4 100593163merged
4 100593163merged
Authors
y
Abstract
l og
Image classi cation is just a way that will balance the visual dissimilarity between machines and humans, which will help the
computer to interpret an image in a way as humans perceive. Despite its rampant application in practice, the conventional
no
methodology has been plagued by numerous problems. Its accurate classi cation precision is highly and arti cially limited
while unsatisfactory results are yielded. Thus, deep learning has been put forward as a possible improvement to the
h
accuracy of the solution to challenges involving image classi cation. This article addresses fundamental principles related to
ec
picture classi cation employing deep learning, examines different methodologies for image classi cation, highlights the
T
bene ts associated with deep learning, and explores the diverse applications that utilize deep learning in the context of
image classi cation. Arti cial neural networks, computational learning, deep learning techniques, and visual categorization.
nd
ea
nc
cie
fS
lo
na
ur
n Jo
dia
In
Ind
ian
Jo
urn
al
of
Sc
ien
ce
an
dT
ec
hn
olo
gy
Image Classification Using Deep Learning
y
l og
Dr. Shambu Kumar Jha
no
Assistant Professor, Galgotias University
h
Plot No.-2, Sector-17A, Yamuna Expressway, Gautam Buddha
ec
Nagar,
T
Greater Noida, Uttar Pradesh, India
nd
ea
Abstract: Image classification is just a way Introduction:- As reported by the Internet
nc
that will balance the visual dissimilarity Data Center (IDC), the entirety of global data
between machines and humans, which will help It will add up to 45ZB by 2020. Much of it,
cie
the computer to interpret an image in a way as some 70%, will seem like images and videos.
humans perceive. From computer vision, big data can now be
fS
Despite its rampant application in practice, the extracted from these pictures and videos.
conventional methodology has been plagued by Computer vision is getting more and more
lo
numerous problems. Its accurate classification common these days in the picture classification
precision is highly and artificially limited while space. Since its development in the late 1950s,
na
unsatisfactory results are yielded. Thus, deep image classification has been widely used in
learning has been put forward as a possible
ur
learning, examines different methodologies for Classification is the systematic structuring and
image classification, highlights the benefits grouping of items based on its attributes. In an
In
associated with deep learning, and explores the attempt to bridge the perceptual chasm between
diverse applications that utilize deep learning in man and machine, image categorization is used
the context of image classification. to provide data that enables machine learning.
Artificial neural networks, computational
learning, deep learning techniques, and visual
categorization.
images is based can be generalized into four and thousands of images in datasets are used to
categories. train
y
geographical traits, and traditional color
l og
schemes.
no
3. Image classification using shallow
learning.
h
ec
4. Deep learning powered categorization by
images.Machine learning is the standard
T
technique for classifying images. A feature
nd
extraction module in machine learning extracts
significant features such as edges and ea
texturessuggestion for categorizing
nc
5. In this work image classification using
deep learning is covered. For this task, the
cie
y
l og
Owing to the data-driven nature of the
analysim of ever-growing volumes of varying
data, transfer, and semi-supervised learning, as
no
well as the automated construction and
h
assessment of deep learning models using AI-
ec
driven methods, it is sometimes difficult to
overemphasize the role and significance of
T
artificial intelligence in this regard. It is no
nd
secret that AI-powered technologies are
establishing themselves in numerous areas of ea
human activity, and most researchers or
entrepreneurs claim that Every technology
nc
sector has its own ‘big expectation’, so the
artificial intelligence’s ‘big hope’ rests many
cie
y
encoding and generate predictions in a Neural Networks (CNNs) for picture
l og
hierarchical structure. categorization.
A CNN consists of many layers.
no
The process begins with an input image. 1. The convolutional layer is responsible
h
for transmitting data to the following layer.
ec
This generates a feature map through the
application of several filters, subsequent
T
2. The pooling layer pools a group of neurons
application of ReLU functions to augment non-
to one neuron in the next level.
nd
linearity, implementation of pooling layers for
each feature map, and transforming the images ea
into a singular elongated vector. The vector then 3. A fully connected layer where each neuron
gets input in a fully connected artificial neural in one layer is connected to every neuron of the
nc
network after flattening. Train the network by another layers.
using both forward and backward propagation. CNN uses many hidden layers, sometimes
cie
Repeat this process until the network is well- hundreds of them, to enable feature recognition.
defined and has feature detectors and learned The characteristics learned are somewhat more
fS
using CNN
na
ur
n Jo
dia
In
Let's take a closer look at how CNN is used for
picture categorization.
