Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)

IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

Skin Cancer Detection using Deep Learning

2022 International Conference on Electronics and Renewable Systems (ICEARS) | 978-1-6654-8425-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICEARS53579.2022.9751826

Dr.R.Senthil Kumar Dr.Amarjeet Singh Sparsha Srinath, Department of

Associate Professor, Department of Dean Academics, Department of Computer Science and Engineering
Computer Science and Engineering Computer Science and Engineering New Horizon College of Engineering
New Horizon College of Engineering New Horizon College of Engineering M arathalli,Bengaluru
M arathalli,Bengaluru M arathalli,Bengaluru [email protected]
[email protected] [email protected]

Nimal Kurien Thomas, Department of Vishal Arasu, Department of Computer

Computer Science and Engineering Science and Engineering
New Horizon College of Engineering New Horizon College of Engineering
M arathalli,Bengaluru M arathalli,Bengaluru
[email protected] [email protected]

Abstract— Identifying mel anoma at the early stages of utilized the k-means clustering algorithm as well as the
di agnosis is imperati ve as early detection can Support Vector Machine(SVM ) so as to determine a certain
exponentially increase one's chances of cure. The paper classification method for the types of melanoma. Abbas et
first proposes a literature survey of multi ple methods al. [3], and Germán C et al. [7] looked into AdaBoost MC to
used for performing skin cancer classification. Our create an image classificat ion method of classificat ion of
methodology consists of using Convolutional Neural different kinds of cancers and skin lesions image
Network (CNN) to i dentify and di agnose the skin cancer classification methods separately. Michael B et al. [23], and
using the IS IC dataset containing 2637 images. The Vicente B et al. [5] factored into the account that lesion
proposed model gi ves an accuracy of 88% for classifying texture, colour, affected area and degree of damage and
the training dataset as either benign or malignant. developed a system of support for the decision for the
differentiation of melano ma. An Android application was
Keywords—Convolutional Neural Networks, Deep also developed by Kiran et al. [20] fo r the melanoma
Learning Models, Image Processing, Support Vector determination in a clin ical approach. In this paper we have
Machine, K Means Clustering decided to employ a CNN technique on the ISIC dataset
fro m Kaggle within the Keras python library in order to
classify the given input or data as either malignant or
benign.

I. INTRODUCTION II. LITERATURE REVIEW

Melanoma is identified by the changing or unusual on S. R. Guha et al. [20] - The author started off the research
both sun-exposed and sun-protected areas of the body. with the intent of studying the existing accuracy of a certain
Melanomas usually appear on the legs of women, and the skin disease detection algorithm that used CNN and wished
most common area of occurrence on men is the trunk. to expand the accuracy which could have been obtained.
Therefore, it is important to keep in mind, that melano mas Since the goal was determined, the writers had moved on to
can arise anywhere on the skin, even in areas where the sun preprocess the input image to detect the lesion part of the
doesn’t shine. The recurring method for skin cancer skin and wrench out the features of the afo rementioned part.
classification has been variations of Convolutional Neural two of the chosen models were prepared and compared and
Networks. Deep Learning models are mo re beneficial as they claimed to obtain better accuracy with the transfer
they require no hu man intervention wh ile performing image learning technique of the model 'VGG16'. The model has
processing, feature ext raction and has a significant been compared to show accuracy against existing
advantage over other strategies like Machine Learning architecture models wh ich have been selected for reference.
algorith ms wh ich can solve well-structured problems. But they had come to the conclusion of an 11% greater accuracy
for more co mplex data applications we require Deeper rate to the advantage of the transfer learning method and
Architectures. The automated machines and systems used further developments have been said to be in development
for the classification and differentiat ion of the diseases like to fine-tune such a model.
melano mas and Anas et al. [16] The research that was built
on the specified paper published by Hitoshi et al.[9] by the A. W. Setiawan [23]- The author had taken the inspiration
classification of melanomas into various different of the effects of UV lights rays which penetrate the ozone
classification (four to be exact), Almansour et al. [ 6] had layer and attack the hu man skin, which is one of the ways

