Wireless Personal Communications (2022) 122:349–378

https://doi.org/10.1007/s11277-021-08903-4

Handwritten Devanagari Character Recognition Using

Modified Lenet and Alexnet Convolution Neural Networks

Duddela Sai Prashanth1,2 · R. Vasanth Kumar Mehta1 · Kadiyala Ramana3 ·

Vidhyacharan Bhaskar4 

Accepted: 8 August 2021 / Published online: 24 September 2021

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2021

Abstract
Despite many advances, Handwritten Devanagari Character Recognition (HDCR) remains
unsolved due to the presence of complex characters. For HDCR, the traditional feature
extraction and classification techniques are limited to the datasets developed in the respec-
tive laboratory that are not available publicly. A standard benchmarking dataset is not avail-
able for HDCR that helps to develop deep learning models. To progress the performance
of HDCR, in this study, we produced a dataset of 38,750 images of Devanagari numerals,
and vowels are generated and made publicly available for fellow researchers in this domain.
This data is collected from more than 3000 subjects of different age groups. Each character
is extracted by a segmentation technique proposed here, which is limited to this applica-
tion. Experiments are conducted on the dataset; three different Convolution Neural Net-
works (CNN) architecture is developed. 1. CNN, 2. Modified Lenet CNN (MLCNN) and
3. Alexnet CNN (ACNN). A Modified LCNN is proposed by changing the architecture of
Lenet 5 CNN. Regular Lenet 5 has tanh(x) as its activation function. Since the Devangari
characters are nonlinear, non-linearity is introduced in the Networks by using Rectified Lin-
ear Unit. This solves the problem of vanishing gradient problem by tanh(x) . We achieved
a recognition rate of 96% on training data and 94% on unseen data using CNN. MLCNN
obtained an accuracy rate of 99% and 94% with less computational cost. Whereas, ACNN
attained a recognition rate of 99% and 98% on unseen data. A series of experiments were
conducted on the data with different combination splits of data and found a minimum loss
of 0.001%. Such developments fill a significant percentage of the huge gap between real-
world requirements and the actual performance of Devanagari recognizers.

Keywords  Devanagari character recognition · Convolution neural network · Computer

vision · Pattern recognition

* Vidhyacharan Bhaskar
vcharan@gmail.com
Extended author information available on the last page of the article

13
Vol.:(0123456789)
 350 D. S. Prashanth et al.

1 Introduction

In the digital world, the demand for accessing multimedia is emerging. It involves a pro-
cess of adapting the offline entities into the digital domain. The analysis of the scanned
documents deals with the images. In real-time, character recognition on those images is a
complex process. The automation of character segmentation and recognition of characters
in real-time is a complex process—it includes character segmentation and recognition of
characters with the features extracted [1]. Optical Character Recognition (OCR) is catego-
rized into two methods, i.e., online and offline.
Character recognition is easier for the documents with Latin script with the form of
the English language. A lot of research is carried for handwritten character recognition
in various scenarios like English, Chinese, and Arabic scripts [2]. In any case, for Indian
Languages, OCR framework work is as yet slacking when compared with others due to its
complicated structure and computations. Devanagari script is base for many languages like
Hindi, Marathi, Sanskrit, and many more [3]. It contains 10 numerals, 13 vowels, and 33
consonants. In this research, majorly concentrate on the recognition of offline handwritten
Devanagari numeral and vowels.
Offline handwritten Devanagari character recognition (HDCR) is an emerging research
field over decades because of the challenges of variant character classes. Despite many
advances, HDCR remains unsolved due to the presence of complex characters. Most of the
traditional recognizers have not to lead better performance, or it is limited to their data-
set, because of not having big datasets as a benchmark. The variety of handwriting styles
makes it more difficult due to the similarities between the characters [4]. The progress in
this research is improved recognition accuracies with the help of traditional approaches.
Though, the number of studies increased—crafted features and robust classifiers still rec-
ognition accuracy is far behind the human ability. The reason is the lack of benchmark
datasets [5]. Many researchers have developed their dataset in their laboratory and clas-
sifiers developed, accuracies are limited to the respective dataset, and most of the data-
sets are not available publicly [6]. Implementing HDCR systems is much difficult than
printed characters, especially for the Indian scripts [7]. The algorithms used for the printed
Devanagari characters can often be used on the handwritten characters. Most of the time,
it fails to work on the handwritten characters because of the variation in writing style and
sizes of different writers.
Most of the OCR techniques developed on HDCR includes a time-consuming pre-pro-
cessing step. After noise removal and character segmentation, a process that extracts the
header line and separates the upper part and lower part of the line for every character. Clas-
sifiers are built on the basis of features extracted from these two parts of the character.
The primary objective of this study is to develop a CNN-based model with a lower com-
putational complexity and memory space while maintaining the required accuracy. As a
result, the paper is divided into two parts: (a) We investigate and compare/analyze several
different state-of-the-art CNN models; and (b) We develop a new CNN model, which we
refer to as the "Modified LeNet Convolutional Neural Network (MLCNN)," that requires
less time and memory space than existing CNN models. On some publicly available data-
sets, “MLCNN” is compared to other popular CNN models for validation purposes.
In this paper, a new dataset is presented for Devanagari numerals and vowels. A total
of 38,750 images are collected from 2400 subjects of different age groups. This dataset is
publicly available [8]. A segmentation algorithm is proposed for the dataset collected, and
Convolution Neural Network (CNN) architectures are proposed for the HDCR. First, CNN

13
Handwritten Devanagari Character Recognition Using Modified… 351

obtained 96% accuracy and 94% for unseen data. Secondly, LeNet CNN architecture is
modified by introducing max-pooling instead of average pooling to reduce the parameters
of fully connected layers then added non-linearity by using Rectified Linear Unit. Modi-
fied LeNet Convolution Neural Network (MLCNN) obtained an accuracy of 99% accuracy
on Training data and 94% on unseen data. Thirdly, AlexNet CNN (ACNN) architecture
achieved an accuracy of 99% and 98% on the unseen dataset and made a detailed compari-
son between these architectures by separating the data into different ratios of training and
test (i.e., 80:20, 70:30, 60:40) with varying epochs of count (i.e., 20, 30, 40 and 50 epochs).
The rest of this paper is organized as follows. Section 2 reviews the related works about
Offline HDCR. Section  3 provides the method: the basic theory of CNN, ACNN, and
MLCNN and details about the proposed architecture. Section  4 presents the experiment
results and discussion that includes the experimental data, settings for the experiment, and
training strategy along with the comparison of the results. Finally, the conclusion is given
in Sect. 5.

