INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 02, FEBRUARY 2020 ISSN 2277-8616

A Survey On Deep Learning Methods And Tools

In Image Processing
B.Meena ,Prof.K.Venkata Rao ,Prof.Suresh Chittineni

Abstract: Deep Learning brings provocative light on various fields and research areas, such as image processing, Artificial Intelligence, and natural
language processing. it is an essential method applied to different applications in future. In recent years, the image processing and has been attracting
and increasing attention as it is one of the emerging areas of research. There are various applications for image processing in various domains such as
agriculture, space agencies, medical field, forensics and many others. As there is a progress in deep learning, many modern methods in deep learning
are recommended to improve image processing and image analysis performance. This document provides a broad overview on deep learning methods
used in image processing. Initially, an introduction on deep learning and image processing is given. Then, architectures of deep learning are introduced.
Then, the applications of deep learning on image processing like the image detection, image segmentation, and classification image classification are
explained. Thereafter, the benefits and weaknesses of deep learning tools common tools are mentioned along with the deep learning tools used in image
processing applications.

Index Terms: Artificial Intelligence, Convolution Neural Networks, Deep learning, Image Processing, Image segmentation, Machine Learning, Tools.
——————————  ——————————

1. INTRODUCTION network generates data, connections and discrete layer

Artificial Intelligence [1] is one of the branches of computer propagation direction . In 2011, the first-generation learning
science and Information Technology that creates computers systems DistBelief was launched by Google. Google has
or machines that behave as intelligent as humans. John been able to analyze thousands of their data from various
McCarthy, who is the father of artificial intelligence said that data centers Google is broadly using Deep learning in
"The Science and engineering of creating intelligent machines processing its products such as Google search and Google
particularly artificial intelligence attempts to simulate photos etc. .One of the important aspects in image processing
intelligence of humans and results with a new intelligent and analysis is feature representation. The following are the
machine that shall be able to produce information as human two advantages that deep learning yields: (1) Given a
awareness and behavior . Artificial Intelligence has the dataset, deep learning finds features naturally for a definite
application in various fields such as natural language application. Usually semi automatic learning method is a
processing, Neural Networks, robotics and image processing. classical feature extraction method that uses a prior
The heart of artificial intelligence is machine learning [2] and knowledge to extract the features (2) Deep learning can find
the basic approach for designing intellectual systems new features that are relevant to the suitable applications.
.Machine learning includes a many practices on statistics and According to traditional extraction methods , they are always
probability theory, convex analysis and approximation as well defined by previous knowledge, which only collect certain
as algorithm complexities. the concepts of machine learning features associated with an appropriate application. In
essentially adopts intuition and integration, in order to create addition to this, there are two other components that alter
the contemporary learning by reproducing the human nature image processing results . they are image acquisition and
and performance ,Then restructures the new knowledge to interpretation : Image Acquisition: It is demonstrated
improve the performance regularly. Machine learning has experimentally by researchers that the higher the image
been extensively applied in various areas such as space quality, the more the results obtained in image processing and
agencies, automation, industry Finance, Industry Healthcare, analysis. Nonetheless, image quality depends on the image
Government Organization, Marketing, etc. The concept of acquisition
deep learning [3] has been proposed with the deepening of Image interpretation: It is a procedure of reviewing an aerial
artificial neural network[3],[4],[5] Deep learning uses Deep image or a digital remote sensing image and manually
neural networks and it’s a software that simulates human determines image features. This approach is very trust
brain neural network. Hence Deep learning is a subdivision of worthy and broad features can be collected from an image.
Machine learning which is bringing the machine learning and some of them can be riparian features ,anthropogenic
artificial intelligence closer to create innovative applications. features etc . Yet it takes lot of time and only expert people
The human brain is an interdependence of network of like image analyst having sound knowledge can do it.
neurons , this scheme has lead to the emergence of artificial Although image interpretation works best on the built in
neural networks. Artificial neural network has a dissimilarity attributes of image which are composed of seven aspects
with human brain, that the human brain neurons are such as texture, size, association, tone, pattern and shape .
These aspects are used to know the information about the
———————————————— objects present in the image.
 B.Meena ,(Research Scholar, Department of CS &SE ,AUCE, However to get better results during image interpretation, it is
Visakhapatnam) ,Associate Professor Department, Raghu very important to design and develop a system that would
Engineering College,Dakamarri,Visakhapatnam, E-mail: interpret the images automatically. This kind of systems
[email protected]
should incorporate abundant operations like image
 Prof.K.Venkata Rao,Professor,Department of CS & SE,Andhra
University,Visakhapatnam.E-mail: [email protected] classification[20,27], image registration, image
 Prof.Suresh Chittineni ,Professor,CSE Department ,GITAM segmentation[20,27] and image detections. This paper
University,Visakhapatnam.E-mail: [email protected] focuses on deep learning by providing a broad inspection
connected through physical path especially the neural about the architectures and applications of deep learning over
1057
IJSTR©2020
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 02, FEBRUARY 2020 ISSN 2277-8616