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 1724

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on March 06,2024 at 04:27:55 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

humans get skin cancer. The writer had decided to develop a of Natural Language Processing or NLP for short such that
simp le image processing tool to detect early signs of cancer. there exists a user interface with the patients and their
The technique the author chose to use included image treatment head whenever required through the internet. The
improvement using MSRCR and CLAHE as a contradiction chosen data model had yielded an accuracy rate of 75% with
to a convoluted neural network. The result shows us that the the chosen input data set and deployment of the app would
former is more suitable in color enhancement of pictures to allo w further investigations and required data through back
detect early signs using convoluted neural networks but the tested data.
margins of co mparison 0.8476 & 0.8161 for the enhanced
datasets against the normal dataset did not prove to change Keun-Kwang et al. [6] The premise of this paper is to
the accuracy of validation or training. However, it d id prove devise a highly efficient CNN for classifying skin cancer.
to make the computational load allo wing it to be easily The above study classifies ECOC SVM into 4 different
portable to a smaller device. types of skin cancer. It uses an already existing AlexNet
CNN model in order to mine the features. The results
H. K. Kondaveeti et al. [8]- Th is paper has introduced the showed a tremendous accuracy for each kind of skin cancer.
technique of transfer learning in neural networks as a The highest values for maximu m accuracy was found while
method of train ing them to detect early signs of detection of classifying the squamous cell carcino ma(95%) and lo west
cancer. The proposed models in a HAM 10000 data samp le while classifying melano ma (90.74%). This study when
were given by data models of MobileNet, Inception V3, and compared with previous research showed higher results for
such. The altercations to the class imbalance had been made sensitivity, accuracy as well as specificity.
an attempt to resolve it by the augmentation of input data
and the adjustments to its bias weights. After such Yi-Xin Cai et al. [20] suggests a new and improved
methodologies were taken and tested, the selected model version of CNN which co mbines CNN and a whale
which is the ResNet50 had yielded a mu lti-class precision optimization algorith m wh ich was conceived by Lewis and
rate of about 90% (with weighted recall and accuracy being Mirjalili. Th is technique combines the high efficiency of
90% and 89% respectively). The application of such models CNNs with an optimizat ion algorithm to fu rther enhance the
within the first layer of the networks helps us to understand efficiency of the diagnosis. The whale algorith m was
how convoluted neural networks interpret images and how inspired by the behavior of hu mpback whales when trapping
they can be modelled after to present a better yielding their prey. The performance of th is novel technique was
output. The paper also introduced a way to reduce noise and analyzed against two primary parameters i.e DermIS and
external disturbances from input data before it is processed. Dermquest. The subsequent results revealed that the
This method helped to achieve higher performance o f the proposed method achieved the best results when compared
said model and saves on computational resources. with 10 d ifferent methods. Some of the methods in the
comparison were ResNet, AlexNet, Ord inary CNN, semi-
A. A. Ibrahim et al. [10]- The primary objective of the supervised method and Spot-mole.
Presented thesis work presented by the authors was to find
and build appropriate and acceptable CNN models which
are to be used for the detection of skin cancer post-training. III.CONVOLUTIONA L NEURAL NETWORKS
A large part of the research done by the authors was on
convoluted neural networks. To obtain a trained model with
CNN with the desired arch itectures of the author, the A Multifunctional network which can be applied in many
parameters such as the batch size, dropout value and epochs, different areas, A convolutional neural network is a certain
and other variables were ad justed to meet specificat ions. type of artificial neural network wh ich is primarily used to
process pixel data and such image data. In itself, it is a
The best-observed data model(for the given dataset) with an
multilayer perceptron made in a particular manner to
accuracy of 85% was retrieved by Xception. The authors
accommodate picture and image-related info rmation
also pointed out the risk of poor computational power and classification and processing. Inclusive
quality of images wh ich hinder the working and precision of to the prior fact, the architecture of a deep neural network
the convoluted neural network and hence can cause an error includes multip le layers to process more information and do
and extend the time required to train the dataset. The authors more co mp lex functionalities at the expense of
have moved on to implement computational power. Regardless of the restriction other
such work with real-life applications but might produce data model machines face, CNN is not as constricted for
devalued accuracy. adjoining layers of the connections in-between them. They
share weights and work the feel of a part icular local area of
Y. Huo [24] -The author of the presented thesis had the image data set. There are two p rimary object ives for any
emp loyed the use of convoluted neural networks to find given convolutional neural network, which are depicted in
early signs of cancer to t reat it with ease. The application o f Fig 3.1.
the CNN by the author required a Keras CNN model
through the TensorFlow, Java Script to classify and detect (i) A data set selection process, also referred to as
the processed image. the model had a backend built on sampling. We obtain 'n' number of data points(which in our
Exp ress JavaScript and the front-end GUI was implemented case would be pixels) for each of the neighborhood through
by Swift user interface. The data set introduced as the input a series of steps known as pooling. These steps start off with
value was one of HAM10000 with integrated functionalities identifying the pixel and then adding a bias before