2 Related Work

Offline Handwritten Devanagari Character Recognition (HDCR) has received intense

attention from the past few decades. The traditional OCR method includes three stages:
pre-processing, feature extraction, and classification. Pre-processing consists of the steps
of size normalization and resampling [6]. The variations in the input may adversely impact
recognition. Features characterize the best representation of the shape of a character.
Character segmentation is classified into word, line, and character segmentation [1, 42].
It is more accessible to segment printed characters when compared to handwritten char-
acters. There are many standard segmentation techniques developed for printed text. For
HDCR, segmentation-based classifiers are popular before pattern-based neural implemen-
tation. Different approaches of segmentation as per the dataset is implemented.
Feature extraction techniques developed are of shape-based or non-shape based [10,
41]. These techniques generate a feature vector. These vectors are used as input for any
classifiers, unlike deep learning methods, feed a pre-processed image as input along with
its label for supervised learning. Hybrid features—a combination of shape-based and non-
shape-based features classify the input image more accurately. Still, the automatic features
generated by CNN’s are better.
In the recent review article [11], structured based and statistical features and their accu-
racies are compared. Chain code and gradient features achieved an accuracy of 98.51% by
the SVM classifier for the printed Devanagari characters [12]. Few Structural features with
different classifiers like syntactic pattern analysis-based classifier [13] (90% accuracy),
Binary tree classifier [14] (95% accuracy), Distance-based classifier [15] (93% accuracy)
for printed text. Different feature extraction techniques implemented for HDCR, quadratic
based classifier on the histogram of directional chain code features [16] achieved 98.86%
accuracy with the dataset size of 11,270 handwritten Devanagari Characters. The most
commonly used classifiers are MLP based classifiers [17–19], SVM [18, 20, 21] Neural
Networks [22, 23] or Modified Neural Networks for feature-based classification.
There is no standard dataset that is implemented nor used for Devanagari characters,
unlike MNIST for Latin numbers. The dataset size used to develop feature-based classifiers
vary. Few datasets developed are 11,270 [16] and achieved an accuracy of 98.86%, 25,000
[18] characters with an accuracy of 90.116% by SVM and 87.56% by MLP classifier.

13
 352 D. S. Prashanth et al.

Zernike Moment based feature [21] on 27,000 handwritten characters with an accuracy of
98.37% by SVM.
Deep learning models has increased its acceptance in the field of Computer Vision and
Pattern Recognition due to CNN [24]. CNN includes deep layers of convolutions, subsam-
pling layers, feature extraction over pooling layers, stride, and padding operations to con-
trol the size of image representation. It also comprises different activation functions like
sigmoid, softmax, to introduce non-linearity—Rectified Linear units, and more. To avoid
overfitting, many regularization techniques are included in CNN’s. The classification meth-
ods are categorized into Kernal based, statistical based and Neural network based [25]. In
[37], different language scripts numerals data is used and classified using CNN and able to
achieve 99% accuracy. The numerals dataset used in [37] is Arabic (MADBase), MNIST,
and Persian. In [38], a comparison of CNN-based classification and feature-based classi-
fication is presented. The maximum accuracy able to reach by HoG based features and
MLP classifier is 82.66% where as using CNN, 94.91% accuracy, and 96.09% accuracy
by Bidirectional Long Short Term Memory (BLSTM). In [39, 40], CNN with linear func-
tion—tanh(x) , accuracy achieved is 93.8% and 85.1%. When nonlinearity is introduced to
CNN 96.9% accuracy is achieved.
This brief inre survey reveals that there is a lack of standard datasets for Devanagari
Scripts. With larger datasets, CNN models can be designed with high accuracies than the
classification techniques. Hence, in the present work, we present a dataset and developed
a unique segmentation algorithm that is limited to this collected data. Three types of CNN
models are developed and compared, CNN, ACNN, and MLCNN.

3 Proposed System

3.1 Preprocessing

The traditional method of recognizing individual characters consists of three steps, pre-
processing, feature extraction, and classification. In deep learning models, features are
extracted by the Neural model automatically [43]. There is no exclusive process involved,
but a few techniques are required in pre-processing. The below-mentioned steps are imple-
mented on the data collected by different subjects, as shown in Figs. 1, and 2. that extracts
the images into a folder. Those images are manually identified and separated into respec-
tive folders for classification.

13
Handwritten Devanagari Character Recognition Using Modified… 353

Fig. 1  Sample sheet used to collect Devanagari numeral data

13
 354 D. S. Prashanth et al.

Fig. 2  Sample sheet used to collect Devanagari vowel data

13
Handwritten Devanagari Character Recognition Using Modified… 355

The size of each original character varies from one image to another. Since the bounding
box is plotted on the recognized character and crops it by the original coordinates. Each
subject writes in their own ways leads to the variation of the size of characters on the paper.
All the characters extracted are resized to 28 × 28. Following the standards of one of the
most standard and benchmark datasets is MNIST for numerals, the size of the character is
chosen as 28 × 28.

13
 356 D. S. Prashanth et al.

3.2 Deep Learning Models

From the past few decades, research in Deep Learning models (DLM) has delivered remark-
able outcomes in the field of computer vision, machine learning, and pattern recognition.
DLM’s are more efficient because of their deep architecture. Unlike any conventional classifi-
ers, DLM’s accept raw or pre-processed images as inputs. These models do not entail any par-
ticular feature extraction. The deep layers, extract features from the input at the lower and mid-
dle levels and high-level layers perform classification. These models integrate everything into
a single network that can be optimized as per the objective function. These type of integrated
models results better than traditional classifiers. The following concepts are being utilized in
this research to build such models.

3.3 Convolutions

Convolutions were employed for automatic feature extraction [26]. For an image I in size of
(p, q) , the convolution is defined as
∑∑
C(p, q) = (I ∗ K)(p, q) = I(p − m, q − n)K(m, n) (3)
k l

where K is the kernel is the size of (m, n) . Convolution is a mathematical function to
describe the process of combining two functions to produce a third function. The output is
called a feature map, and by using a filter or kernel to the input image, extract features for
training.

3.4 Dropouts and L2 Regularization

Devanagari vowels and numerals are highly sensitive and complex. Characters in Table 5 out
of 13 different vowels six vowels are identical with minor changes. Similarly, numerals in
Table 6 out of 10 numerals, two couple of numerals, and three numerals are identical with
changes. This Features extracted vary with small differences. It leads to over tune of weights
to our dataset used for training that may lead to the problem of overfitting in the data, and
classifier performs poorly on the unseen data. Regularization methods reduce overfitting [27].
Different regularization methods are L1 & L2 regularization, cross-validation, early stopping,
drop out, dataset augmentation. In this research, dropout is adopted and well performed on the
data.
Dropout refers to dropping nodes to reduce overfitting. This regularization ignores some
neurons in the network during forwarding and backward propagation. This approach helps in
reducing interdependent learning among the neurons. CNN extracts the most robust features.
In dropout a parameter ‘ p ’ that sets the probability of which nodes are kept or (1 − p) for
those that are dropped.
( n
)2 n
1
(4)
∑ ∑
pi 1 − pi w2i Ii2
( )
E= t− pi wi Ii +
2 i=1 i=1

L2 regularization is also called Ridge regression. It helps to penalize weights. Any

larger weights are gradients with abrupt changes in the model are manifest with a decision
boundary by penalizing them it becomes smaller. 𝜆 controls the degree of penality.

