Paper 8863

ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
Volume 2, Issue 4, June 2020

Impact Factor: 4.819
Review for Handwritten Devanagari Character

Recognition using ML Algorithms
Manisha Bhatnagar
Assistant Professor
ISBA Institute of Professional Studies, Indore, M.P., India
Abstract: Devanagari Character Recognition is a system in which handwritten Image is recognized and
converted into a digital form. Devanagari handwritten character recognition system is based on Deep
learning technique, which manages the recognition of Devanagari script particularly Hindi. This
recognition system mainly has five stages i.e. Pre-processing, Segmentation, Feature Extraction, Prediction
and Post processing. This paper has analyzed the approach for recognition of handwritten Devanagari
characters. There are various approaches to solve this. Some of the methods along with their accuracy and
techniques used are discussed here. Depending upon the dataset and accuracies of each character the
techniques differs.
Keywords: Devanagari Script; Optical Character Recognition; Segmentation; Convolutional Neural

Network (CNN); Support Vector Machine; Image Processing.
I. INTRODUCTION
Handwriting recognition has been one of the most enchanting and demanding research areas in today’s digitalized
world, which has evolved through the combination of artificial intelligence and machine learning. It contributes
exceptionally to the advancements of the interface between humans and machines. Handwritten Character Recognition
is basically ability of a system to identify human handwritten input. In general, it is classified into two types an on-line
and an off-line handwriting recognition system. The handwriting can be from many sources, such as images, paper
documents, or other devices; this is considered to be as offline system. Non-Indian languages, such as English, Chinese,
German, Japanese, Korean, etc. are already grown-up as compared to Indian scripts. Although, Indic scripts have some
more challenges in handwriting recognition than Latin, Chinese and Japanese because of the presence of variations in
the order of strokes or symbols, half consonant, etc.
II. SOME CONCEPTS OF CHARACTER RECOGNITION

2.1 Devanagari Characters
Devanagari the ancient Brahmi script, used in the Indian subcontinent and is the national font of India. Hindi is written
in Devanagari script, which is also used to write Marathi, Konkani, Nepali, possibly with little modifications.
Devanagari composed of 10 numeral characters (०, १, २, ३, ४, ५, ६, ७, ८, ९) and primary characters including 13 vowels
and 36 consonants and is the fourth most widely adopted writing system in the world. Also, Devanagari Characters have
some characters with similar structures.
For Example, the ‘ङ’ and ‘ड’ have only difference is dot. Similarly, ‘ब’ and ‘व’ are different in only the tiled line
inside circle. Also, the character ‘प’,’ म ‘, and ‘य’ appear almost to be similar. The character ‘२’ and ‘र’ appears almost
same. Some characters like ‘ ’, ‘ ’, ‘ ’ are derived from other previous characters like ‘ग’, ‘य’ etc. The Fig. 1 shows
Hindi script consists of 13 vowels, 36 consonants, and 10 numerals.
Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 71

www.ijarsct.co.in
IJARSCT

(a) (b)
(c)
Fig 1. (a) Vowels of Devanagari Script (b) Numerals of Devanagari Script (c) Consonants of Devanagari Script
Fig 2. Vowels with Corresponding Modifiers of Devanagari Script

Vowels’ modifiers have a crucial role in Hindi script. Fig.2 depicts with corresponding modifier in Devanagari script.
Hindi language words can be classified as a combination of three components: a middle component (core), a top
component and a bottom component. Top and bottom components consist of only swar modifiers and diacritic signs and
the core component contains all, the characters, punctuation and special symbols. The top and core components are
divided by a “shirorekha” (top line). A Purnaviram (full stop) is used to mark the end of a sentence or phrase which is
depicted by a vertical line.
Fig 3. Half Form of Consonants of Devanagari Script

In Hindi, when two vyanjans are combined, a straight bar with a “vyanjan” can appear as a half-form. The half-form of
“vyanjan” is the left side of the original “vyanjan” with a straight bar. Figure 3 shows half the “vyanjan” type of Fig.1c.
The vacant positions in this figure indicate that the corresponding consonant has no half form.
2.2 Character Recognition

Character Recognition is the process of extracting digitized text from images of scanned documents. Character
Recognition systems have already matured in various languages, but they still have scope in other Indian languages like
Devanagari, Bengali, Marathi,etc. Character recognition could be done through various techniques, such as

www.ijarsct.co.in
IJARSCT

using quadratic classifier [34], Curvelet Transform [35], Transfer Learning [36], Linear Discriminant Analysis [37], and
many more.
2.3 Optical Character Recognition