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 1725

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on March 06,2024 at 04:27:55 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

weighting them by scalar property 'W x + 1'. Finally, the whether or not to fire a neuron resides upon the activation
activation function comes into play to produce a narro w 'n ' function. The most frequently used ones happen to be
times feature map. Sig mo id, ReLu and Tan H. The output is restrained to values
between 0 & 1 for all sig moid function inputs which happen
(ii) We employ the use of a trainable filter for the to be real numbers. TanH on the other hand is like sig mo id,
reconstruction of the given data set to be entered after wh ich but the output for this function is between -1 and 1. ReLU is
we add our personal bias to retrieve our final convolution widely considered as the most used activation function
layer. because of its Lower co mputational load. The end of the
convoluted neural network has a co mpletely covered layer.
Since the layer consists of neurons, each interconnected
with each other in the prior layer of the network. This
henceforth mimics an approach in wh ich it is fully
connected and is used as a classifier for CNN. The final
output layer is also shown to us by this output layer. The
output layer also emp loys the use of some loss functions to
find out the percentage of predicted error in the entire
sample in the model which had been made. This erro r
represents the increasing normalization of the actual and
predicted output. After which we finally use 'Re lu' to rectify
as an activation function. It is given to add non -linearity to
the system. The final output probability of each class will be
found to be taken in the final layer. The results fro m the
models used had yielded a 86% accuracy for our data
model. For the model in which we used 50 epochs with a
Fig 3.1: Convolution Neural Networks target size of (150,150) and a batch size of 32. It had
compared the pred icted probability to the actual class output
The advantages of the employment for our
in a binary cross-entropy to show an output of
implementation include the fully connected layers, which
allo w the computation of highly co mplex problems. It can 0.86(predicted) and a 1(actual). Given that our model
also be used for image and ob ject classification, self-driving showed consistent accurate results after execution, it proved
cars, etc. And yield very h igh-performance rates and to yield high efficiency and performance.
efficiency.

IV. NETWORK ARCHITECTURE

The CNN architecture is a culmination of d ifferent layers

hence the name mu lt i building b locks depicted in the Fig
3.1.