13
Handwritten Devanagari Character Recognition Using Modified… 357

1( )2 𝜆 2
E= targetO1 − outO1 + w (5)
2 2 i

3.5 Subsampling

Subsampling reduces the parameters or the dimensions of the feature map. It helps to dis-
card the spatial information. For input to the further layers in CNN, the parameters are
reduced by taking a collection of neurons. In this research, max pooling is used by taking
the whole image and is partitioned into rectangles. From each box, a maximum value is
taken and forms a new box [28]. Only one value is considered from a group of values. It
helps to avoid further overfitting of data.
( )
(6)
n n−1
X(p,i,j) =f max X(p,iS +u,jS +v)
0≤u,v≤Mn −1 n n

3.6 Activation Functions

The role of activation functions is crucial in CNN to segregate the useful and not much
helpful information from the neurons. Non-linearity is introduced into the network using
activation functions. Over the choices of activation functions like logistic function, tanh(x)
function, arctan function, and more, these functions lead to vanishing the gradients [4].
Since the input to the neurons is a range of 0, whereas the gradients are significant values.
Rectified Linear Unit (ReLU) is used to overcome this problem.
ReLU(x) = max(x, 0) (7)

4 Optimizer

4.1 Stochastic Gradient Descent (SGD)

SGD optimizers update the parameter for every training example and its label [30]. Lets
xi is the input and its labelsyi . J(𝜃;xi ;yi ) in the equation represents the objective function.
Gradient Descent is a way to minimize the cost. J(𝜃;xi ;yi ) is parameterized by the model’s
parameters 𝜃 ∈ ℝd . 𝜃 represents the weights, that gets updated for every iteration. 𝜂 repre-
sents the learning rate that reflects the size of the steps that take to reach a local minimum.
𝜃 = 𝜃 − 𝜂.∇𝜃 J(𝜃;xi ;yi ) (8)
Batch Gradient Descent (BGD) performs excess calculations for massive datasets, as it
recomputes gradients for similar examples before every parameter gets updated. SGD gets
rid of this repetition by performing each update at a time. It is in this manner typically a lot
quicker and can likewise be utilized to learn online.
SGD performs updates frequently with a large difference that causes the target function
to change vigorously as in image. SGD enables a better local minimum than BGD.

13
 358 D. S. Prashanth et al.

4.1.1 Adadelta

Adagrad algorithm adapts the lower learning rate of the parameters with lower learning
rates where features occur frequently and the higher learning rate for the parameters
with infrequent features. Adelta is an extension of Adagrad that optimizes the learning
rate. Rather than aggregating all past squared gradients, it limits the window of col-
lected past gradients to some fixed size w

E g2 t = 𝛾E g2 t−1 + (1 − 𝛾)g2t (9)

[ ] [ ]

where E g t is moving average, t is the time step, and 𝛾 is the momentum. Here, 𝛾 value is
[ 2]

set to 0.9.

4.2 Fully Connected Layers

All the nodes in one layer are connected to the output of the other layers is the Fully
Connected (FC) layer. The FC layer outputs the class probabilities, where each class
is assigned a probability number. All probabilities must sum to 1 . Softmax functions
produce the probability values for all the categories. All the outputs of a fully connected
layer are probabilities. Applying softmax squashes the real value numbers into prob-
abilities that sum to 1 . Softmax comprises the output range of (0,1)
exp(xi )
softmax(x)i = ∑n (10)
j=1 exp(xj )

4.3 Need of ReLU (Non‑linearity) and Modification in LeNet 5

Non-linear functions are required for universal approximation. Adding more layers into
the neural network doesn’t increase the approximation power. Nonlinearity in the deci-
sion function increases with the ReLU activation function. As per the universal approxi-
mation algorithm, using only linear activation function in the Neural Networks is less
valid. When more layers are added into the network, ReLU is faster to train than CNN
with tanh(x) . The reason to make the network train slower is—Gradient Descent based
optimizers are used for weight optimization and to achieve the local minima. Use of
tanh(x) may lead to vanish the Gradient Descent of the network. Whereas in the tanh(x) ,
x is the independent variable and it has the problem of reaching ±∞ and tanh(x) deriv-
ative goes to 0. The gradients become smaller and smaller which means the network
cannot be trained anymore, whereas in ReLU, the derivative is a constant if x > 0. The
relationship between the Network and input is non-linear with the use of non-linear
activation function, which makes trained networks take complex decisions. In a regular
LeNet 5 the activation function used is tanh(x) which is linear in nature, where a modi-
fication in LeNet 5 by updating Linear function to ReLU a non-linear function gives a
better classification accuracy and achieved max local minima (discussed in the results
session). in the below Table 6.

13
Handwritten Devanagari Character Recognition Using Modified… 359

5 Experiment Results and Discussion

5.1 Experimental Data

5.1.1 Dataset

The data contains handwritten samples of Devanagari numerals and vowels (i.e., 10 numer-
als and 13 consonants). Thus, the dataset includes a total of 23 different Devanagari char-
acters, as shown in Tables 1 and 2. The data is collected on a regular A4 sheet which was
distributed to the subjects. The subjects wrote numerals and vowels in Devenagari Scripts.
Targeted subjects are from schools and colleges. These sheets were scanned at 300 dpi
using Epson DS – 150 is shown in Figs. 1, and 2. The numerals and vowels were collected
from 2400 and 1400 subjects of different age groups (Figs. 3, 4, Tables 3, 4).
Characters are extracted from the scanned images using the segmentation algorithm
mentioned in Sect. 2. Further, segmented data is pre-processed. This characters are manu-
ally segregated since noise in the scanned images are also obtained as characters. All the
images were resized to 28 × 28 pixels, verified manually and converted into black and
white. Where background was converted to black and character was converted to white, as
shown in Tables 1 and 2. These are stored in a publicly accessible location. By removing
the occluded images and scribbles. the final data set contains a total of 38,750 digitized
images where 22,500 Devanagari Numerals (2250 each) and 16,250 Vowels (1250 each).
Extracted images of Devanagari characters were arranged into different folders—a total of
23 folders where each folder for each character (Figs. 5, 6).

Table 1  Devanagari Numerals Representation

0 1 2 3 4 5 6 7 8 9

10 1 2 3 4 5 6 7 8 9

Table 2  Devanagari Vowels Representation

a aa e ee u uu ru ye i o au am aha

11 12 13 14 15 16 17 18 19 20 21 22 23

Table 3  Devanagari Characters that looks similar

Vowels
a aa u uu o au am aha
(pronunciation)

Vowels Sample
Image

Class Number 11 12 15 16 20 21 22 23

13
 360 D. S. Prashanth et al.

Table 4  Devanagari Numerals that looks similar

Numerals 1 2 3 5 6 7 8

Numerals Sample
Image

Class Number 1 2 3 5 6 7 8

The success rate of research on the recognition of handwritten English characters is

high when compared to the Indian script like Devanagari. The state-of-the-art techniques
in deep learning are efficient in automatic recognition of Devanagari handwritten charac-
ters, but this requires large samples of data with labels. This data will fill that data-gap for
Devanagari numerals and vowels. This data is publicly available at https://​data.​mende​ley.​
com/​datas​ets/​pxrnv​p4yy8/1 (Figs. 7, 8).
In this research, CNN, MLCNN, and ACNN is used for implementing HDCR. Dataset
used in this work is collected from 3,800 subjects of different ages from schools, colleges,
and other professions. The number of images filtered after pre-processing is 38,750. The
data is split into different ratios of 80:20, 70:30, and 60:40 for training and testing, as men-
tioned in Table 10 (Figs. 9, 10).