image processing. Later the pros and cons of various tools y= f (x; ɵ). This mapping function y acquires the value of ɵ
are deliberated along with the tools that are applied to image settings to compute the best equivalent functions. Usually, a
processing. Finally the objective of deep learning in image network of direct action in extent consists of an input layer,
processing is expressed along with forth coming tools different hidden layers and an output layer. In addition to that
the information flow in deep feed forward network occurs in
2. ARCHITECTURES OF DEEP LEARNING one direction and does not move backward . This is depicted
in the sample figure . In this example the deep feed forward
2.1 Artificial neural networks (ANN) : network has few input layers, hidden layers and one output
ANN is important because just like human brain network it layer. Hence the deep feed forward network is one the ancient
grasps the information which is processed to establish various architecture of deep learning
models pertaining to neural network. The concepts of artificial
neural network evolved during 1980’s. These models of 2.3 Auto encoders Stacked
artificial neural network are calculation models having huge A simple deep feed forward network which is simple and
finite nodes and links among them. every node in an having input layers , hidden layers and an output layer is
distinguished ANN present an image of particular output referred as an auto encoder[6],[7],[8],[9] . A decoder and an
function . The link between the nodes communicate encoder are two parts of an auto encoder according to its
information between them . This information is represented as functionalities .An encoder is notated as f(x) that produces a
a weight . Hence ANN is an algorithm whose output rely upon decreased characteristic description from x,h where x is a
weights and activation function (output function) . In primary input h is a hidden layer. the notation of decoder is
accordance with, ANN algorithms are mainly classified into g(f(x)) . the decoder is familiar with reorganizing the basic
feedback neural networks and feed forward neural networks. input from the encoder output. the decoder minimizes the loss
Few of them are described below : Feed Forward neural function during reorganizing process:
networks: It is graph which is directed and having no cycles
and no feedback in the network. It is applied for the L(x,g(f(x))
conversion the input space information to output space. As it
is combined with various simple nonlinear functions; it has got The high-dimensional data is converted to low-dimension by
a good capability of processing information. its network encoding and decoding technique. Hence auto encoder is
topology is simple and easy to apply. much beneficial in image classification and various identical
Feedback neural networks: It is a complete graph having a tasks. The auto encoder is derived from sparse auto encoder,
feedback and no directions .it has state transitions while denoising auto encoder and contractive auto encoder. sparse
processing information. In order to handle the information auto encoder is dissimilar to auto encoders. the sparse
processing it uses dynamic system theory. the network constraint Ω(h) is added to the hidden layer. Hence its
cohesion is firmly connected to the function of associative restoration error is computed as
memory. It’s a decade that the research on artificial neural
network is getting deepened and a lot of advancement has L(x,g(f(x)))+ Ω(h)
been made in various domains in biology, smart robots,
finance and medical. In addition to these, artificial neural De noising auto encoder : the objective of denoising auto
networks have been strongly used to solve countless encoder is to minimize the loss function , they are dissimilar to
problems by showing many intellectual characteristics. sparse auto encoder
Commonly the attributes of artificial neural networks are L(x,g(f(x)))
primarily mirrored on associative memory, self-learning
capability and quick fixing the solution using optimization. Contractive auto encoder: Similar to sparse auto encoders,
they decrease the specific regularizer by accumulating Ω(h) to
hidden layer .The explicit regularizer is computed as as :