The layers are as follo ws-1) convolutional layer, 2) pooling

layer, 3) activation functions, 4) Fu lly connected layer and
5) loss functions. The convolutional layer is arguably the
most significant layer in the architecture. It is co mprised of
a collection of convolutional filters also termed as kernels.
The generation of the output feature map is created by the
convolution of N - dimensional metrics of the input image
by convolutional filers. Since the shrinking of the large size
map features are done by the pooling layers to create a
reduced feature map. The happen to be multiple methods of
utilizat ion. So me examples are gated pooling, g lobal
pooling average, average, tree, min, max pooling. Out of the
following, the most used ones happen to be the max and the
min pooling. There also happens to be a specific activation
function you map the input to the output in which the value Fig 4.1: Convolution Neural Networks
is determined by the calculation of the weight’s summation
of a particular bias of a neuron. Hence the final choice o f

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 1726

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on March 06,2024 at 04:27:55 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

V. IMPLEMENTATION of in formation fro m the given dataset, such as edges,

various shapes present, which are all combined to form the
The initial step taken for the imp lementation is the complete image. Convolution operations involve different
convolutional filters wh ich is a set of layers used understand mathematical operations, such as matrix arithmet ic
and get knowledge to train from the image input data operations, and they are presented in the form of an array o f
through the Keras library in python for a Keras sequential values. This is fo llo wed by pooling, wh ich is applied
data model. A certain number of filters were used at each sequentially to the data in the hopes of ext racting features.
layer each being 32 filters, 64 filters,128 filters,256 filters Following this is the flattening of the data into a feed -
for each of the first, second, third, and final layer. These forward neural network, referred to as a multi-layer
filters are often referred to as kernel filters and can perform perceptron. To move ahead, we went on to import the
very useful functionalities wh ich include but are not limited necessary libraries which were required for performing our
to the transformation of the images. The convoluted neural tasks and creating our model as such.
network can proceed to separate chosen features from this
which is necessary for classification. We then proceed to
downsize in the pooling layer of the convolution neural
network to pick the maximu m value of nearby pixels to Increasing Conv filters of 32,64,128,256 and kernel size=3
minimize co mputational cost. The effect of overfitting is along with a pool size of (2,2).
also hindered due to this while increasing normalization. We The Keras library layers are as follows
use the help of an activation function 'Relu' wh ich is used as Dropout(0.5),
a rectifier. It is given to add non-linearity to the system. The Flatten(),
final output probability of each class will be found to be Dense(units=1024, activation='relu'),
taken in the final layer. The results fro m the models used Dropout(0.25),
had yielded a 86% accuracy for our data model. The input Dense(units=1024, activation='relu'),
layer was activated using the ‘ReLU’ function while the Dropout(0.25),
output layer used ‘Soft max’. For the model in which we Dense(units=7, activation='softmax')
used 50 epochs with a target size o f (150,150) and a batch
size of 32. It had compared the pred icted probability to the We are then presented with the summary of our created
actual class output in a binary cross -entropy to show an model, and we specify the optimizers. We then calcu late our
output of 0.86(predicted) and a 1(actual). Given that our errors with our specified loss function.
model showed consistent accurate results after execution, it
proved to yield high efficiency and performance. Conv2D: it's the layer to convolve the picture into nu merous
pictures enactment is that the actuation work.
ISIC dataset was published by the International Skin MaxPooling2D: it 's acclimated max pool the cost fro m the
imaging co llaboration as a dataset of dermoscopy images. given size matrix and the same is utilized for the
Currently the ISIC Archive contains around 13000 accompanying 2 layers.
dermoscopic images, which were co llected fro m mu ltip le Flatten: it's utilized to flatten the elements of the picture got
med ical clinic centres. The inco ming images to the Archive subsequent to convolving it. Dense: it's utilized to make this
are screened for privacy as well as quality. Fro m the large an extremely associated model and is the hidden layer.
dataset, a subset of the images would have undergone Dropout: it's acclimated keep away fro m over fitting on the
annotation to include dermascopic features by many dataset and dense is that the result layer contains only one
recognised cancer experts. neuron which attempt to which classification picture has a
place.
Image Data Generator: it's that rescales the picture, applies
The general CNN algorithm has been inspired to be created shear in some reach, zoo ms the picture and does flat flipping
by the visual part of the brain, which is the visual cortex. It with the picture. The Image Data Generator incorporates
possesses various filters, each of which extracts some type each conceivable direction of the picture.

