See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/332114646

Handwritten Hindi Character Recognition using Deep Learning Techniques

Article  in  INTERNATIONAL JOURNAL OF COMPUTER SCIENCES AND ENGINEERING · February 2019

DOI: 10.26438/ijcse/v7i2.17

CITATIONS READS
5 1,193

2 authors, including:

Vijaya Kumar Reddy .R

Lakireddy Bali Reddy College of Engineering
22 PUBLICATIONS   105 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Vijaya Kumar Reddy .R on 12 June 2020.

The user has requested enhancement of the downloaded file.

 International Journal of Computer Sciences and Engineering Open Access
Research Paper Vol.-7, Issue-2, Feb 2019 E-ISSN: 2347-2693

Handwritten Hindi Character Recognition using Deep Learning

Techniques
R. Vijaya Kumar Reddy1*, U. Ravi Babu2
1
Dept. of CSE, Acharya Nagarjuna University, Guntur, India
2
DRK College of Engineering & Technology, Hyderabad, India

Corresponding Author: [email protected]

DOI: https://doi.org/10.26438/ijcse/v7i2.17 | Available online at: www.ijcseonline.org

Accepted: 12/Feb/2019, Published: 28/Feb/2019

Abstract - In this paper we present a handwritten Hindi character recognition system based on different Deep learning
technique. Handwritten character recognition plays an important role and is currently getting the attention of researchers
because of possible applications in assisting technology for blind and visually impaired users, human–robot interaction,
automatic data entry for business documents, etc. In this work, we propose a technique to recognize handwritten Hindi
characters using deep learning approaches like Convolutional Neural Network (CNN) With Optimizer RMSprop (Root Mean
Square Propagation) , Adaptive Moment (Adam) Estimation and Deep Feed Forward Neural Networks(DFFNN). The proposed
system has been trained on samples of large set of database images and tested on samples images from user defines data set and
from this experiment we achieved very high recognition results. Experimental results are compared with other neural network
based algorithm.

Keywords: DFFNN, CNN, Softmax classifier, RMSprop and Adam Estimation, Deep Learning.
which was not easily available at that time. Since then, the
I. INTRODUCTION researchers have drawn their attention to finding the
technique which needs less power by converting the images
Deep learning Techniques have been effectively applied to into feature vectors. In the last few decades, a lot of feature
various areas like image classification, speech recognition, extraction technique have been proposed such as HOG
Medical Images detection, face detection, satellite images, (histogram of oriented gradients) [4] and many others
recognizing traffic signs and pedestrian detection and so on. techniques are used as prominent feature extraction methods,
The outcome of deep learning techniques is also prominent, which have been experimented for many problems like
and in some cases the results are superior to human experts image recognition, character recognition, face detection, etc.
[1,2] in the past years. From last few years most of the Feature extraction [5] is one type of dimensionality
problems are also being re-experimented with deep learning reduction technique that represents the important parts of a
techniques with the view to achieving improvements in the large image into a feature vector. These features are
existing findings. Different architectures of deep learning handcrafted and clearly designed by the research
have been introduced in recent years, such as convolutional community. The robustness and performance of these
neural networks, deep networks, and recurrent neural features depend on the skill and the knowledge of each
networks. The entire architecture has shown the expertise in researcher. There are the cases where some vital features
different areas. Character recognition is one of the areas may be unseen by the researchers while extracting the
where machine learning techniques have been extensively features from the image and this may result in a high
experimented. The first deep learning technique, which is classification error. Deep learning inverts the process of
one of the leading machine learning techniques, was handcrafting and designing features for a particular problem
proposed for character recognition in 1998 on MNIST into an automatic process to compute the best features for
database [3]. that problem. A convolutional neural network has multiple
convolutional layers to extract the features automatically.
The deep learning techniques are basically composed of The features are extracted only once in most of the shallow
multiple hidden layers, and each hidden layer consists of learning models, but in the case of deep learning models,
multiple neurons, which compute the suitable weights for the multiple convolutional layers have been adopted to extract
deep network. A lot of computing power is needed to discriminating features multiple times. This is one of the
compute these weights, and a powerful system was needed, reasons that deep learning models are generally successful.
And also in Deep feed forward neural networks the features