Hindi OCR is a model which is basically used to recognize handwritten Hindi characters. Till today, the models
developed for Indian languages like Hindi, have not shown quite good accuracy due to the complexity of the languages.
The text or characters of Hindi are little difficult to segment and evaluate mainly due to their complex structures.
2.4 Working Principle of OCR

In real life applications, OCR software accepts the document image as input and produces a written text file as output.
So, we are required not only to develop the recognition process, but also to develop the preprocessing and post-
processing parts. The flow control of the Hindi OCR processes is shown in Figure 4. In the design of the proposed
system the following steps are followed:
Fig 4. Flow control for Hindi OCR System
A. Image Acquisition
Image acquisition part plays the main role in OCR problem. No matter how accurate model is on testing, the real- world
image will never be same. In real world image, there will be lots of noise, blur, and many other quality degradations of
image.
B. Data Pre-processing
In this stage the image is converted into grayscale, and a NumPy array is prepared to store the image pixels. After this
the intention is to find foreground and background colors. Removing some noise and doing threshold makes it easier for
image to recognize text, and find foreground color. Here combination of threshold of Otsu and Binary is used. It is a
way to create a binary image from grayscale or full-color image. This is mainly done in order to separate “object” or
foreground pixels from background pixels to aid in image processing.
There are various processes in preprocessing of data such as:
1. Binarization
2. Noise Elimination
3. Skew Correction
4. Size Normalization
5. Thinning.

www.ijarsct.co.in
IJARSCT

D. Segmentation
Here breaking of the entire image into small fragments is done. Segmentation [38, 39] process is done here on unique
way by defining the possible rows for top most part of character and the possible percentage of space between
characters. Each segment individually is further passed to a prediction process.
E. Feature Extraction, Detection and Error Correction

In Feature Extraction most relevant information from the raw data is extracted and is used. The features should be
selected in such a way that it reduces the intra class variability and increase the feature space. CNN has the best feature
extraction Neural Network and scope. First network is FFNN (Feed Forward Neural Network) and it is the simple
among the NN class.
Some feature extraction methods are:
1. Fourier Transforms
2. Wavelets
3. Moments
4. Zoning
5. Crossings and Distances
6. Projections
7. Coding
8. Graphs and Trees
F. Classification
Each segment is passed to prediction process. Before doing actual prediction, the shape of Segment must be resized as
that of the neural network input. Thus, each segment is converted into 30 by 30 sized image and in addition, we added a
1-pixel borders around it with background color. Then our segments will be of 32 by 32 shape which is the input shape
for our model. Then it is to the neural network. If the segment has high prediction then is assumed that the character
should be shown. Prediction can be wrong also depending upon the image quality due to the false segmentation.
Character Recognition techniques can be classified as:
1. K-nearest neighbors [24,35]
2. Support Vector Machine [11,14,15]
3. Convolution Neural Networks [1,32]
4. Hybrid Network [18]
5. Artificial Neural Network [28]
These methods are mainly used but more methods could be taken under consideration.
G. Prediction
This is the overall collection of previous processes. The actual recognition is seen after there is bordered around the
found character and its corresponding label as well. The final detection takes some time and gives accurate prediction.
If poor image is given then it gives false prediction. So, high resolution image is recommended.
2.5 Types of Classifiers

A. Convolutional Neural Networks (CNN)
Convolutional neural networks are mainly deep neural nets primarily used to classify images, clustering them by
similarity, and object recognition. CNN performs OCR to digitize text and make natural language processing possible
on any sort of document like hand-written document. The efficiency of convolutional nets (ConvNets or CNN) in image
recognition is one of the main reasons why people are confident in using in-depth learning.
Various layers of CNNs are:
1. Convolutional layer is the layer where convolution operation occurs that is same as image processing. A filter
of same row and column or square size matrix is taken and multiplied across the window that fits filter. The
element-wise product is done and then summation is done. The concept of stride is generally used, as how
www.ijarsct.co.in
IJARSCT

much pixel shifts after doing one convolution. Here more the number of filters, more accuracy can be achieved
but computational complexity increases.
2. Max-Pooling layer take some pixels from previous layers. Pool size is defined and then that pool size is used
on input pixels. The pool matrix is moved over entire input and max value within the overlapped input is taken.
3. Dropout layer is mainly used to avoid overfitting. This layer randomly cuts the unnecessary connection
between two neurons of different layers.
4. Flatten layer is one where multiple sized input is converted into 1Dvector
5. Dense layer is used to do classification after doing whole convolution process
2.6 Architecture of CNN