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 1727

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on March 06,2024 at 04:27:55 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

Training Process: Train_datagen. Flo w_fro m_directory is

that the capacity that is acclimated get ready information
fro m the train_dataset index Target_size determines the
objective size of the picture. test_datagen.
flow_ fro m_directory is utilized to orchestrate test
informat ion for the model and everyone is similar as above.
fit_generator is utilized to suit the information into the
model made above, different variab les utilized are
steps_per_epochs enlightens us concerning the amount of
times the model will execute for the preparation
information. Fig 5.2: Training Data Classification
Epochs: It lets us know the times model are prepared in
forward and in reverse pass. Table 5.1 g ives the comparative
analysis that depicts the relat ionship between accuracy and
epochs for every successive 50 epochs for the CNN model
used.
Validation process: Validation data is used to take care of
the approval/test information in the model.
The dataset is preprocessed utilizing Image reshaping,
resizing and transformat ion to an exhibit structure.
Co mparative handling is moreover done on the test picture.
A dataset containing threatening and harmless pictures are
utilized, out of wh ich any picture is usually utilized as a test
picture for the product. The train dataset is utilized to
instruct the model (CNN) so as that it can recognize the test
picture. After the model is prepared effectively, the product
can distinguish the sickness assuming the plant species is
held inside the dataset. After effectively preparing and
preprocessing, examination of the test picture and prepared
model happens to anticipate the infection
Table 5.1: Comparative Analysis of Epoch with
val_accuracy

VI. RESULTS

The model that we have created and implemented in this

paper showcases 86% accuracy along with differentiating
between malignant and benign form of skin cancerous cells.
It classifies the input image as either of the two types
mentioned in the previous statement. The Keras Tensor flo w
package is utilized for modelling the convolutional neural
network. The belo w graphs depict the model accuracy and
model loss respectively. They summarize the accuracy and
loss of our model over 50 epochs. It is observed that
throughout the training process the highest accuracy attained
was 86%. The training process shows that the testing dataset
is pretty accurate in classifying images for skin cancer
diagnosis. The accuracy is fairly less than what is usually
expected from a CNN, but we suspect that is due to the size
of the dataset which is smaller than what we intended to us.

Fig 5.1: Training Data

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 1728

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on March 06,2024 at 04:27:55 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

[4] Daniel R, Vicente B, Antonio S, Belén S,

“Decision support system for the d iagnosis of
melanoma- A comparative approach”.

[5] Dorj, UO., Lee, KK., Choi, JY. et al., “The skin
cancer classificat ion using deep CNN”, Published:
22nd Feb, 2018.

[6] Ebtihal A, Arfan JM , “The Classification of

Dermoscopic skin cancer images using color and
hybrid texture features”, Int J Co mputer Sci Net w
Security

[7] Germán C, Andrés C, Anabella B, Rodrigo A,

Pablo M, “ Study on : To ward a co mbined tool to
assist dermatologists in melanoma detection”

[8] H. K. Kondaveeti and P. Edupuganti, "Study:

Classification of Skin Cancer using Transfer
Learn ing, IEEE International Conference on
Advent Trends in Mult idisciplinary Research and
Innovation, 2020

[9] Harshitha R., R. Beham, Tho mas B., “A study on a

novel approach to detect acute lymphoblastic
leukemia”, IEEE,2018.

[10] Ibrahim and H. A. Hasan, “Hybrid Detection

Techniques for Skin Cancer Images”, Symposiu m
on Multidisciplinary Studies and Innovative
Technologies, 2020

[11] Ku mar A., Adhikary P., Bandyopadhyay S.,

Mazu mdar A., “A case study on chilled water
pump trouble shooting by AI”, 2019, ARPN
Journal.