© 2019, IJCSE All Rights Reserved 1

 International Journal of Computer Sciences and Engineering Vol.7(2), Feb 2019, E-ISSN: 2347-2693

are compute automatically by using different number of and used Euclidean similarity measure to classify the
hidden layer in it. numerals. KNN classifier obtained 90 % as recognition rate
whereas PCA obtained recognition rate of 84%. After the
The LeNet [3] is an example of deep convolutional neural comparison it is observed that KNN classifier has shown
network for character recognition. Recently, many other better results as compared to PCA classifier.
examples of deep learning models can be listed such as
AlexNet [2], ZFNet [6], VGGNet [7] and spatial transformer Sonu Varghese K et al. [14] proposed a Novel Tri-Stage
networks [8]. These models have been successfully applied Recognition Scheme for Handwritten Malayalam Character
for image classification and character recognition. Owing to Recognition. In the first stage we are grouping characters
their great success, many leading companies have also into different classes based on the number of corners,
introduced deep models. Google Corporation has made a bifurcations, loops and endings. In the second phase we are
GoogLeNet having 22 layers of convolutional and pooling identifying exact character in the class based on the different
layers alternatively. Apart from this model, Google has also feature extraction technique specially defined for each class.
developed an open source software library named In the third stage we are checking the probability of
Tensorflow to conduct deep learning research. Microsoft occurrence of the current character in the given position
also introduced its own deep convolutional neural network based on defined rules for the formation of words. we are
architecture named ResNet in 2015. ResNet has 152-layer implementing a three stage feature extraction technique
network architectures which made a new record in detection, which uses structural, statistical and moment variant features
localization, and classification. This model introduced a new of the character. Recognition conducted in different stages
idea of residual learning that makes the optimization of improves the efficiency, recognition rate and accuracy of the
different data. given system.

Character recognition is a field of image processing where Parshuram M. Kamble [15] proposed a method for
the image is recognized and converted into a machine- handwritten Marathi character recognition using R-HOG
readable format. As discussed above, the deep learning Feature. The system has been evaluated on a large amount of
techniques and especially convolutional neural networks handwritten Marathi characters. From the results as shown it
have been used for image detection and recognition. It has can be concluded that the use of R-HOG based feature
also been successfully applied on Roman (MNIST) [3], extraction method and FFANN based classification will be
Chinese [9], Bangla [10] and Arabic [11] languages. In this more effective with increased processing speed and
work, a convolutional neural network and Deep Feed accuracy.
Forward neural network is applied for handwritten Hindi
characters recognition. Verma [3] used radial basis function and multilayer
perceptron neural networks for recognizing the handwritten
Rest of the paper is organized as follows; Section 2 contains characters of Devanagari script. Back propagation error
related work. The proposed algorithm is presented in Section algorithm is also used to improve the recognition rate. In this
3. The experimental details and results obtained are proposed system, they compare the results obtained from
presented in Section 4. Section 5 contains the conclusion radial basis Function (RBF) networks and multi-layer
part. perceptron. Dataset consists of 245 samples written by five
different users. The results so obtained show that multilayer
II. RELATED WORK perceptron (MLP) networks performs better than that of
radial basis functions. But MLP networks training time is
Amit Choudhary et al. [12] proposed an Off-Line more as compared to radial basis function networks. Highest
Handwritten Character Recognition using Features Extracted recognition rate so obtained using radial basis function and
from Binarization Technique. This work is to extract multilayer perceptron (MLP) networks are 85.01% and
features obtained by Binarization technique for recognition 70.8% respectively.
of handwritten characters of English language. The
recognition of handwritten character images have been done III. PROPOSED APPROACH
by using multi-layered feed forward artificial neural network
as a classifier. This algorithm delivers outstanding The proposed method is mainly consists of 4 phases steps. In
classification accuracy of 85.62 %. the primary phase, collecting the characters data from kaggle
dataset and gathering images from different users. After
Baheti M. J et al. [13] proposed a comparison of the offline collecting the character data of gray scale images will be
handwritten character recognition system for the isolated pre-processing by checking null and missing values. Using
Gujarati numerals. They used affine invariant moments the normalization techniques to convert the gray level values
based model for the feature extraction. They used KNN to range of 0 to 1 values, and then labelling Hindi characters
classifier and PCA to reduce dimensions of feature space from 0 to 35 with one hot coding which will generate vector

© 2019, IJCSE All Rights Reserved 2

 International Journal of Computer Sciences and Engineering Vol.7(2), Feb 2019, E-ISSN: 2347-2693

form of data. In the 3rd phase, we extracted the features Many attempts, such as decreasing an adaptive learning rate,
atomically from Different Deep learning algorithm like adopting a big batch size, incorporating a temporal
Convolutional Neural Network (CNN) and Deep Feed decorrelation technique, seeking an analogous surrogate,
Forward Neural Networks (DFFNN) for recognition of etc., In addition, we illustrate that Adam is essentially a
handwritten character system. In finally phase we applied an specifically weighted AdaGrad with exponential moving
optimization technique like Optimizer RMSprop (Root average momentum, which provides a novel perspective for
Mean Square Propagation), Adaptive Moment (Adam) understanding Adam and RMSProp. At last, we validate the
Estimation to get very promising results. The proposed sufficient condition by applying Adam and RMSProp to
method block diagram is shown below in Figure 1. tackle the counterexamples and train deep neural networks.

