2022 2nd Asian Conference on Innovation in Technology (ASIANCON)

Pune, India. Aug 26-28, 2022

Transfer Learning: Convolutional Neural Network-

AlexNet Achieving Face Recognition
Shatvik Singh Sugandha Sharma Amit Jain
Department of Computer Science & Department of Computer Science & University Institue of Computing
Engineering Engineering Chandigarh University
2022 2nd Asian Conference on Innovation in Technology (ASIANCON) | 978-1-6654-6851-0/22/$31.00 ©2022 IEEE | DOI: 10.1109/ASIANCON55314.2022.9908650

Chandigarh University Chandigarh University Mohali, Punjab

Mohali, Punjab Mohali, Punjab [email protected]
[email protected] [email protected]

Pritpal Singh Animesh Kudake

Department of Computer Science & Department of Computer Science &
Engineering Engineering
Chandigarh University Chandigarh University
Mohali, Punjab Mohali, Punjab
[email protected] [email protected]

Abstract— Nowadays, Machine-based face recognition is association to the underlying calcified and skeletal
becoming very commonplace, robust and dependable system components. Each human face is distinct owing to variation
that is widely employed in numerous cases for access control. in relative size, form, and spatial arrangement among the
As in traditional approach, face recognition needs the numerous facial features (e.g., forehead, nose, chin, eyes,
extraction of face features prior to classification and etc.). As a result of its advantages in the recognition area,
recognition, which affects recognition rate. We employ Face face structure is gaining the attention of an increasing
Verification checks whether the pictures are associated with a number of researchers. Traditional face recognition methods
single individual, whereas Face Identification must identify a can be classified into two types: holistic and local feature
specific face from a collection of known profiles in the system.
techniques for face recognition.
To tackle this question, this paper incorporates the CNN
structure Alexnet to obtain face identification. Since its inception in 2012, the convolutional neural
network approach has produced impressive success in a
Throughout this article, we perform facial recognition variety of computer vision challenges. Researchers have
using transfer learning in a Siamese network composed of 2 studied facial recognition using machine learning thoroughly.
comparable CNNs. A pair of 2 face picture is fed into the Scholars used the convolutional neural network approach to
Siamese network as input, after which the network learns the perform relevant research trials in the area of face
traits of this pair of pictures.
recognition, with positive findings. This paper employs the
Next the network is trained using the PRelu activation Alexnet framework., and tests and enhancements are done on
function to find the ideal learning algorithm and maximal the spectral resolution and 2D face picture datasets. The
values. Then, the face was identified and categorized. Library reliability of each spectrum is really high, demonstrating the
Multi-Spectral Face Data - set and Library 2D Faceprint deep learning technique's effectiveness in face recognition.
Database were used to test the methodology, it enhances the The remainder of the paper is laid out as follows: Section II
accuracy of face recognition when compared to algorithms explains the facial preprocessing approach and the deep
trained on datasets with a particular dataset and a specific learning-based facial recognition approach; Section III
spectrum's recognition rate peaked up to 98 percent. describes the experimental findings acquired by validating
the technique on a facial input database, section IV present
Keywords— Deep Learning, Face Recognition, experimental values and Section V wraps up the study.
Convolutional Neural Network, Alexnet, Image procession LFW
(Labeled face in the wild) II. METHODOLOGY
I. INTRODUCTION A. Image preprocessing
Facial recognition technology, a sort of machine learning, Face picture with rich texture and structural
has become increasingly popular in a wide range of characteristics such as nose, lips, eyes, brow, forehead, and
applications, including security agencies. Facial recognition detail points. Images of the same photograph gathered at
is a method of detecting or verifying a person's identification different periods will have various levels of rotation and
through using their image. People can be identified using translation, and the size of the face captured at the same
facial recognition algorithms in pictures, films, or in real - moment may also differ, as a result, prior extracting features
time basis. There are two types of face recognition: face and recognizing faces, it is required to extract the effective
validation and face recognition. Face verification is a 1:1 Return on Investment section of a face containing primary
matching that simply determines whether two pictures features. Figure 1 depicts the entire processing procedure.
belong to the same person or not. Face identity 1: X The extraction of Return on Investment is a critical stage,
matching, on either hand, requires determining who this and the accurate Return on Investment extraction is favorable
person is in the image from all potential outcomes. to picture alignment, boosting the efficiency of feature
matching, and ultimately having a good impact on
The facial surface, in comparison to certain other human recognition outcomes.
physiology, is easily visible and recognizable, with a close

978-1-6654-6851-0/22/$31.00 ©2022 IEEE 1

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on June 13,2024 at 13:39:09 UTC from IEEE Xplore. Restrictions apply.
 720*576 180*180 227*227
Fig. 1.