[12] Mohd A, et al. ,”Study: Skin cancer classification

using K-means clustering” Int J Tech Res Apple.
Fig 5.4: Model accuracy and model loss graphs
[13] Oka H, Iyatomi H,Celebi M E, Ogawa K,
Argenziano G, Soyer HP, Koga H, Saida T, Ohara
K, Tanaka M, “Study: Co mputer-based
classification of Dermoscopy images of
REFERENCES melanocytic lesions on Acral volar skin”

[14] R. S. Ku mar and C. Ramesh, "A study on

prediction of rainfall using datamining technique,"
[1] Balasubramaniam, Vivekanadam. "Artificial
2016 International Conference on Inventive
Intelligence Algorith m with SVM Classification
Computation Technologies (ICICT), 2016.
using Dermascopic Images for Melano ma
Diagnosis." Journal of Art ificial Intelligence and
[15] R. S. Ku mar, C.Ramesh, “Extreme Precipitation
Capsule Networks 3, no. 1 (2021): 34-42.
Events in Chennai Metro City Using Data M ining”,
International Journal of Innovative Tech and
[2] Bandyopadhyay S., Kundu S., Adhikary P, “Case
Exploring Engineering.
Study: Application of AI in Energy Efficient
HVAC System”, ARPN Journal.
[16] Ramlakhan K, Shang Y, “A Mobile Automated
Skin Lesion Classification System”, Proceedings of
[3] Celebi M E, Qaisar A, Fondón GI, Carmen S, Ma
23rd IEEE International Conference on Tools with
G, “The pattern classificat ion of dermoscopy
Artificial Intelligence.
images: A perceptually uniform model”.

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 1729

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on March 06,2024 at 04:27:55 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

[17] Ramlakhan K, Shang Y, “A Mobile Automated

Skin Lesion Classification System”, 23rd IEEE.

[18] Roy P. K., Mazu mdar A., Adhikary P.,”Study on

Fuzzy Logic based optimu m penstock design”,
ARPN Journal.

[19] S. Krishna B.V., Abinaya T., Jijin Godwin J.,

Tharanee Shree S., Sreen idhi B., “Detection o f
Leukemia and Its Types Using Co mb ination of
SVM and KNN A lgorith m”, 2021, Lecture Notes
in Network and Systems.

[20] S. R. Guha et al., "CNN Based Skin Lesion

Analysis for Classifying Melanoma," Published at:
The International Conference on Sustainable
Technologies for Industry 4.0, 2019.

[21] Sander L, Ioannis G, Nynke M, Michael B,

Jonkman Marcel F, Nicolai P , “M ED-NODE: a
computer-assisted melanoma diagnosis system
using non-dermoscopic images.”, 2015, An Expert
System Application.

[22] Vijayaku mar, Dr T. "SELECTIVE IMA GE

ENHA NCEM ENT AND RESTORATION FOR
SKIN CA NCER IDENTIFICATION." Journal o f
Innovative Image Processing 1, no. 1 (2019): 1-10.

[23] W. Setiawan, "Study: The Effect of Colo r

Enhancement on Early Detection of Skin Cancer
using Convolutional Neural Net work," Date Added
to IEEE Xplore: 11 May 2020, presented at-IEEE
International Conference on Informatics, IoT, and
Enabling Technologies, 2020.

[24] Y. Huo, “Study: A Full-Stack Application of Skin

Cancer Diagnosis Based on CNN Model”, IEEE -
International Conference of Co mputer Science,
2021, pp. 754-758/

[25] Yong-Yong Wang, Xiao-Li Wang, Yi-Tao Tian,

Benjamin Badami, Ni Zhang, Yi-Xin Cai,
“Diagnosis of skin cancer based on optimized
CNN”, ISSN 0933-3657.

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 1730

Authorized licensed use limited to: Andhra University College of Engineering. Downloaded on March 06,2024 at 04:27:55 UTC from IEEE Xplore. Restrictions apply.