B. RMSprop Optimizer
The RMSprop (Root Mean Square Propagation) optimizer is
similar to the gradient descent algorithm with momentum.
The RMSprop optimizer restricts the oscillations in the
vertical direction. Therefore, we can increase our learning
rate and our algorithm could take larger steps in the
horizontal direction converging faster. It has shown success
for training Recurrent Models.
(1)

Figure 1: Block diagram proposed handwritten digit

recognition system (2)
√( )

A. Convolutional neural network ss - the exponentially weighted average of past squares of

Computer Vision and pattern recognition is a major growing gradients
field in area of image processing. In that Convolutional - cost gradient with respect to current layer weight tensor
Neural Network (CNNs) plays major role in computer
W - weight tensor
vision. CNNs is working on many applications in
β - hyperparameter to be tuned
Image Classification and it is the core of most Computer
α - the learning rate
Vision and pattern recognition systems today, from
ϵ - very small value to avoid dividing by zero
automatic tagging of photo in Facebooks to self-driving
RMSprop as well divides the learning rate by an
cars, recognizes digits, alpha-numerals, traffic signal boards,
exponentially decaying average of squared gradients.
and the other object class[7]. We used five layered
Convolutional Neural Networks (CNN) model. On them one
layers for convolutional, one layers for max pooling or sub C. Adam Estimation Optimizer
Adaptive Moment (Adam) Estimation is another method that
sampling, one Flatten layer which converts 2D array into 1D
computes adaptive learning rates for each parameter. Adam,
array and finally two fully connected layers for
an algorithm for first-order gradient-based optimization of
classification. The initial layer is convolutional (Conv2D)
stochastic objective functions, based on adaptive estimates
layer has 32 output mapping and the next max pooling layer
of lower-order moments. The method is straightforward to
has 14 output mapping.
implement, is computationally efficient, has little memory
requirements, is invariant to diagonal rescaling of the
gradients, and is well suited for problems that are large in
terms of data and/or parameters.

(3)
(4)
mt and vt are estimates of the first moment (the mean) and
the second moment (the uncentered variance) of the
gradients respectively, hence the name of the method.
As mt and vt are initialized as vectors of 0's, the authors of
Figure 2: The overall structural design of the CNN Model of our Adam observe that they are biased towards zero, especially
proposed system with different layer. during the initial time steps, and especially when the decay
Adam and RMSProp, as two of the most influential adaptive rates are small (i.e. β1 and β2 are close to 1). They counteract
stochastic algorithms for training deep neural networks.

© 2019, IJCSE All Rights Reserved 3

 International Journal of Computer Sciences and Engineering Vol.7(2), Feb 2019, E-ISSN: 2347-2693

these biases by computing bias-corrected first and second

moment estimates:
̂ (5)
̂ (6)
They then use these to update the parameters just as we have
seen in Adadelta and RMSprop, which yields the Adam
update rule:
̂ (7)
√

D. Deep Feed-Forward Neural Network

This neural network is known as feed-forward networks, this Figure 3: The overall structural design of the DFFNN Model
network is simplest to analyze. The feed-forward Neural of our proposed system with different layer.
Network input layer contains n-dimensional vector as input
to the network and contains L-1 hidden layers as middle IV. RESULTS AND DISCUSSION
layers mostly two hidden layers are used and may increase
based upon the requirement. Finally there is one output layer In general Feed forward Neural Network consists of
containing k number of output classes. Each neuron in the different hidden layer. In most of FFNN will have two
hidden layer and output layer can be split into two parts: pre- hidden layers with 16 or 32 neurons and more, Hidden layers
activation and activation. In Deep feed-forward network is are multiplied with different random weight of image pixel
the hidden layer are taken of size m- dimensional vector size data which is between 0 to 1. But in Deep Feed forward
which quite larger than the normal hidden layers of feed- Neural Network was design with same two hidden layers
forward network, so that the hidden layer can generated a and each hidden layers consists of large set of neurons i.e.
huge network of features. From this huge number of features we used 512 neurons are taken and this are multiplied with
we can extract right match for predicated classes [17][18]. random weights. By using this deep network we got
A deep feed-forward network is used to extract features of promising results. Here we displaying some labeling of
image automatic to recognition the classes of testing image characters in Table 1.
of the data set. Figure 3 is shown below how the DFFNN
used in this work.