5.2 Experimental Settings

In this research, the input image characters are resized to 28 × 28, following the standards
of the MNIST dataset. If the size of the character has increased, the accuracy may also
increase, which is directly proportional to the computational cost. So, to avoid more com-
putational cost, the size is fixed to 28 × 28. For training the network, the batch size is set to
32. Both the numerals and alphabets are trained separately as per Table 8.

5.2.1 For CNN

Along with the above settings, SGD is used to train the model at the learning rate of 0.01.
In the first layer, the network is initialized with 36 kernels and increased to 64 kernels.
Dropouts are added to avoid overfitting with the ratio of 25%, and 50% of neurons are
dropped before the final layer. The architecture of CNN is as mentioned in the Table 5.

5.2.2 For MLCNN

Convolution kernels used in the MLCNN network is 6 and 16 in different layers, which are
very useful. Strides are used with values 1 and 2 to control the size of the convolution layer
in subsampling. The padding of [1 1 1 1] is implemented in the second and fourth layers. In
this NN, SGD is used with a learning rate of 0.01. It is prepared to update the value of the
learning rate if the error rate is compromised. The architecture of MLCNN is as mentioned
in the Table 6.

13
 Table 5  CNN Architecture
Layer (type) Image input Param No. of params Activation function

1 ‘conv1’ Convolution 32 kernel 3 × 3 and input shape 320 ReLU

28 × 28
2 ‘pool1’ Max pooling 2 × 2 Avg Pooling, input shape – –
26 × 26 × 32
2 ‘conv2’ Convolution 64 kernel 3 × 3 and input shape 18,496 ReLU
13 × 13 × 32
3 ‘pool1’ Max pooling 2 × 2 Avg Pooling, input shape – –
11 × 11 × 64
Handwritten Devanagari Character Recognition Using Modified…

4 flatten Flatten 1600, input shape 5 × 5 × 64 – –

5 ‘fc1’ Fully connected Fully connected with dense 128 1,179,776 ReLU
6 ‘drop2’ Dropout 50% dropout – –
7 ‘fc2’ Fully connected Fully connected with dense 10 1290 Softmax
Total param: Trainable params: 1,199,882 Non-trainable
1,199,882 params: 0
361

13
 362

13
Table 6  Modified Lenet CNN Architecture
Layer (type) Image input Param No. of params Activation function

1 ‘conv1’ Convolution 6 kernel 5 × 5 and input shape 156 ReLU

28 × 28
2 ‘pool1’ Avg pooling 2 × 2 Avg Pooling, padding [1 1 1 1],
Stride 2, input shape 27 × 27 × 6
3 ‘conv2’ Convolution 16 kernel 5 × 5, Stride 2, input shape 2416 ReLU
23 × 23 × 6
4 ‘pool2’ Average pooling 2 × 2 Avg pooling, padding [1 1 1 1],
Stride 1, input shape 11 × 11 × 16
5 ‘fc1’ Fully connected Fully connected with dense 120, 232,440 ReLU
input shape 5 × 5 × 16
6 ‘fc2’ Fully connected Fully connected with dense 84 10,164 ReLU
7 ‘fc3’ Fully connected Fully connected with dense 13 / 10 1105 Softmax
Total param: Trainable params: 246,281 Non-trainable
246,281 params: 0
D. S. Prashanth et al.
Handwritten Devanagari Character Recognition Using Modified… 363

Fig. 3  CNN Architecture

Fig. 4  Modified LeNet CNN Architecture

Fig. 5  Loss for 40 Epochs for ACNN for 80:20 Vowels Dataset

5.2.3 For ACNN

ACNN has its first layers with 28 kernels that are regularized with the L2 regularization
technique. L2 regularization helps to avoid the overfitting of data from the first layer itself.
Batch normalizations are added to all the convolution layers with activation function ReLU
that introduces non-linearity to the data. Kernels that are used in each layer are 56, 112,
224, 448 with the window size of (3,3) (5,5) (7,7) (3,3), and (3,3), respectively. Since there
are essential parameters in Alexnet CNN – to avoid further overfitting, dropout is added at
the ratio of 50% for the first two fully connected layers. The final output layer is softmax.
The architecture of MLCNN is as mentioned in the Table 7.

13
 364 D. S. Prashanth et al.

Fig. 6  Accuracy for 40 Epochs for ACNN for 80:20 Vowels Dataset

Fig. 7  Loss for 40 Epochs for MLCNN for 80:20 Vowels Dataset

5.3 Results Comparison and Training strategy of CNN, MLCNN and ACNN

Three different CNN architectures are designed before performing classification and
trained as per the ratios in Table  8. Table  11 shows the training accuracies achieved by
CNN, MLCNN, and ACNN for alphabets and numerals (Fig. 11, Tables 9, 10).
A simple CNN architecture is used in this research for both alphabets and numer-
als. Alphabets achieved an accuracy of 96.88% of efficiency for 70:30 split ratio at 50
epochs and 96.29%, 96.00% for 80:20 and 60:40 as shown in the Fig.  12 split ratio.
ACNN architecture obtained an accuracy of 99.99% is collected for 40 epochs training
of alphabets at 80:20 ratio split as shown in the Fig.  8. The other accuracies obtained

13
Handwritten Devanagari Character Recognition Using Modified… 365

Fig. 8  Accuracy for 40 Epochs for MLCNN for 80:20 Vowels Dataset

Fig. 9  Loss for 50 Epochs for CNN for 60:40 Vowels Dataset

at 70:30 and 60:40 is 99.98% and 99.97%, respectively. The training of alphabets is
stopped at 40 epochs; the reason for not recording the 50 epochs training is—loss values
beyond 40 epochs are increasing and crossing the local minima in the opposite direc-
tion. A similar problem was raised for MLCNN. The training strategy used in MLCNN
is, where SGD is using as the optimizer, the learning rate is updated to 0.001 from 0.01.
The advantage of using SGD over the Adadelta optimizer is tuning the learning rate.
By slowing down the learning rate, the accuracy is compromised. In MLCNN, at 33rd
epoch loss is increasing so, the learning rate is tuned to 0.001. Later, it is observed that
the cost value is not reversed when it is trained for 50 epochs. When the network is
trained for 40 epochs as shown in the Fig. 10, the accuracy obtained is 99.98%, 99.98%,

13
 366 D. S. Prashanth et al.