L(x,g(f(x)))+ Ω(h)

where Ω(h) is the squared Frobenius norm of the Jacobian

partial derivative matrix of the encoder function f(x) and λ is a
free parameter. The neural network with many auto encoder
layers is referred as stacked autoencoder . In addition to that
lateral layer input is taken from the former layer output in
stacked auto encoders . An auto encoder usually consists of
ternary layers only without having any stacked auto encoders
like the architecture of deep learning . It is much noted that
Figure 1 : Sample deep feed forward networks stacked auto encoder can only finish the task when they are
trained . As an example in order to train a nmk network ,
2.2 Deep feed forward networks the stacked auto encoder need to be trained with nmn for
The traditional deep learning models are deep feed forward obtaining a conversion nm and then train mkm network
networks. deep feed forward networks are also called as feed to obtaining a conversion of mk . At the end it stacks the
forward neural networks. deep feed forward networks aim to transformations (conversions) to form nmk auto encoder .
train the network for computing the equivalent objective This series of actions are referred as unsupervised pre-
functions. A deep feed forward network is described as training in layers.

1058
IJSTR©2020
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 02, FEBRUARY 2020 ISSN 2277-8616

In this portion of the paper we provide an analysis on the

2.4 Deep belief networks advancement of putting deep learning architecture in all the
Unlike stacked auto encoders , deep belief networks are above areas mentioned.
likewise referred as neural networks and is a special form of
Boltzmann machine[10],[11],[12] with some restrictions . 3.1.Deep learning in image segmentation
Furthermore , the input of next layer comes from output of One of the important and interesting concepts in image
previous layer in deep belief networks. In order to pre-train processing is image segmentation
every restricted Boltzmann machine hierarchically, the deep The major categories of computer vision are:
belief Boltzmann machine accustoms an unsupervised greedy 1. Image Classification
method which is pre trained . the output of this used as an 2. Image Detection
basic input for supervised learning probability model. 3. Image segmentation
Boltzmann machine is a method of modeling that uses The classification of images by identifying an existence of an
functions of energy and has been derived from statistical image in a picture is known as image classification. Image
physics. This characterizes the higher interactions among the detection follows image classification and localization. the
variables .when compared it is complex having total tangible concept localization finds the area if a given object is located .
analysis and is supported by strict statistics concepts. It is A boundary is drawn around it to highlight the region. The
symmetrically joined by arbitrary binary unit neural network difficult and the most beneficial type among the above three is
with feedback having visible and hidden units. A special image segmentation. This is related to classify every pixel into
form of Boltzmann machine is referred as restricted a group of objects in a given picture.
Boltzmann machine that includes only a visible layer and a
hidden layer . In restricted Boltzmann machine the
connections between the hidden layer nodes and visible layer
are bi directional. the restricted Boltzmann machine have
faster calculation speed when compared to Boltzmann
machine as it has only one hidden layer .The restricted
Boltzmann machine has two main functions : (1) Used to
reduce the dimensions of data. (2) A weight matrix is acquired
that is a basic input to other neural networks.

2.5 Convolutional neural networks

Convolutional neural networks [24][25] are broadly used in
tasks recognition like image classification, handwriting
recognition and document recognition. The dissimilarity in Figure 2 : Image Classification of single object and multiple
convolution neural networks and fully connected feed forward objects
is their connection type among the adjacent layers. In feed
forward neural networks some nodes are connected where as Image segmentation is classified into simultaneous detection
in convolution neural network, all the nodes are completely and semantic segmentation. Both of the type of image
connected. Convolution neural networks consists of two segmentation seems to be similar but the dissimilarity is
kernel layers namely the Convolutional and pooling layers. simultaneous detection identifies exclusive objects where as
Convolutional Layer: the input of each node is only a tiny semantic segmentation deals with pixel categorization. The
area of the previous layer and the size of this tiny patch is above figure is considered to quote the dissimilarity of these
usually 3 X 3 or 5 X 5. the convolution layer analyzes deeply two types of segmentation. as per simultaneous detection
each small portion of the neural network and results a better which is also known as instance aware segmentation find
feature representation. Pooling Layer: The layer followed after each cat image separately where as the semantic
the conventional layer is pooling layer .The purpose of this segmentation deals with categorizing all cats as a single
layer is to decrease the matrix size from convolution layer, the instance .These kind of problems can be solved easily using
result of this layer is also to decrease number of parameters deep learning techniques.
in neural network The advantage of this pooling layer is to
prevent over fitting problem and speed up the calculation. 3.2. Image Segmentation using Convolution Neural
Usually the convolution neural network architecture are Networks (CNN):
categorized into two types basing on their connection modes Here in this procedure the pieces of images which are marked
.first type connects various convolutions layers in an order .As are inserted as an input of convolution neural network. This
an example LeNet-5,AlexNet and ZfNet , while the second procedure is done again and again on the complete array of
one connects in parallel like Inception and later versions pixel. This architecture has a disadvantage of performing this
set of operations only over well consistent image. Passing
3. APPLICATION OF LEARNING IN DEPTH IN complete images to CNN:  In this architecture a complete
THE IMAGE PROCESSING image is given to the network where the pixels are labeled in
There are many deep learning methods that have been a single step rather than many steps. It is an improved
proposed for applying in Image analysis and Image architecture than traditional one . here the mask of
processing areas, like Image classification, Image registration, segmentation is decreased in size due to the convolution and
Image segmentation and Image detection. The problems in pooling technique . if the input of an image is 512 X 512 , the
the above mentioned areas can be solved using feature result may be 28 X 28. Fully CNN: The next improved
representation in collaboration with Deep learning techniques. architecture than the earlier one is fully convolution neural