Table 1: Sample data representation of labeling.

Layer Type layer Operation No of feature Maps feature maps Size window Size Total parameters

C1 Conv2D 32 32×32 5×5 832

MP1 Max-pooling2 32 32×32 2×2 0
FL1 Flatten 6272 1×1 N/A 0
F1 Fully connected 256 1×1 N/A 1605888
F2 Fully connected 36 1×1 N/A 9509

In this testing phase we used five layered Convolutional 2D array into 1D array and finally two fully connected
Neural Networks (CNN) model with Adam Optimization. layers for classification. All the parameters with respect to
On them one layers for convolutional, one layers for max the corresponding layers are stated in Table 2.
pooling or sub sampling, one Flatten layer which converts
Table 2. Parameters setup for CNN with Adam Optimization.
Layer Type layer Operation No of feature Maps feature maps Size window Size Total parameters

C1 Conv2D 32 32×32 5×5 832

MP1 Max-pooling2 32 32×32 2×2 0
FL1 Flatten 8192 1×1 N/A 0
F1 Fully connected 256 1×1 N/A 2097408
F2 Fully connected 36 1×1 N/A 9252

In this testing phase we used five layered Convolutional Optimization. On them one layers for convolutional, one
Neural Networks (CNN) model with RMSProp layers for max pooling or sub sampling, one Flatten layer

© 2019, IJCSE All Rights Reserved 4

 International Journal of Computer Sciences and Engineering Vol.7(2), Feb 2019, E-ISSN: 2347-2693

which converts 2D array into 1D array and finally two fully 28 waw 378 352 26 93.12
connected layers for classification. All the parameters with 418 404 14 96.65
respect to the corresponding layers are stated in Table 3. 29 mot
30 pet 422 406 16 96.21
Table 3. Parameters setup for CNN with Adam 31 pat 391 368 23 94.12
Optimization. 383 354 29 92.43
32 ha
bha ra dhaa dha ta
33 chhya 430 418 12 97.21
34 tra 378 360 18 95.24
35 gya 382 360 22 94.24
Average recognition percentage 95.57
Table 5: Individual results of the test database using CNN-Adam Estimation
Table 4: Individual results of the test database DFFNN Not
Labele No. Correctl Correctl
Label No. Not
d Hindi of y y %
ed Hindi of Correctly Correctly %
numbe Charact imag Classifi Classifi Accura
numb Charact image Classifie Classifie Accu
r er es ed ed cy
er er s d d racy
396 383 13 96.72 0 Ka 396 383 13 96.72
0 Ka
413 401 12 97.09 1 kha 413 401 12 97.09
1 kha
2 ga 418 403 15 96.41 2 ga 408 403 5 98.77
3 gha 402 387 15 96.27 3 gha 402 387 15 96.27
4 kna 368 352 16 95.65 4 kna 368 352 16 95.65
5 cha 396 370 26 93.43 5 cha 396 370 26 93.43
6 chha 394 374 20 94.92 6 chha 394 374 20 94.92
7 ja 427 413 14 96.72 7 ja 417 413 4 99.04
8 jha 389 370 19 95.12 8 jha 389 370 19 95.12
9 yna 430 415 15 96.51 9 yna 420 415 5 98.81
10 taa 438 421 17 96.12 10 taa 428 421 7 98.36
11 thaa 405 389 16 96.05 11 thaa 405 389 16 96.05
12 daa 413 400 13 96.85 12 daa 403 400 3 99.26
13 dhaa 402 390 12 97.01 13 dhaa 402 390 12 97.01
14 adna 430 414 16 96.28 14 adna 420 414 6 98.57
15 ta 395 384 11 97.22 15 ta 395 384 11 97.22
16 tha 380 358 22 94.21 16 tha 380 358 22 94.21
17 da 409 394 15 96.33 17 da 409 394 15 96.33
18 dha 378 353 25 93.39
18 dha 378 353 25 93.39
19 na 405 383 22 94.57
19 na 405 383 22 94.57
20 pa 393 370 23 94.15
20 pa 393 370 23 94.15
21 pha 407 393 14 96.56
21 pha 407 393 14 96.56
22 ba 395 379 16 95.95
22 ba 395 379 16 95.95
23 bha 401 372 29 92.77
23 bha 401 372 29 92.77
24 ma 398 387 11 97.24
24 ma 398 387 11 97.24
25 yaw 410 387 23 94.39
25 yaw 410 387 23 94.39
26 ra 421 409 12 97.15
26 ra 421 409 12 97.15
27 la 405 390 15 96.30