Fig. 10  Accuracy for 50 Epochs for CNN for 60:40 Vowels Dataset

and 99.97% for 80:20, 70:30 and 60:40 at 33rd, 35th, and 35th epoch simultaneously.
The problem of overfitting was raised for alphabets, because of 13 classes in HDCR
of alphabets. Eight categories of characters look similar to minor changes, as shown in
Table 3.
{1,2}, {3, 6}, and {5, 7, 8} sets of numbers in Devanagari have structural similarity
as shown in Table 4. CNN architecture obtained training accuracy of 97.22%, 96.86%,
and 96.77% at different splits. ACNN obtained top accuracy of 99.87% for 80:20 ratio
at 40 epochs as show in the Fig. 12 and Fig. 13. MLCNN has shown promising results
with the least loss for both training and testing, as shown in Table 12. It obtained an
accuracy of 99.84%, 99.83%, and 99.82% at different splits and epochs (Fig. 14).
A trained model efficiency can be calculated not only based on training data accu-
racy. But, based on accuracy, it gives, when it encounters unseen data. Table 12 indi-
cates the accuracy achieved by the sensed data by the neural network model. Overall,
the training accuracies of all the three models are similar when it is compared to test-
ing accuracy/ validation accuracy or accuracy on the unseen dataset; there are some
differences in the values. ACNN for alphabets obtained an accuracy of 98.25% accu-
racy for 80:20 split at 40 epochs. The regularization techniques L2 and dropouts are
used in this ACNN helped to give better accuracy. 98.25% and 98.32% for 70:30 and
60:40 split of data. Whereas, numerals are considered 97.22% of accuracy is achieved
at 70:30 for 40 epochs. The loss calculated over epochs for vowels dataset are shown
in the Figs. 7, 9 and 11 for ACNN, MLCNN and CNN respectively. The loss calculated
for MLCNN and CNN for numerals dataset is shown in the Figs. 11 and 14.
Each Deep Learning Network is trained and tested for 12 number of times for alpha-
bets data and 12 number of times for numbers data. This process is categorized into 4,

13
 Table 7  Alexnet CNN Architecture
Layer (type) Image input Param No. of params

1 ‘conv1’ Convolution 28 kernel 3 × 3 and input shape 28 × 28 280

2 ‘norml’ Batch normalization Batch normalization 112
3 ‘activation1’ ReLU ReLU –
4 ‘pool1’ Max pooling 2 × 2 max pooling –
5 ‘conv2’ Convolution 56 kernel 5 × 5 and padding [1 1 1 1] 39,256
6 ‘norm2’ Batch normalization Batch normalization 224
7 ‘activation2’ ReLU ReLU –
8 ‘pool2’ Max pooling 2 × 2 max pooling –
9 ‘zeropadding3’ Zero padding Zero Padding [1 1 1 1] –
10 ‘conv3’ Convolution 112 kernel 7 × 7 and padding [1 1 1 1] 307,440
11 ‘norm3’ Batch normalization Batch normalization 448
12 ‘activation3’ ReLU ReLU –
13 ‘pool3’ Max pooling 2 × 2 max pooling –
14 ‘zeropadding4’ Zero padding Zero Padding [1 1 1 1] –
Handwritten Devanagari Character Recognition Using Modified…

15 ‘conv4’ Convolution 224 kernel 3 × 3 and padding [1 1 1 1] 226,016

16 ‘norm4’ Batch normalization Batch normalization 896
17 ‘activation4’ ReLU ReLU –
18 ‘zeropadding5’ Zero padding Zero padding [1 1 1 1] –
19 ‘conv5’ Convolution 448 kernel 3 × 3 and padding [1 1 1 1] 903,616
20 ‘norm5’ Batch normalization Batch normalization 1792
21 ‘activation5’ ReLU ReLU –
22 ‘pool5’ Max pooling 2 × 2 max pooling –
23 ‘fc6’ Fully connected Fully connected Layer with dense 786 5,620,496
24 ‘norm6’ Batch normalization Batch normalization 3136
25 ‘activation6’ ReLU ReLU –
26 ‘drop6’ Dropout 50% dropout –
367

13
27 ‘fc7’ Fully connected Fully connected layer with dense 1204 945,140
 Table 7  (continued)
368

Layer (type) Image input Param No. of params

13
28 ‘norm7’ Batch normalization Batch normalization 4816
29 ‘activation7’ ReLU ReLU –
30 ‘drop7’ Dropout 50% dropout –
31 ‘fc8’ Fully connected Fully connected layer with dense 10 15,665
32 ‘norm8’ Batch normalization Batch normalization 52
33 ‘prob’ Softmax Softmax –
34 ‘Output’ Classification output Crossentropy with Adadelta optimizer –
Total param: Trainable params: 8,063,647 Non-trainable params: 5738
8,069,385
D. S. Prashanth et al.
Handwritten Devanagari Character Recognition Using Modified… 369

Table 8  Data split for training Training: testing Numerals Alphabets

and testing
Training Testing Training Testing

80:20 18,000 4500 13,000 3250

70:30 15,750 6750 11,375 4875
60:40 13,500 9000 9750 6500

Table 9  Accuracy on training data

Training Accuracies for Alphabets (13 Classes) Training Accuracies for Numerals (10 Classes)
Train: Test ACNN MLCNN CNN Train: Test ACNN MLCNN CNN

20 Epochs 20 Epochs
80:20 99.92 99.10 91.95 80:20 99.69 99.82 95.54
70:30 99.89 99.31 90.63 70:30 99.65 99.82 95.37
60:40 99.96 97.81 91.37 60:40 99.61 99.81 94.36
30 Epochs 30 Epochs
80:20 99.99 99.98 94.01 80:20 99.79 99.82 96.10
70:30 97.91 99.96 93.83 70:30 99.86 99.84 96.01
60:40 99.22 99.91 93.22 60:40 99.84 99.83 95.42
40 Epochs 40 Epochs
80:20(39) 99.99 99.98 95.32 80:20 99.87 99.82 96.66
70:30 99.98 99.98 95.50 70:30 99.85 99.81 96.56
60:40 99.97 99.97 94.91 60:40 99.87 99.81 96.04
50 Epochs 50 Epochs
80:20 – 95.32 96.29 80:20 99.88 99.83 97.22
70:30 – 94.28 96.68 70:30 99.83 99.81 96.86
60:40 – 92.53 96.00 60:40 99.81 99.81 96.77

i.e., 20 epochs, 30 epochs, 40 epochs, and 50 epochs. Every category is further trained
and test based on the split of data into different ratios, i.e., 80:20, 70:30, 60:40. The
proposed deep learning architectures (CNN, MLCNN, and ACNN) are developed and
tested on the dataset for 12 times. The dataset is tested for 72 times. For every instance,
four parameters are taken to check the model stability. Training and testing accuracy
and loss values are collected. Among the 72 different sets of values, Table 11 contains
the optimal values. A study measures it, maximum accuracy of both training and test-
ing, along with the minimum costs of training and testing loss.

13
 370 D. S. Prashanth et al.

Table 10  Accuracies calculated on the unseen data

Testing Accuracies for Alphabets (13 Classes) Testing Accuracies for Numerals (10 Classes)
Train: Test ACNN MLCNN CNN Train: Test ACNN MLCNN CNN

20 Epochs 20 Epochs
80:20 98.03 91.41 92.46 80:20 97.07 96.56 96.0
70:30 97.74 88.86 91.85 70:30 96.17 96.49 95.44
60:40 97.40 88.40 91.88 60:40 97.30 96.57 95.79
30 Epochs 30 Epochs
80:20 98.25 91.56 93.85 80:20 97.20 96.64 96.24
70:30 97.91 91.87 93.05 70:30 96.74 96.41 96.04
60:40 98.26 90.88 92.08 60:40 97.27 96.41 95.86
40 Epochs 40 Epochs
80:20(39) 98.25 94.09 93.54 80:20 97.16 96.71 96.18
70:30 98.24 91.73 93.62 70:30 97.22 96.40 96.21
60:40 98.32 91.11 93.05 60:40 97.27 96.49 96.20
50 Epochs 50 Epochs
80:20 – 88.86 94.31 80:20 97.27 96.58 96.44
70:30 – 87.53 94.21 70:30 97.28 96.55 96.44
60:40 – 85.83 93.52 60:40 97.26 96.46 96.28