1059
IJSTR©2020
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 02, FEBRUARY 2020 ISSN 2277-8616

network. Here the final tier of CNN follows Deconvolutional occur simultaneously accompanying every class that has
neural network . Given a picture, image segmentation is very images in number 6000.the images in this dataset has no
essential to figure out varied things and relationship. The noise , small in size and are definitely labeled . Hence this is
utilization of image segmentation differs from simple an ideal dataset to finish the task significantly with a fewer pre
applications in image recognition to spatial images as they are processing steps. the following are some of the pictures
exemplary generous to know the texture . they are applied to collected from the above dataset:
medical images to identify cancer cells and even in surgery
preparations . This is an essential constituent in searching the
images in Fashion and retail enterprises

3.3. Classification of images using Deep Neural Networks

Image classification is a process of drawing out the
information of classes from a multiband raster image .
Here the computer is taught to identify images and categorize
into the some of the given classes. In this example some
images are given to the computer and it could classify into 10
categories are as follows

Figure 4 : Pictures from cifar-10 data set and image

Here are the steps to apply image classification upon the

above data set
1. Pre-processing
2. Splitting the dataset
3. Building a CNN

Figure 3 : Image Classification

Here the computer is taught with the behavior of bird, dog, cat
etc, in order to find out the objects. The more cats the
computer gets trained, it recognizes the more cat objects. This
kind of training is said to be supervised learning. Here the
images are labeled to accomplish such task. Google has
given an open source framework for deep learning called as
Tensor Flow. The developer would have a better control at
micro level over each node (neuron) in Tensor Flow. To gain
the best performance the weights have to be set. For image Figure 5 : Image after applying above three steps
classification Tensor Flow has built in APIs. Convolution
neural networks is one of the well know techniques to improve Step 1: Pre-treatment: In first step of preprocessing, the data
image classification accuracy. In this technique the image is does not have any noise and it is organized so variance is
broken up into several parts where the machine tries to added to the data. The noise can be added using python
identify each part rather than sending the image array as an library named imgaug. Then a random combination of
input . this technique of sending the parts to machine allows to cropping the image, flipping and applying saturation and hue
accomplish the task parallel and identify the objects adjustments is done.
irrespective of the object location. Step 2: Splitting our data set: Computing gradient for the given
large dataset model is a time taking process . Hence the
3.4 . Sample data images are processed in tiny batches at every optimizer
In this example a dataset CIFAR-10[20] having images iteration. Generally the size of the batch is considered as 32
calibrated 60 000 in number each of size 32 X 32 pixels is or 64. In this example 64 is the batch size as it has big
taken. this dataset is categorized into ten classes that do not number of image. Further the dataset is broken up into
training set each having 50000 image and test set having 10