© 2019, IJCSE All Rights Reserved 5

International Journal of Computer Sciences and Engineering Vol.7(2), Feb 2019, E-ISSN: 2347-2693

27 la 395 390 5 98.73 25 yaw 408 385 23 94.36

28 waw 378 352 26 93.12 26 ra 407 395 12 97.05
29 mot 408 404 4 99.02 27 la 421 416 5 98.81
30 pet 412 406 6 98.54 28 waw 411 407 4 99.03
31 pat 391 368 23 94.12 29 mot 411 395 16 96.11
32 ha 383 354 29 92.43 30 pet 377 370 7 98.14
33 chhya 420 418 2 99.52 31 pat 406 393 13 96.80
34 tra 378 360 18 95.24 32 ha 384 365 19 95.05
35 gya 382 360 22 94.24 33 chhya 402 384 18 95.52
Average recognition percentage 96.02 34 tra 416 407 9 97.84
35 gya 375 360 15 96.00
Table 6: Individual results of the test database using CNN-RMSProp
Average recognition percentage 97.33
Lab
ele
d
nu Hindi No. of Correctly Not %
mb Charact image Classifie Correctly Accu
er er s d Classified racy
0 Ka 430 427 3 99.30
1 kha 396 392 4 98.99
2 ga 433 426 7 98.38
3 gha 401 386 15 96.26
4 kna 391 375 16 95.91
5 cha 405 397 8 98.02
6 chha 369 349 20 94.58
7 ja 365 355 10 97.26
Figure 4: Graphical representation of the % recognition of the DFFNN and
8 jha 419 410 9 97.85 other existing method
9 yna 392 387 5 98.72
10 taa 387 384 3 99.22
11 thaa 401 395 6 98.50
12 daa 405 402 3 99.26
13 dhaa 390 378 12 96.92
14 adna 406 400 6 98.52
15 ta 394 393 1 99.75
16 tha 395 373 22 94.43
17 da 420 415 5 98.81
18 dha 401 386 15 96.26
19 na 420 408 12 97.14
20 pa 399 386 13 96.74
21 pha 410 396 14 96.59
Figure 5: Graphical representation of the % recognition of the
22 ba 360 354 6 98.33
CNN-Adam method and other existing method
23 bha 406 391 15 96.31
24 ma 387 376 11 97.16

© 2019, IJCSE All Rights Reserved 6

 International Journal of Computer Sciences and Engineering Vol.7(2), Feb 2019, E-ISSN: 2347-2693

