Validation Test set
models set (%) set (%) (%) set (%) set (%)
MLR 100 100 92.52 4.068 7.053 4.196
D-ANN 96 78 95 2.357 2.5533 2.2408
M-ANN 99 90 97 0.6629 0.3968 0.5275

Fig. 2 Soil classification as a service in Heroku cloud environment

Table 2 depicts the comparative results between MLR, D-ANN, and M-ANN based
on R2 and PPE values.
MLR model attained 100% accuracy on training and validation sets. A decrease
of 7.48% accuracy was seen on the test set. This was due to the inability of MLR
model in grasping the nonlinear association between input parameters. The decrease
in performance of D-ANN was observed on validation set as the network failed
to generalize. Superior performance was observed for M-ANN model on test set
showing improvements of 4.48% over MLR and 2% over D-ANN models. It is
observed that the average percentage prediction error for the M-ANN model is the
lowest for all the data sets. Hence, the M-ANN model is capable to forecast the yield
of wheat better than the MLR and D-ANN models.

[email protected]
694 K. Aditya Shastry and H. A. Sanjay

Fig. 3 Crop yield prediction as a service in Heroku cloud environment

With respect to deployment of the soil classification and crop yield prediction
applications on Heroku cloud, Heroku cloud infrastructure was used along with
Amazon S3. The entire code of MATLAB was converted into Vaadin Java program.
The datasets were stored in Amazon S3. Figures 2 and 3 depict the snapshots of
soil classification and crop yield prediction provided as a service in Heroku cloud
environment, respectively.

7 Conclusion

Though India is progressing in various sectors, agriculture is a sector in which tech-

nology is not being completely utilized. In this regard, we have developed an agri-
cultural cloud framework for providing soil classification and crop yield prediction
as a service to end users. End users of this framework may be agricultural agen-
cies, government organizations, and farmers (landlords). As per our survey, no cloud
service exists for providing soil classification and crop yield prediction as a service
to end users. In this regard, we have deployed the soil classification and crop yield
prediction applications to Heroku cloud. The main benefits of this include all time
available to access the applications from anywhere at anytime. Further, data loss does

[email protected]
Cloud-Based Agricultural Framework for Soil Classification … 695

not happen as data gets stored in cloud storage and is secure. The soil classification
system can be used by soil experts, while the crop yield prediction system can be
used by agro industries, which in turn can provide valuable suggestions to farmers.
The soil classifier (M-SVM) and crop yield predictor (M-ANN) developed from our
earlier works [10, 16] were deployed on Heroku cloud. The agricultural framework
was able to provide reliable and accurate services to end users. In future, we plan to
develop mobile agricultural apps with sophisticated functionalities.