5.4 Comparison of Different Methods HDCR

To show the pre-eminence of the proposed architecture and dataset efficiency, we com-
pare the performances of different models from the recent review articles [2, 6, 8]. The
parameter considered for comparison is the size of the dataset, because, lower the size
of the dataset leads to cover fewer variant styles and may achieve higher accuracies but
when the dataset size and more variants of writing styles are introduced in the dataset
may lead to a drop of accuracies. In the Table 12, the highest accuracy achieved is 99%
in [32], whereas the size of the dataset is 5424. Similarly, the dataset of sized 10,000
[31], 25,000 [18], 22,556 [23], and 32,413 [32], got accuracy of 97.18%, 90.11%, 93.4%
and 61.8% respectively. It is observed that the accuracy is reduced when more variants
of data or data size increase. In this article, the dataset size of 22,500 images could able
to achieve 99.9% training data and 98.25% for unseen datasets for numbers. For the
dataset size of 16,250 images, could able to achieve 99.87% for training data and 97.16
for unseen data. our model has significant improvement.

13
 Table 11  Top Accuracies
ACNN for Alphabets (13 Classes) ACNN for Numerals (10 Classes)

40 Epochs 40 Epochs
Train: Test Training Accuracy Testing Accuracy Train Loss Test Loss Train: Test Training Accuracy Testing Accuracy Train Loss Test Loss

80:20 99.99 98.25 0.001 0.14 80:20 99.87 97.16 0.003 0.20
70:30 99.98 98.24 0.001 0.13 70:30 99.85 97.22 0.004 0171
60:40 99.98 98.32 0.001 0.11 60:40 99.87 97.27 0.003 0.17
MLCNN for Alphabets (13 Classes) MLCNN for Numerals (10 Classes)

40 Epochs 30 Epochs
Train: Test Training Accuracy Testing Accuracy Train Loss Test Loss Train: Test Training Accuracy Testing Accuracy Train Loss Test Loss

80:20 (33) 99.98 94.09 0.001 0.34 80:20 99.82 96.64 0.002 0.21
Handwritten Devanagari Character Recognition Using Modified…

70:30 (35) 99.98 91.73 0.0010 0.50 70:30 99.84 96.41 0.002 0.21
60:40(35) 99.97 91.11 0.001 0.50 60:40 99.83 96.41 0.002 0.20
CNN for Alphabets (13 Classes) CNN for Numerals (10 Classes)

50 Epochs 50 Epochs
Train: Test Training Accuracy Testing Accuracy Train Loss Test Loss Train: Test Training Accuracy Testing Accuracy Train Loss Test Loss

80:20 96.29 94.31 0.10 0.21 80:20 97.22 96.44 0.07 0.12
70:30 96.68 94.21 0.09 0.23 70:30 96.86 96.44 0.08 0.12
60:40 96.00 93.52 0.11 0.23 60:40 96.77 96.28 0.09 0.12
371

13
 372 D. S. Prashanth et al.

Fig. 11  Loss for 50 Epochs for MLCNN for 80:20 Numerals Dataset

Fig. 12  Accuracy for 50 Epochs for MLCNN for 80:20 Numerals Dataset

6 Conclusion

In this paper, we provide a dataset of 38,750 images of Devanagari numerals and vow-
els collected from 2400 subjects of different age groups. We also developed a segmenta-
tion algorithm suitable for the dataset. A modified CNN is proposed to solve the HDCR

13
Handwritten Devanagari Character Recognition Using Modified… 373

Table 12  Different methods and accuracy of HDCR

Method Feature Accuracy (%) Dataset

Tree based classifier and Tem- Structural features 97.18 10,000

plate matching based approach
[31]
KNN and neural network [32] Gradient features 61.8 32,413
MLP and SVM classifier [18] Chain code, Kirsch Directional edges, 90.116 by SVM 25,000
Gradient, distance transform and 87.56 by
MLP
Neural network classifier [23] Combination of structural and statistical 93.4 22,556
features
MQDF [32] Directional features 99.0 5424
MLCR [34] Geometric distance based 98.1 3100
CNN [35] Convolution 95 4282
KNN [36] Hu’s seven variants and zernike moment 89 2600
features
MLCNN Modified LeNet convolutional neural 99.9 38,750
network

Fig. 13  Loss for 50 Epochs for CNN for 80:20 Numerals Dataset

problem. The dataset developed followed the standards of MNIST and served as a bench-
mark dataset for Devanagari numerals and vowels. Experiments are conducted by using
three different deep learning architectures, 1. CNN, 2. MLCNN, and 3. ACNN. It achieved
an error rate of 0.001% and 99.98% accuracy. Through a series of experiments, we have
shown that MLCNN and ACNN are highly efficient (99.9% and 99.8%). ACNN’s compu-
tational cost is high when compared to MLCNN but can overcome by GPU based systems.

13
 374 D. S. Prashanth et al.

Fig. 14  Loss for 50 Epochs for CNN for 80:20 Numerals Dataset

We implemented a CNN by modifying LeNet 5 CNN, which reduces the time significantly,
where these models can be used in online HDCR. The performance of ACNN can be fur-
ther improved through the fine-tuning of its structure and its parameters. In future work,
we plan to develop a combined model for all the Devanagari characters of more than 50
classes that include numeral (10), vowels (13), and consonant (33).

Authors’ contributions Duddela Sai Prashanth: Conceptualization, Methodology, Software, Visualization

R Vasanth Kumar Mehta: Data curation, Writing—original draft, Data Analysis, Investigation. Kadiyala
Ramana: Software, Validation, Editing. Vidya Charan Bhaskar: Supervision, Writing—review & editing.

Funding  This research did not receive any specific grant from funding agencies in the public, commercial,
or not-for-profit sectors.

Availability of data and material  The datasets generated during and/or analysed during the current study are
available in the Mendeley repository, https://​data.​mende​ley.​com/​datas​ets/​pxrnv​p4yy8/1

Code availability  The code that supports the findings of this study are available on request from the cor-
responding author. The code is not publicly available due to containing information that could compromise
the privacy of research participants.

Declarations 

Conflict of interests  The authors declare that they have no known competing financial interests or personal
relationships that could have appeared to influence the work reported in this paper.

References
1. Sahare, P., & Dhok, S. B. (2019). Robust character segmentation and recognition schemes for mul-
tilingual indian document Images. IETE Technical Review, 36(2), 209–222.
2. Parekh, K. A., Goswami, M. M., & Mitra, S. K. (2020). Handwritten numeral recognition using
polar histogram of low-level stroke features. In Proceedings of 3rd international conference on
computer vision and image processing (pp. 169–181). Springer.