[9] Cire¸san, D.; Ueli, M. “Multi-column deep neural networks for

offline handwritten Chinese characterclassification”, In
Proceedings of the 2015 International Joint Conference on Neural
Networks (IJCNN),Killarney, Ireland, 12–17 July 2015.
[10] . Sarkhel, R.; Das, N.; Das, A.; Kundu, M.; Nasipuri, M. “A
Multi-scale Deep Quad Tree Based Feature Extraction Method for
the Recognition of Isolated Handwritten Characters of popular
Indic Scripts”, Pattern Recognition.2017, 71, 78–93.
[11] . Ahranjany, S.S.; Razzazi, F.; Ghassemian, M.H. “A very high
accuracy handwritten character recognition system for
Farsi/Arabic digits using Convolutional Neural Networks”, In
Proceedings of the 2010 IEEE Fifth International Conference on
Bio-Inspired Computing: Theories and Applications (BIC-TA),
Changsha,China, 23–26 September 2010.
[12] Choudhary, A., Rishi, R., and Ahlawat, S., “Off-Line
Handwritten Character Recognition using Features Extracted from
Figure 6: Graphical representation of the % recognition of the CNN- Binarization Technique”, AASRI Conference on Intelligent
RMSProp method and other existing method Systems and Control, 2013, pp. 306-312.
[13] Baheti M. J., Kale K.V., Jadhav M.E., “Comparison of Classifiers
V. CONCLUSION for Gujarati Numeral Recognition”, International Journal of
Machine Intelligence, Vol. 3, Issue 3, pp. 160-163, 2011.
[14] Sonu Varghese K, Ajay James, Dr.Saravanan Chandran , “A
In this paper we proposed different neural network approach for Novel Tri-Stage Recognition Scheme for Handwritten Malayalam
Recognition of Handwritten Hindi characters. We evaluated the Character Recognition”, International Conference on Emerging
performance using Convolutional Neural Network (CNNs) with Trends in Engineering, Science and Technology (ICETEST -
optimization techniques and Deep Feed Forward Neural 2015),2016 pp-1333-1340.
Network. These techniques are train and test on a standard user [15] Parshuram M. Kamble, Ravinda S. Hegadi, “Handwritten Marathi
define dataset which is collect from different users. From character recognition using R-HOG Feature”, International
experimental results, it is observed that DFFNN, CCN-Adam Conference on Advanced Computing Technologies and
and CNN-RMSprop yield the best accuracy for Handwritten Applications (ICACTA- 2015),2015 pp- 266 – 274.
[16] B. K. Verma, "Handwritten Hindi character recognition using
Hindi characters compared to the alternative techniques. We multilayer perceptron and radial basis function neural networks in
achieved promising results from proposed method with high Neural Networks”, 1995. Proceedings, IEEE International
accuracy rate. Conference on. vol. 4, pp. 2111-2115, 1995.
[17] Hailiang Ye, Feilong Cao, Dianhui Wang, Hong Li , “Building
REFERENCES feed forward neural networks with random weights for large scale
datasets”, Expert Systems with Applications Volume 106, 2018,
pp. 233-243
[1] Ciregan, D.; Meier, U.; Schmidhuber, J, “Multi-column deep
[18] Feilong Cao, Dianhui Wang, Houying Zhu, Yuguang Wang, “An
neural networks for image classification”, In Proceedings of the
iterative learning algorithm for feedforward neural networks with
IEEE Conference on Computer Vision and Pattern Recognition
random weights”, Information Sciences, Volume 328, 2016, pp.
(CVPR), Providence, RI,USA, 16–21 June 2012.
546-557.
[2] . Krizhevsky, A.; Sutskever, I.; Hinton, G.E, “Imagenet
classification with deep convolutional neural networks” .In
Proceedings of the Advances in Neural Information Processing Authors Profile
Systems, Lake Tahoe, NV, USA,3–8 December 2012. R. Vijaya Kumar Reddy Research Scholar, Dept. of. CSE,
[3] Lecun, Y.; Bottou, L.; Bengio, Y.; Haffner P, “Gradient-based Acharya Nagarjuna University, Guntur. He has published more
learning applied to document recognition”, Proc. IEEE 1998, 86, than 10 research paper in International journals. He as 8 years
2278–2324. of teaching experience and 4 years of Research Experience. He
[4] Navneet, D.; Triggs, B. “Histograms of oriented gradients for
human detection”, In Proceedings of the CVPR2005 IEEE
is a life member of ISSE and ISTE.
Computer Society Conference on Computer Vision and Pattern
Recognition, San Diego, CA, USA,20–25 June 2005; Volume 1 Dr. U Ravi Babu obtained his Ph.D in computer science
[5] Wang, X.; Paliwal, K.K. “Feature extraction and dimensionality engineering from Acharya Nagarjuna University. He completed
reduction algorithms and their applications in vowel recognition”, MSc Information Systems (IS) from AKRG PG College,
Pattern Recognit. 2003, 36, 2429–2439. Andhra University in the year 2003 and M. Tech in Computer
[6] Zeiler, M.D.; Rob, F. “Visualizing and understanding Science & Engg from RVD University in the year 2005. He is a
convolutional networks”, In Proceedings of the member of SRRFGIET, Rajahmundry. He has published 30
EuropeanConference on Computer Vision, Zurich, Switzerland,
research papers in various Inter National journals and 10 papers
6–12 September 2014.
[7] Simonyan, K.; Andrew, Z. “Very deep convolutional networks for in various National and International conference proceedings.
large-scale image recognition”. arXiv, 2004. He is working as a Principal, DRK College of Engineering &
[8] Jaderberg, M.; Simonyan, K.; Zisserman, A, “Spatial transformer Technology Hyderabad from 2017 to till date. He is a life
networks”, In Proceedings of the Advances inNeural Information member of ISCA, IAENG and IRS.
Processing Systems, Montreal, QC, Canada, 11–12 December
2015.

© 2019, IJCSE All Rights Reserved 7

View publication stats