What we see
y
l og
h no
T ec
nd
ea
nc
cie
experience it.
na
Instead of seeing an image, a computer sees an The convolution layer is inserted before the
Jo
array of pixels; for example, if an image has 200 nonlinear layer. In the nonlinear layer, an
by 200 pixels, the array's size is 200 by 200 activation function is added, which provides with
n
pixels, with 200 being width, 200 for height, nonlinear properties; otherwise, without this
dia
y
pass through various stages:
Python comes with TensorFlow, it was employed
l og
in this manner. Approximately thousands of
Model building is actually started with the object flowers were gathered, and five different
no
model = Sequential(). varieties were present, each with hundreds of
compose layers based on their model types.add photos. All of these data sets are trained by the
h
the type_layer function. DNN until the system recognizes every image, at
ec
Once sufficient numbers of layers are included, which point each image is classified. Their
the model is compiled. In addition to specifying accuracy rate was about 90%. Ram, S., &
T
an optimization algorithm, a loss function will Chauhan, K. (2018)
nd
need to be defined. Once developed, the model is
[2]
applied to a previously untouched data set and its
ea
effectiveness is measured. They used, in the article, a deep learning
nc
convolutional network based on keras and
After the training of the model is complete and it tesnsorflow with Python for the picture
cie
produces the correct outcome, the model may be categorization. Four different CNN structures on
saved.save("file_name"). The model applies in CPU systems were tested, as well as four
fS
reality. This step is termed as model evaluation, different combinations of classifiers and
meaning that the data can be evaluated using the activation systems
lo
only issue was that it took the model around 50 parameters for image classification in Python,
minutes to display the 60 epochs. then created a CNN with two convolutional
ur
y
feature extraction techniques. The authors used the research called "Image Classification
l og
the CUDA library for the GPU and used Python Algorithm based on Deep Learning – Kernel
and TensorFlow for the CPU. Their Function". This way of computer vision includes
no
performance evaluation covered a range of two tasks - first, the exact identification of
aspects such as accuracy and the time taken to pictures and then the extraction of features from
h
execute both for CPU and GPU. M. Neelima them. The course of action is introduced by
ec
and M. Manoj Krishna (2018). utilizing the sparse representation of the deep
learning network architecture. The main
T
advantage of using the representation is that it is
nd
an effective way of solving the function
[4] They used the architecture of AlexNet and approximation problem, which can be very
ea
the convolutional neural network for complicated.
categorization in order to classify pictures in the
nc
proposed system. They selected four test photos Build an adaptive approximation capable model
from their database, resized the photos to
with deep learning.
cie
even for the portion of test images it shows deep learning algorithm. The system through
how deep learning the autonomous classification of images, makes
ur
model that had been pre-trained on image-net. training consists of four primary steps:
The deep learning CNN model's performance preprocessing the data, feature extraction,
In
was enhanced by the application of the transfer normalization of the data set, and SVM image
learning technique, which allowed for more classification. Xin Mingyuan and Wang Yong
accurate classification even with shorter (2019).
computation times and less hardware. Results
from the SVM technique, the region ranking
SVM method, and the super vector coding of
the local image descriptor approach were
compared with the acquired results.
[8] proposed a research effort known as the fine-tuning, a hybrid, and the CNN model with
Deep Convolution Neural Network-Based Image pre-training. The CNN model was passed
Classification Model. The research into the through the network after pre-training and fine-
backpropagation method provided a tuning the model, whereas in the hybrid
foundational basis for the training of a deep approach, patch-based feature extraction is
neural network intended to reach the maximum used. These methods provide efficiency with
span of minimal classification error. For this high performance and speed as their prime
experiment, they explored three approaches: advantages.
y
l og
Published by Leonardo da Vinci and Sumanth Reference Advantages
Roychowdhary, a research paper on "Image Mohd Azlan 1- An accuracy
no
classification using deep learning and prior Abu, Nurul of 90% was
knowledge" was only present in their 2019 work. Hazirah found .
h
Indra(2019) [1]
ec
[9] In their work, they brought historical Karan Chauhan, 1- A good
ShrwanRam(2018) [2] accuracy was
T
knowledge into deep learning methods by
found in
simply using semantic-based regularization.
nd
result.
ea 2- Model took
This method can be utilized for group very small
classification or as part of a learning process. amount of
nc
The CIFAR-10 dataset, which consists of time
60,000 images, formed the basis of their fo
cie
were the very low costs associated with test 2017) [3] classifying
classification and the impressively quick images on
na
even portion
dia
Muthukrishnan 1- Good
1. Mohd Azlan Abu, Nurul HazirahIndra(2019)
Ramprasath, M. Vijay accuracy of
[1]- Image classification using deep neural
Anand(2018) [7] almost 98%
network.
Mingyuan Xin 1- Better
2. Karan Chauhan, ShrwanRam(2018) [2]-
y
and Yong spee
l og
Wang(2019) [8] d and Image classification using deep learning
performance based on keras and tensorflow.
3. Dr. Vinayak Bharadi, Misbah
no
Michelangelo 1- Collective
Diligenti, classification NaimuddinPanchbhai(2017) [3]- Image
h
Soumal shows classification using deep learning.
ec
iRoychowdhary(2019) improvement 4. M Manoj Krishna, M Neelima(2018) [4]-
[9] over Image classification using deep neural
T
individual network.
nd
classification 5. W. A. Ezat, M MDesouky and N A
. ea Ismail(2020) [5]- Image
2- Cost of classification using convolutional neural
collective network.
nc
classification 6. Jun-e-Liu, Feng-Ping An(2019) [6] Image
was
cie
7. MuthukrishnanRamprasath, M. Vijay
reduction was
larger than Anand(2018) [7]Image Classification using
lo
knowledge.
dia
Conclusion:
After examining several approaches, it can be said 10. https://in.mathworks.com/matlabcent
that each strategy may have a few benefits over ral/fileexchange/59133-neural- network-
In