The main architecture of this model involves the use of Convolutional Neural Network (CNN) layers for image
processing and feature extraction and detection and fully connected layers are added for recognition purpose. Fig 5
shows the general architecture of the proposed CNN model.
Fig 5. Architecture Diagram of CNN
A. Support Vector Machine (SVM)

Support vector machine (SVM) is considered one of the most popular and powerful supervised tool of machine
learning, which can be used for both classification as well as regression. Basically, SVM takes a set of input data and
gives predictions, for each given input. Given a set of training points, each marked as one of two phases, the SVM
training algorithm creates a model that provides new points in one phase or another. More precisely, SVM constructs
one or many hyper planes in a high or infinite dimensional space, basically used for classification, regression or other
tasks. In Fig. 6 [30], a good separation is obtained by a hyper flight at a greater distance to the nearest point of any
training stages because the limit is usually greater when reducing the general error of that separation.
Fig 6. SVM Algorithm
B. Deep Learning
Deep learning [1, 27, 40] is a branch of machine learning in artificial intelligence (AI) that has networks capable of
learning unsupervised from data that is unlabeled. Other name for Deep learning is deep neural network. Convolutional
Neural Networks (CNN) along with various other Neural Networks such as ANN, RNN and other hybrid models are
used on Deep Learning techniques.

www.ijarsct.co.in
IJARSCT

III. LITERATURE REVIEW
Acharya [1] propose a deep learning architecture mainly deep CNN for recognition of Devanagari Handwritten
characters. The main focus was on the use of Dropout and dataset increment approach to improve test accuracy and
hence, able to increase test accuracy by nearly 1 percent. Lakshmi [2] incorporates an effective method for recognition
of isolated handwritten Devanagari numerals, in which edge directions histograms and splines along with PCA for
enabling recognition accuracies is proposed. In [3], Tamil handwritten characters are recognized using Support vector
machine (SVM) in which data is collected from various A4 sized documents and then preprocessed to enhance the
quality of the image, finally achieved an accuracy up to 82.04%. Alternatively discussed by Hanmhunga et al. [4] s
based on modified descriptive membership functions embedded in unambiguous sets found in features with standard
distances obtained using the Box method. The colorful classification of Hindi characters is done using structural
features such as the position of the bar, the alignment of the letter parts, and the sides open on which side etc. They
tested on a database of 4750 samples and the total recognition rate was found to be 90.65%. Jawahar [5] describes the
character recognition for Hindi and Telugu text. He basically used Principal Component Analysis followed by support
vector classification for his bilingual OCR. The experiment was done on approximately 200000 characters, which gave
an overall accuracy up to 96.7%.
The accuracy of Stroke Recognition is important for the performance of character Recognition [6]. Sekhar used SVM
for the stroke recognition engine for Devanagari and Telugu scripts and applied Normalization, smoothing and
interpolation for preprocessing of stroke data. He had talked about various feature extraction and stroke classification
techniques such as Single recognition engine approach, Multiple Engines for Recognition and Stroke Recognition using
HMMs for Devanagari Script and Telugu Script. Chaudhuri [7] experimentally did optical character recognition
systems for different languages using soft computing techniques. Bansala and Sinha [8] presented a complete method
for segmentation for Devanagari printed text. They have used a set of filters those are robust and used a two-level
partitioning scheme and search algorithm. Various segmentation techniques were explained in his analysis. An attempt
to recognize Telugu script using KNN and a compositional approach using connected components and Fringe distance
is reported in [13]. Pankaj kale et. al.[28] proposed an ANN based recognition system for handwritten Marathi
characters and experiments were applied on 50 handwritten characters from 10 different peoples. The data was
preprocessed, and features were extracted. The accuracy obtained is 92%.
Sinha and Mahabala [29] attempt to recognize Devanagari automatically according to syntactic pattern analysis system.
They choose 26 symbols and extract the structural information from the characters in terms of primitives and their
relationships. However, their study is limited by the sample size, no attempt has been made to make it a commercial
product and it couldn’t achieve any quantitative recognition rate. Najwa et.al.[32] proposed CNN technique on self-
prepared Arabic dataset called Hijja Dataset and Arabic Handwritten Character Dataset (AHCD) and achieved an
accuracy of 97% and 88% on the AHCD dataset and the Hijja dataset, respectively. The Arabic language was chosen
because very few studies is being done for Arabic language. Bharat et.al.[33] described the challenges in recognizing
online handwriting in Indic scripts and provide an overview of the state of the art for isolated character and word
recognition. Also, it shows various resources such as tools and data sets, currently available in the Indic script visual
study online.
Sharma et.al.[34] proposed a Quadratic classifier-based system used for the recognition of unconstrained off-line
Devanagari handwritten characters. Feature vector had dimension of 64, and the features are obtained based on the
directional chain codes of the contour of the character. Encouraging results are obtained with this experiment.
Gyanendra et.al.[35] experimentally proved a technique, Curvelet transform also known as the curved singularities of
images. It is very useful for feature extraction to character images. Devanagari script characters have a lot of curve.
Firstly, segmentation of image is done then by applying curvelet transform curvelet features are obtained after the
calculation of the statistics of thick and thin images. K-Nearest Neighbor classifier is used for training the system. 200
in house images of character set is used and the model achieved an accuracy of 90%. Aneja et.al.[36] used Transfer
learning technique for the recognition of handwritten Devanagari alphabets using pre trained model for Deep
Convolution Neural Network. As a fixed feature extractor AlexNet, DenseNet, Vgg, and Inception ConvNet are
implemented. 15 epochs for each of AlexNet, DenseNet, Vgg, V3 etc has been implemented. From the results it
concluded that Inception V3 performs better for accuracy and able to achieve 99% with average epoch time 16.3