The AlexNet rollout begins here. AlexNet requires a images. This paper includes a generated the database using
huge database with at least hundreds different faces. This is the (MATLAB) toolbox. All of the photographs gathered
why a fresh database is created rather than using a database were stored in a subdirectory and then loaded. A method was
available on the internet and after creating database, all used to accomplish face detection and scaling on input
photos should be 227x227 in size, as illustrated in Figure 1. photos. The outcomes of the uploaded photographs are then
We began by gathering images. Then, using MATLAB's preserved. A class is a collection of photos of a specific
Image Batch Processor tool, I obtained the necessary facial individual which is shown in FIGURE 2.

Fig. 2. Dataset

B. Convolutional Neural Network- AlexNet before immediately outputting the results via the entire
A CNN is a Deep Learning algorithm that can taking in a network. It is divided into 4 sections: Convolution layer,
file as input, allocate significance (learnable support vectors) pooling layer, connection layer and output layer.
to diverse components inside the picture, and distinguish one In table 1 and figure 3, this article describes the system
from another. Convolutional Neural Network continually structure of the eight neural network layers, five Conv., three
applies the convolutional operations to the identical image to pool., and three fc.
produce progressively complicated characteristic maps

Fig. 3.

2
Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on June 13,2024 at 13:39:09 UTC from IEEE Xplore. Restrictions apply.
 TABLE I. ALEXNET ARCHITECTURE

Author [9] introduced a new nonlinear correcting input

feature PReLU, which is defined in equation. PReLU
converges more quickly than ReLU. And, despite the
addition of the variable "n," PReLU has little to fear from
generalization. That shown in fig. 4.

Fig. 4. PRelu and Relu

III. EXPERIMENT
Fig. 5. Small-sample dataset
A. Small- sample Dataset
This research employs a 2D database in the form of a 2D IV. EXPERIMENTAL RESULTS
image of a face. Because the range of various energies, skin
absorption and reflectance are different, you can collect We chose Database in the facial photos Database for the
numerous ranges of face images, and these pictures are not experiment in order to check and pick the right learning rate.
easily fabricated, face may gain more abundant Facial We choose all 13 data classes inside the database. After
characteristics and a greater recognition result. As a result, preparing the data, we separated it into 50 percent train, 25
we turned to the small-sample of face Image Dataset that percent test, and 25 percent validation sets to process
shown in fig.5. The Facial Image Dataset includes 2000 efficiency.
example photos taken from 12 distinct faces. Each face was TABLE II. TRAINING DATA
sampled in a different session, with 150 samples gathered in
each trial. In the study, samples obtained from the same face Base
can be classified as belonging to a certain class, whereas Epoch Iteration Time Elapsed Mini- Mini- Learning
samples taken from various faces can be classified as (hh: mm: ss) Batch Batch Rate
belonging to distinct classes. As a result, this sample base
has a total of 13 classes, each with 150 samples. This report Accuracy Loss
makes use of a segment from the dataset, and the image size 1 1 00:00:04 20.31% 3.1689 0.0010
is 227*227. 6 50 00:02:41 100.00% 0.0292 0.0010
B. Experimental Environment 12 100 00:05:12 98.44% 0.0284 0.0010
This study's training equipment involves an eight-core 17 150 00:05:12 98.44% 0.0461 0.0010
Intel i5 processor, an NVIDIA GEFORCE RTX 3090 20 180 00:09:08 100.00% 3.0724e- 0.0010
graphics card, and 8GB of memory. This article's test was
05
carried out in the MATLAB deep learning framework, and
the programming language is MATLAB.