1060
IJSTR©2020
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 02, FEBRUARY 2020 ISSN 2277-8616

000 images TABLE 2: DEEP LEARNING TOOLS FOR IMAGE

Step 3: Building a convolution neural network : After pre PROCESSING AND REFERENCES
processing and splitting up the neural network , the
implementation begin with three convolution layers having a Tool name URL references
max pooling of 2 X 2 .max pooling - a technique that reduces OpenCV http://www.opencv.org/
the image dimensions . It collects maximum value of a pixel in
BoofCV http://www.boofcv.org/
the grid . Hence it makes the model generic and over fitting is
reduced . Below is an example that show the working of 2 X 2 NASA Vision http://t1.arc.nasa.gov/tech/asr/intelligent-robotics/nasa-
Workbench vision-workbench/
max pooling . Here is a max pool with 2 X 2 filters and stride
SimpleCV http://simplecv.org
2.
Tesseract http: //github.corn/tesseract-OCR / Tesseract
SLIC http://ivrl.epfl.ch/supplementary_material/RK_SLICSSu
superpixels perpixels/
OpenMVG https://github.com/openMVG/openMVG
LIBVISO http://www.cvlibs.ne/software/libviso
VisualSFM http://hornes.cs.washington.edu/~ccwu/vsfm/
MeshLab http://meshlab.sourceforge.net/
Bundler http: //phototour.cs/washington.edu/bundler/
Vid.stab https://github.com/georgmartius/vid.stab

Figure 6: Two cross two max pooling 5. CONCLUSION:

The aim of this paper is to present relevant information for
Now the neural network is trained and ready to use . An investigators and researchers, with respect to the deep
artificial convolution neural network can recognize the image learning applications in image processing research area. In
more accurately than previous approach using tensor flow . order to acquire better results in image processing research
area, the architectures of deep learning are extensively used
4. DEEP LEARNING TOOLS : in sub domains like image segmentation, image classification,
image analysis, image recording and image sensing. Deep
4.1. Common tools of Deep Learning learning has faced various challenges and proved that these
Various tools methods give improved and more appropriate results for
[13],[14],[15],[16],[17],[18],[19],[20],[21],[22],[23],[24],[25] of researches working in image processing. To achieve the best
deep learning have been evolved that are convenient for results the large data sets are to be given to the network,
analyst and researchers. this advancement helps a lot for all limited data set would lead to imbalances in classes. The
those working in deep learning. The table below introduces large the data set, deep learning learns in depth. Deep
various deep learning tools. learning with image processing would present outstanding
achievements in future.
TABLE 1 : SOME OF THE DEEP LEARNING TOOLS AND
REFERENCES REFERENCES :
[1] RABrooks Intelligence without representation. Artif.Intell., Vol 47
Tool Name URLs 139-159 1991
DeepLearnTool https://github.com/PierreHao/Deep-Learning-ToolBox- [2] B.Bishop, Recognition and Machine Learning.Springer-Verlag,
box CNN 2006
[3] Y.Bengio, learning deep Architechtures for AI, Found. Trends
Caffe http://caffe.berkeleyvision.org/
Mach. Learn., Vol.2 no.1,2009, (1-127)
Torch http://torch.ch/ [4] B.Yegnanaryana, artificial neural networks. New Delhi, India: PHI
Theano http://deeplearning.net/software/theano Learning Pvt. Ltd. 2009
Pylearn2 http://deeplearning.net/software/pylearn2/ [5] H.Bourlard and Y.Kamp, Self-organization by multi preceptrons
Keras https://github.com/EderSantana/keras and singular value decomposition, Biol.Cybern .. PP 291-294
tensorflow https://www.tensorflow.org/ [6] RS GEHinton and Zemel, Autoencoders, minimum description
CNTK https://www.microsoft.com/en- length and Helmholtz free energy in Proc. 6 Int.Conf.Newural
us/research/product/cognitive-toolkit/ 1994 Information processing systems
[7] G.Alain and Y.Bengio, this regularized self-learning encoders from
MXNet https://github.com/dmlc/mxnet
the distribution data generating, Mach.Learn J. Res., Vol 15,
chaining http://chainer.org/ No. 1, pp 3563 to 3593.2014
Deeplearning4j https://deeplearning4j.org/ [8] MARanzato, C.Pultney, S.Chopra and Y. LeCun, effective
learning of the few performances with an energy-based model
MatConvNet http://www.vlfeat.org/matconvnet/
in the 20th annual conference of information processing
maxDNN https://github.com/eBay/maxDNN systems Neural. 2006
[9] P. Vincent H. Larochelle, I. Lajoie, Y. Bengio and PA Manzagol,
autoencoders denoising Stacked: Learning useful
representations in a deep network with a criterion of local
1061
IJSTR©2020
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 02, FEBRUARY 2020 ISSN 2277-8616

noising, J. Mach. Learn. Res., Vol. 11, pp. 3371-3408, 2010.