References

1. Deichmann, U., Goyal, A., & Mishra, D. (2016). Will digital technologies transform agricul-
ture in developing countries? Agricultural Economics, 47, 21–33. https://doi.org/10.1111/agec.
12300.
2. Shawish, A., & Salama, M. (2014). Cloud computing: Paradigms and technologies. In Inter-
cooperative collective intelligence: Techniques and applications (Studies in Computational
Intelligence, Vol. 495, pp. 39–67). Berlin Heidelberg: Springer-Verlag. https://doi.org/10.1007/
978-3-642-35016-0_2.
3. Mekala, M. S., & Viswanathan, P. (2017). A survey: Smart agriculture IoT with cloud comput-
ing. In 2017 IEEE International Conference on Microelectronic Devices, Circuits and Systems
(ICMDCS) (pp. 1–7). Vellore. https://doi.org/10.1109/icmdcs.2017.8211551.
4. Milovic, B., & Radojevic, V. (2015). Application of data mining in agriculture. Bulgarian
Journal of Agricultural Science, 21, 26–34.
5. Mucherino, A., Papajorgji, P., & Pardalos, P. (2009). A survey of data mining techniques
applied to agriculture. Operational Research, 9(2), 121–140. https://doi.org/10.1007/s12351-
009-0054-6. Springer-Verlag.
6. Kibblewhite, M. G., Ritz, K., & Swift, M. J. (2008). Soil health in agricultural systems. Philo-
sophical Transactions of the Royal Society B: Biological Sciences, 363, 685–701. https://doi.
org/10.1098/rstb.2007.2178.
7. Mattsson, Berit, Cederberg, Christel, & Blix, Lisa. (2000). Agricultural land use in life cycle
assessment (LCA): Case studies of three vegetable oil crops. Journal of Cleaner Production,
8, 283–292. https://doi.org/10.1016/S0959-6526(00)00027-5.
8. Karthik, D., Vijayarekha, K., & Manickkam, V. (2014). Land characterizations based on soil
properties using clustering techniques. World Applied Sciences Journal (Data Mining and Soft
Computing Techniques), 29, 60–64. https://doi.org/10.5829/idosi.wasj.2014.29.dmsct.11.
9. Hristov, Biser. (2013). Importance of soil texture in soil classification systems. Journal of
Blakan Ecology, 16, 137–139.
10. Shastry, K. A., Sanjay, H., & Deexith, G. (2017). Quadratic-radial-basis-function-kernel for
classifying multi-class agricultural datasets with continuous attributes. Applied Soft Computing,
58, 65–74. https://doi.org/10.1016/j.asoc.2017.04.049.
11. Wani, S. P., Sahrawat, K. L., Sarvesh, K. V., Baburao, M., & Krishnappa, K. (Eds.). (2011).
Soil Fertility Atlas for Karnataka, India (pp. 312). Patancheru 502 324, Andhra Pradesh, India:
International Crops Research Institute for the Semi-Arid Tropics. ISBN 978-92-9066-543-4.
12. National Bureau of Soil Survey and Land Use Planning, Soil Data. https://www.nbsslup.in.
13. Paul, M., Vishwakarma, S. K., Verma, A. (2015). Analysis of soil behaviour and prediction of
crop yield using data mining approach. In: 2015 IEEE International Conference on Compu-
tational Intelligence and Communication Networks (CICN) (pp. 766–771). https://doi.org/10.
1109/cicn.2015.156.
14. Ramasubramaniyam, V. (2005). Forecasting techniques in agriculture. Journal of the Indian
Society of Agricultural Statistics. IASRI, New Delhi.

[email protected]
696 K. Aditya Shastry and H. A. Sanjay

15. Stastny, J., Konecny, V., & Trenz, O. (2011). Agricultural data prediction by means of neural
network. Agricultural Economics, 57, 356–361. https://doi.org/10.17221/108/2011-agricecon.
16. Aditya Shastry, K., Sanjay, H., & Deshmukh, Abhijeeth. (2016). A parameter based customized
artificial neural network model for crop yield prediction. Journal of Artificial Intelligence, 9,
23–32. https://doi.org/10.3923/jai.2016.23.32.
17. Savu, L. (2011). Cloud computing: Deployment models, delivery models, risks and research
challenges. In International Conference on Computer and Management (CAMAN) (pp. 1–4),
Wuhan. https://doi.org/10.1109/caman.2011.5778816.
18. Tan, L. (2016). Cloud-based decision support and automation for precision agriculture in
orchards. In 5th IFAC conference on Sensing Control and Automation Technologies for Agri-
culture.
19. Balaji Prabhu, B. V., & Dakshayini, M. (2016). A novel cloud based data analytics framework
for effective crop management. IJCTA, 9(22), 257–264.
20. Khattab, A., Abdelgawad, A., & Yelmarthi, K. (2016). Design and implementation of a cloud-
based IoT scheme for precision agriculture. In 28th IEEE International Conference on Micro-
electronics (ICM) (pp. 201–204). https://doi.org/10.1109/icm.2016.7847850.
21. Gill, Sukhpal S., Chana, I., & Buyya, R. (2015). Agri-info: Cloud based autonomic system for
delivering agriculture as a service.
22. Murakami, Y., Utomo, S. K. T., Hosono, K., Umezawa, T., & Osawa, N. (2013). iFarm: Devel-
opment of cloud-based system of cultivation management for precision agriculture. In 2013
IEEE 2nd Global Conference on Consumer Electronics (GCCE) (pp. 233–234), Tokyo. https://
doi.org/10.1109/gcce.2013.6664809.
23. IndiaStat, Fertilizer and crop yield data, https://www.indiastat.com/agriculture.
24. India Water Portal, Rainfall data, http://www.indiawaterportal.org.
25. Heroku, Slugs and dynos, https://www.logicline.de/en/blog/2015/11/a-more-technical-look-
at-heroku-about-procfiles-dynos-and-the-slug/.
26. Shi, L., Duan, Q., Ma, X., & Weng, M. (2012). The research of support vector machine in
agricultural data classification (pp. 265–269). Berlin Heidelberg: Springer.
27. Lichman, M. (2013). UCI machine learning repository. http://archive.ics.uci.edu/ml.

[email protected]