3
Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on June 13,2024 at 13:39:09 UTC from IEEE Xplore. Restrictions apply.
 REFERENCES
[1] Palmprint Recognition Based on Convolutional Neural Network-
Alexnet Weiyong Gong, Xinman Zhang, Bohua Deng School of
Electronics and Information Engineering MOE Key Lab for
Intelligent Networks and Network Security Xi’an Jiaotong University
No.28 xianning west road Xi’an, China
[2] Recent Advances in Deep Learning Techniques for Face Recognition
MD. TAHMID HASAN FUAD1 , AWAL AHMED FIME1 ,
Fig. 6. Face detection DELOWAR SIKDER1 , MD. AKIL RAIHAN IFTEE1 , JAKARIA
RABBI 1 , MABROOK S. AL-RAKHAMI 2 (MEMBER, IEEE),
Figure 6 depicts the outcome of HaarCascade face ABDU GUMAE2 , OVISHAKE SEN1 , MOHTASIM FUAD1 ,
AND MD. NAZRUL ISLAM1
detection.
[3] Recognition of face photos on the basis of sketch using deep CNN
and transfer learning Mayuri B. Bahadure, Medha A. Shah
Department of computer science Walchand college of Engineering
Sangli,India
[4] Transfer Learning for Face Identification with Deep Face Model*
Huapeng Yu, Zhenghua Luo College of Information Science and
Engineering University of Chengdu Chengdu, China
[5] Transfer Learning of a Neural Network Using Deep Suleman Khan,
Ehtasham Ahmed, M. Hammad Javed, Syed A A Shah, Syed Umaid
Ali Dept. of Electrical Engineering Bahria University Islamabad
Campus Islamabad, Pakistan
[6] Multiple-Instance Hidden Markov Model for GPR-Based Landmine
Detection Achut Manandhar, Peter A. Torrione, Leslie M. Collins,
Senior Member, IEEE, and Kenneth D. Morton, Member, IEEE
[7] Facial Expression Recognition Based on Local Features of Transfer
Learning Haiqiang Feng, Jingfeng Shao School of Management,
Xi’an Polytechnic University
Fig. 7. Face recognition
[8] Using Siamese Networks with Transfer Learning for Face
Recognition on Small-Samples Datasets Mohsen Heidari, Kazim
Figure 7 Displays the outcome of face recognition using Fouladi-Ghaleh Deep Learning Research Lab Department of
a newly trained model. Computer Engineering Faculty of Engineering, College of Farabi
University of Tehran, Iran
This data is obtained utilizing application MATLAB.
[9] Krizhevsky A, Sutskever I,Hinton G E,“Imagenet classification with
Four classes were trained for recognizing. Every class has a deep convolutional neural networks,” Advances in neural information
total of 150 photos. The output result is in the form of a class processing systems, pp.1097-1105, January 2012.
label and a forecast result. Accuracy rate can reach up to 98 doi:10.1145/306538.
percent, but it can be improved by employing a multitude of [10] J. Seo and H. Park, ‘‘Object recognition in very low resolution images
alternative photos. using deep collaborative learning,’’ IEEE Access, vol. 7, pp. 134071–
134082, 2019.
V. CONCLUSION [11] YOLOv5. Accessed: Oct. 2, 2021. [Online]. Available:
https://github.com/ultralytics/yolov5
The methodology was tested using the Analyzed Face
[12] Rounak Raj, Sandhya Balakrishna, Shruthi M.L.J, Deeksha H Kulal,
Data - set which improves the accuracy of face recognition Sunny Singh, "A survey on machine learning-based facial recognition
when compared to algorithms trained on datasets with a algorithm", 2022 International Conference on Smart Technologies
specific dataset and a specific spectrum, with a recognition and Systems for Next Generation Computing (ICSTSN), pp.1-7,
rate of up to 98 percent. To get face identification, the paper 2022.
employs the CNN architecture Alexnet. Researchers [13] Li Kai, Wang Tianjun, Xiao Jingfeng, "Research on image
thoroughly investigated facial recognition using machine Recognition Technology Based on Artificial Intelligence", 2021 2nd
International Conference on Computer Communication and Network
learning. The Alexnet framework and tests are used in this Security (CCNS), pp.5-9, 2021
study. [14] Yuan Cao, Yongmao Wang, "Reliability analysis of computer
AlexNet in Mathematica requires a massive database adaptive system based on artificial intelligence recognition
model", Journal of Ambient Intelligence and Humanized Computing,
with at least hundreds of different faces. This article includes 2022.
a database created with the (MATLAB) toolkit. [15] Rachana G Kummar, Suhas J Shetty, S N Vishwas, P J Vismith
We constructed a database in the face database to test and Upadhya, Jyoti R Munavalli, "Edu-bot: An AI based Smart Chatbot
for Knowledge Management System", 2021 IEEE International
determine the right learning rate. Following data Conference on Computation System and Information Technology for
preprocessing, we separated the data into a few parts for Sustainable Solutions (CSITSS), pp.1-6, 2021.
verification of sets. With a max of 180 iterations, we [16] A. W. Senior and R. M. Bolle, "Face recognition and its application"
investigated the effect of different learning rates on in Biometric solutions, Springer, pp. 83-97, 2002.
recognition rate. When the recognition accuracy was 0.0010, [17] K. Simonyan and A. Zisserman, Very deep convolutional networks
this was the highest rate. For training and testing purposes, for large-scale image recognition, 2014.
traditional algorithms such as CNN are utilised. [18] Harsh Jindal , "Analytical approach of Quantum computing based on
thematic studies", In: Raju Pal and Praveen Kumar Shukla (eds),
SCRS CONFERENCE PROCEEDINGS ON INTELLIGENT
SYSTEMS, SCRS, India, 2021, pp. 313-320.
https://doi.org/10.52458/978-93-91842-08-6-31

4
Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on June 13,2024 at 13:39:09 UTC from IEEE Xplore. Restrictions apply.