[10] Yi Pan,Min Li,Ziyue Chen,Lu Tang,Chengqian Lu. and Jianxin
Wang Applications of deep learning to MRI images : A survey
Jin Liu . ISSN 2096-0654 01/06 pp 1– 18 Volume 1, Number 1,
March 2018
[11] SE Fahlman, GE Hinton and TJ Sejnowski, Massively parallel
architectures for AI: NETL, thistle and Boltzmann machines, in
Proc. AAAI 3rd Conf. Artificial intelligence, Washington, DC,
USA, 1983.
[12] RB Palm, forecasting as a candidate for learning deep
hierarchical data models, master thesis, Technical University of
Denmark, Denmark, in 2012.
[13] Y. Jia, Shelhamer E., J. Donahue, S. Karayev, J. Long, R.
Girshick S. Guadarrama and T. Darrell, Caffe Architecture for
embedding fast convolution function, in Proc. 22 ACM Int. Conf.
Multimedia, Orlando, FL, USA, 2014, pp. 675-678.
[14] R. Collobert, K. and C. Kavukcuoglu Farabet, Torch7: Matlab
type learning environment of the machine, BigLearn, Workshop
NIPS, No. EPFL-CONF-192376.2011.
[15] R. Al-Rfou, Alain G., A. Almahairi, C. Angermueller, D. Bahdanau
N. Ballas F. Bastien, J. Bayer, A. Belikov, A.Belopolsky, et al.,
Theano : A python framework for rapid calculation of
mathematical expressions, arXiv preprint arXiv: 1605.02688
2016.
[16] IJ Goodfellow, D-Warde Farley, P. Lamblin, V. Dumoulin, Mr.
Mirza Pascanu R., J. Bergstra F. Bastien and Y. Bengio,
Pylearn2: A machine learning research library , arXiv preprint
arXiv: 1308.4214 2013.
[17] F. Chollet, Keras: Théano-libraryCode learning base long,
https://github.com/fchollet. Documentation: http: //keras.io 2015.
[18] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, ZFChen, C. Citro,
GS Corrado, A. Davis, J. Dean, M. Devin, et al, tensorflow :.
Large scale machine learning on heterogeneous distributed
systems, arXiv preprint arXiv:
[19] D. Yu, A. Eversole, Mr. Seltzer, KS Yao, O. Kuchaiev, Y. Zhang,
F. Seide, ZH Huang, B. Guenter, HM Wang et al., Introduction
to computer networks and computer network transmission
tools, technical report
[20] https://medium.com/@tifa2up/image-classification-using-deep-
neural-networks-a-beginner-friendly-approach-using-tensorflow-
94b0a090ccd4
[21] TQ Chen, Li, Li YT, Lin, NY Wang, J. Wang, TJ Xiao, B. Xu, Z.
Zhang and Zhang CY, MXNet: A flexible efficient Machine
learning library for heterogeneous distributed systems, arXiv
preprint arXiv: 1512.01274 2015.
[22] S. Tokui, K. Oono, S. Hido, J. and Clayton, Chainer: A new
generation open source framework for in-depth learning, in
Proc. Workshop on Machine Learning Systems (LearningSys)
in the Twenty-ninth Annual Conf. Neural Processing Systems
(PNM), in 2015.
[23] Deeplearning4j development team, Deeplearning4j: Open-
source distributed deep learning for JVM, Apache Software
Foundation License 2.0, http://deeplearning4j.org 2017.
[24] BC Ooi, KL Tan, S. Wang, W. Wang, Cai QC, G. Chen, Gao JY,
ZJ Luo, AKH Tung, Y. Wang, et al, SINGAPORE. A learning
platform distributed depth, Proc. 23 ACM Int. Conf. Multimedia,
Brisbane, Australia, 2015, pp. 685-688.
[25] A. and K. Vedaldi Lenc, MatConvNet: convolutional neural
networks for Matlab, in Proc. 23 ACM Int. Conf. Multimedia,
Brisbane, Australia, 2015, pp. 689-692.

1062
IJSTR©2020
www.ijstr.org