13
Handwritten Devanagari Character Recognition Using Modified… 375

3. Bansal, V., & Sinha, R. M. K. (2002). Segmentation of touching and fused Devanagari characters.
Pattern recognition, 35(4), 875–893.
4. Li, Z., Teng, N., Jin, M., & Lu, H. (2018). Building efficient CNN architecture for offline handwritten
Chinese character recognition. International Journal on Document Analysis and Recognition (IJDAR),
21(4), 233–240.
5. Kaur, S., & Sagar, B. B. (2019). Brahmi character recognition based on SVM (support vector machine)
classifier using image gradient features. Journal of Discrete Mathematical Sciences and Cryptography,
22(8), 1365–1381.
6. Setlur, S. (2009). Guide to OCR for indic scripts. In V. Govindaraju (Ed.). Springer.
7. Bathla, A. K., Gupta, S. K., & Jindal, M. K. (2016, March). Challenges in recognition of Devanagari
Scripts due to segmentation of handwritten text. In 2016 3rd International Conference on Computing
for Sustainable Global Development (INDIACom) (pp. 2711–2715). IEEE.
8. https://​data.​mende​ley.​com/​datas​ets/​pxrnv​p4yy8/1
9. Bharath, A., & Madhvanath, S. (2011). HMM-based lexicon-driven and lexicon-free word recognition
for online handwritten Indic scripts. IEEE transactions on pattern analysis and machine intelligence,
34(4), 670–682.
10. Soora, N. R., & Deshpande, P. S. (2018). Review of feature extraction techniques for character recogni-
tion. IETE Journal of Research, 64(2), 280–295.
11. Sharma, A. K., Adhyaru, D. M., & Zaveri, T. H. (2020). A survey on Devanagari character recognition.
In Smart systems and IoT: Innovations in computing (pp. 429–437). Springer.
12. Goswami, M. M., Prajapati, H. B., & Dabhi, V. K. (2011). Classification of printed Gujarati characters
using SOM based k-Nearest Neighbor Classifier. In 2011 International conference on image informa-
tion processing (pp. 1–5). IEEE.
13. Sinha, R. M. K., & Mahabala, H. N. (1979). Machine recognition of Devanagari script. IEEE Transac-
tions on Systems, Man and Cybernetics, 9(8), 435–441.
14. Jayanthi, K., Suzuki, A., Kanai, H., Kawazoe, Y., Kimura, M., & Kido, K. (1989, November). Devana-
gari character recognition using structure analysis. In Fourth IEEE Region 10 international conference
TENCON (pp. 363–366). IEEE.
15. Bansal, V., & Sinha, M. K. (2001, September). A complete OCR for printed Hindi text in Devanagari
script. In Proceedings of Sixth International Conference on Document Analysis and Recognition (pp.
0800–0800). IEEE Computer Society.
16. Sharma, N., Pal, U., Kimura, F., & Pal, S. (2006). Recognition of offline handwritten devnagari char-
acters using quadratic classifier. In Computer vision, graphics and image processing (pp. 805–816).
Springer.
17. Arora, S., Bhattacharjee, D., Nasipuri, M., Basu, D. K., Kundu, M., & Malik, L. (2009, December).
Study of different features on handwritten Devnagari character. In 2009 Second international confer-
ence on emerging trends in engineering & technology (pp. 929–933). IEEE.
18. Kumar, S. (2009). Performance comparison of features on Devanagari hand-printed dataset. Interna-
tional Journal of Recent Trends in Engineering, 1(2), 33.
19. Arora, S., Bhattacharjee, D., Nasipuri, M., Basu, D. K., & Kundu, M. (2010). Recognition of non-
compound handwritten devnagari characters using a combination of mlp and minimum edit dis-
tance. arXiv preprint arXiv:1006.5908.
20. Pal, U., Chanda, S., Wakabayashi, T., & Kimura, F. (2008). Accuracy improvement of Devnagari
character recognition combining SVM and MQDF. In Proceedings of 11th International Con-
ferenece on Frontiers Handwrit. Recognition (pp. 367–372).
21. Kale, K. V., Deshmukh, P. D., Chavan, S. V., Kazi, M. M., & Rode, Y. S. (2013). Zernike moment
feature extraction for handwritten Devanagari compound character recognition. In 2013 Science
and information conference (pp. 459–466). IEEE.
22. Arora, S., Bhatcharjee, D., Nasipuri, M., & Malik, L. (2007). A two stage classification approach
for handwritten Devnagari characters. In International conference on computational intelligence
and multimedia applications (ICCIMA 2007) (Vol. 2, pp. 399–403). IEEE.
23. Hanmandlu, M., Murthy, O. R., & Madasu, V. K. (2007, December). Fuzzy model based recogni-
tion of handwritten Hindi characters. In 9th Biennial Conference of the Australian Pattern Recogni-
tion Society on Digital Image Computing Techniques and Applications (DICTA 2007) (pp. 454–
461). IEEE.
24. Aneja, N., & Aneja, S. (2019, July). Transfer learning using CNN for handwritten Devanagari char-
acter recognition. In 2019 1st International Conference on Advances in Information Technology
(ICAIT) (pp. 293–296). IEEE.

13
 376 D. S. Prashanth et al.

25. Arora, S., Bhattacharjee, D., Nasipuri, M., Malik, L., Kundu, M., & Basu, D. K. (2010). Perfor-
mance comparison of SVM and ANN for handwritten devnagari character recognition. arXiv pre-
print arXiv:1006.5902
26. Lu, S., Lu, Z., & Zhang, Y. D. (2019). Pathological brain detection based on AlexNet and transfer
learning. Journal of computational science, 30, 41–47.
27. Shima, Y., Nakashima, Y., & Yasuda, M. (2018, April). Handwritten Digits Recognition by Using
CNN Alex-Net Pre-trained for Large-scale Object Image Dataset. In Proceedings of the 3rd inter-
national conference on multimedia systems and signal processing (pp. 36–40).
28. Kumar, M., Jindal, M. K., Sharma, R. K., & Jindal, S. R. (2019). Character and numeral recogni-
tion for non-Indic and Indic scripts: A survey. Artificial Intelligence Review, 52(4), 2235–2261.
29. Niu, X. X., & Suen, C. Y. (2012). A novel hybrid CNN–SVM classifier for recognizing handwritten
digits. Pattern Recognition, 45(4), 1318–1325.
30. Jiang, B., He, J., Yang, S., Fu, H., Li, T., Song, H., & He, D. (2019). Fusion of machine vision tech-
nology and AlexNet-CNNs deep learning network for the detection of postharvest apple pesticide
residues. Artificial Intelligence in Agriculture, 1, 1–8.
31. Chaudhuri, B. B., & Pal, U. (1997, August). An OCR system to read two Indian language scripts:
Bangla and Devnagari (Hindi). In Proceedings of the fourth international conference on document
analysis and recognition (Vol. 2, pp. 1011–1015). IEEE.
32. Kompalli, S., Nayak, S., Setlur, S., & Govindaraju, V. (2005, August). Challenges in OCR of
Devanagari documents. In Eighth International Conference on Document Analysis and Recognition
(ICDAR’05) (pp. 327–331). IEEE.
33. Pal, U., Sharma, N., Wakabayashi, T., & Kimura, F. (2007, September). Handwritten numeral rec-
ognition of six popular Indian scripts. In Ninth International Conference on Document Analysis
and Recognition (ICDAR 2007) (Vol. 2, pp. 749–753). IEEE.
34. Soora, N. R., & Deshpande, P. S. (2017). Novel geometrical shape feature extraction techniques for
multilingual character recognition. IETE Technical Review, 34(6), 612–621.
35. Prashanth, D. S., Mehta, R. V. K., & Sharma, N. (2020). Classification of Handwritten Devanagari
number-an analysis of pattern recognition tool using neural network and CNN. Procedia Computer
Science, 167, 2445–2457.
36. Prashanth, D. S., & Panini, C. N. (2017). KNN classification of Kannada Characters using Hu’s
Seven Variants and Zernike Moment. Nagendra Panini Challa on Aug, 10.
37. Latif, G., et  al. (2018). Deep convolutional neural network for recognition of unified multi-lan-
guage handwritten numerals. 2018 IEEE 2nd International workshop on Arabic and derived script
analysis and recognition (ASAR). IEEE.
38. Chakraborty, B., et al. (2018). Does deeper network lead to better accuracy: a case study on hand-
written Devanagari characters. 2018 13th IAPR International Workshop on Document Analysis Sys-
tems (DAS). IEEE.
39. Rojatkar, D. V., Chinchkhede, K. D., & Sarate, G. G. (2013). Design and analysis of LRTB feature
based classifier applied to handwritten Devnagari characters: A neural network approach. In 2013
International Conference on Advances in Computing, Communications and Informatics (ICACCI).
IEEE,.
40. Ashokkumar, P., Siva Shankar, G., Gautam Srivastava, Praveen Kumar Reddy Maddikunta, & Thippa
Reddy Gadekallu. 2021. “A Two-stage Text Feature Selection Algorithm for Improving Text Clas-
sification”. ACM Trans. Asian Low-Resour. Lang. Inf. Process. 20, 3, Article 49 (July 2021), 19 pages.
DOI:https://​doi.​org/​10.​1145/​34257​81
41. Dhanamjayulu, C., Nizhal, U. N., Kumar Reddy Maddikunta, P., Thippa Reddy Gadekallu, Celestine
Iwendi, Chuliang Wei, & Qin Xin. (2021). Identification of malnutrition and prediction of BMI from
facial images using real-time image processing and machine learning. IET IMAGE PROCESSING.
42. Gadekallu, T. R., Alazab, M., Kaluri, R., Reddy Maddikunta, P. K., Bhattacharya, S., Lakshmanna, K.,
& Parimala, M. (2021). Hand gesture classification using a novel CNN-crow search algorithm. Com-
plex & Intelligent Systems, 7, 1–14.
43. Kothai, G., Poovammal, E., Dhiman, G., Ramana, K., Sharma, A., AlZain, M. A., & Masud, M.
(2021). A new hybrid deep learning algorithm for prediction of wide traffic congestion in smart cities.
Wireless Communications and Mobile Computing, 2021, 1.

Publisher’s Note  Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

13
Handwritten Devanagari Character Recognition Using Modified… 377

Duddela Sai Prashanth, received M.Tech from Indian Institute of

Information Technology, Allahabad and pursuing his PhD from SCS-
VMV University. He is serving as Assistant Professor from 2015 in
Sahyadri College of Engineering & Management, Mangaluru. He is
Co-Principal Investigator for the project sanctioned 30,00,000 INR by
Vision Group of Science and Technology to establish Center of Excel-
lence in Artificial Intelligent and Machine Learning.

Dr R Vasanth Kumar Mehta  received M.S from BITS Pilani and the
PhD Degree from SCSVMV University. He Joined SCSVMV Univer-
sity as Assistant Professor in the year 2001, where he is now Associate
Professor. He is serving the University with the different responsibili-
ties, as Board of Studies – Chairman, enabling off-campus internships
and placements, Interaction with industries and academia in India and
abroad for enhancing teaching, placements, industry grade projects
and research. His Area of Interest lies in Data Mining, Distributed
Computing Domain, Pattern Recognition. He is professional member
of IEEE, CSI, ACM, IAENG. He is a member in Editorial Boards and
Reviewer of reputed Journals.

Dr. Kadiyala Ramana  is currently working as Associate Professor in

Department of Artificial Intelligence & Data Science, AITS, Rajam-
pet, India. He obtained his Bachelor of Technology degree in Informa-
tion Technology from JNTUH, Hyderabad, India, M. Tech. from Saty-
abhama University, Chennai, India and completed his Ph.D. from
SRM University, Chennai, India. He has 14 years of experience in
teaching and research. He authored more than 30 international publica-
tions in Hindawi, Springer, IEEE etc. He is editorial board member
and reviewer of several journals of international repute. His research
interests Wireless Sensor Networks, Software-Defined Networking
with Machine Learning and Data Analytics.

13
 378 D. S. Prashanth et al.

Dr. Vidhyacharan Bhaskar  received the B.Sc. degree in Mathematics

from the University of Madras, Chennai, India in 1992, M.E. degree in
Electrical & Communication Engineering from the Indian Institute of
Science, Bangalore in 1997, and the M.S.E. and Ph.D. degrees in Elec-
trical Engineering from the University of Alabama in Huntsville in
2001 and 2002, respectively. During 2002-2003, he was a Post Doc-
toral fellow with the Communications research group at the University
of Toronto, Canada, where he worked on the applications of space-
time coding for wireless communication systems. During 2003-2006,
he was an Associate Professor in the Department of Information Sys-
tems and Telecommunications at the University of Technology of
Troyes, France. From Jan. 2007 to May 2014, he was a Full Professor
in the Department of Electronics and Communication Engineering at
S.R.M. University, Kattankulathur, India. Since 2015, he is a Professor
in the Department of Electrical and Computer Engineering at San
Francisco State University, San Francisco, California, USA. His
research interests include MIMO wireless communications, signal pro-
cessing, error control coding and queuing theory. He has published 130 Refereed Journal papers, presented
around 72 Conference papers in various International Conferences over the past 20 years, published a book
on “Adaptive Rate Coding for A-CDMA Systems” in Jan 2011, a book on “Higher-Order s-to-z mapping
functions for digital filters,” in March 2013, and has also co-authored a book on MATLAB in 1998. He has
868 Google scholar citations. He has advised 50 Masters students, 11 Doctoral students, and 3 Senior design
projects. He is an IEEE Senior member (SM-IEEE) and is a member of IET (M-IET, UK). He is a Fellow of
Institute of Electronics and Telecommunication Engineers (F-IETE), and a Fellow of Institute of Engineers
(F-IE), Kolkata, India. He is also a Life member of the Indian Society of Technical Education (LM-ISTE)
and a member of the Indian Science Congress (M-ISC).

Authors and Affiliations

Duddela Sai Prashanth1,2 · R. Vasanth Kumar Mehta1 · Kadiyala Ramana3 ·

Vidhyacharan Bhaskar4 
Duddela Sai Prashanth
dsaip13@gmail.com
R. Vasanth Kumar Mehta
vasanthmehta@gmail.com
Kadiyala Ramana
ramana.it01@gmail.com
1
SCSVMV University, Kanchipuram, India
2
Sahyadri College of Engineering and Management, Mangaluru, India
3
Department of Artificial Intelligence and Data Science, Annamacharya Institute of Technology
and Sciences, Rajampet, Andhra Pradesh, India
4
Department of Electrical and Computer Engineering, San Francisco State University,
San Francisco, CA 94132, USA

13