www.ijarsct.co.in
IJARSCT

whereas Alexnet performs faster with per epoch with 2.2 minute per epoch and 98% accuracy.
Shitole et.al.[37] comared the performance of recognition system using Principal Component Analysis (PCA) and
Linear Discriminant Analysis (LDA). For the feature extraction three methods are used chain coding, edge detection
using gradient features and direction feature techniques. These are further reduced by using LDA. Classification of
characters is done by using SVM classifier and concluded that LDA has given better result than PCA. Bisht et.al.[41]
proposed a novel technique for offline handwritten modiﬁed character recognition. Two models were used, A single
CNN architecture and in the second model, he used double-CNN architecture for the recognition. Dataset of Hindi
consonants and Matras with acceptable accuracies was considered. Based on the results, it was concluded that duplicate
CNN formats provide better results than CNN single formats, as it uses a smaller number of output classes compared to
existing modes of modified characters in Devanagari text. The results also inferred that double CNN proves to be better
as compared to traditional feature extraction (like histogram of oriented gradients) and classification methods (like
SVM). Gurav et.al.[42] presented a system which works on a set of 29 consonants and one modifier. With the help of
Consecutive convolutional layers extracting higher-level features becomes easy. Here, character-wise segmentation is
done instead modifiers-wise segmentation, which is a standard approach.
IV. ANALYSIS AND DISCUSSION

Handwriting recognition has become one of the most intriguing research areas in today's world, which has evolved into
a combination of artificial intelligence and machine learning, which is discussed in this survey. It contributes
specifically to the development of a visual connection between humans and machines. The handwritten recognition is
used in everywhere for example documents, reading postal addresses, bank check amounts, and forms, etc. In this
survey, we chose a deep learning-based Handwritten Devanagari Character Recognition concept. There are
approximately 42 papers selected interrelated to the survey. The effectiveness of this survey is analyzed and compared
using various parameters and algorithms. In this section, the different kinds of algorithms, methods, based on the deep
learning in Character Recognition concept papers are discussed effectively.
4.1 Comprehensive Study

Below table 1 shows the comprehensive study of different techniques used for handwritten recognition.
Table 1 A comparative study of various papers.
Ref. Paper Name Preprocessing Classifier Used Data Size Accuracy (In
no. Techniques Percentage)
[1] Deep Learning Based Large-Resizing, Gray CNN 92 thousand images 98.47
Scale Handwritten Devanagariscaling, padding
Character Recognition
[2] Handwritten Devnagari numerals- PCA 9,800 94.25
recognition with higher accuracy
[3] A Novel SVM - basedOperations are Support vector Data samples are 82.04
Handwritten Tamil characterperformed on the machine (SVM) collected from
recognition system digitized image to method different writers on
enhance the A4 sized documents
quality of the and then scanned.
image.
[4] Fuzzy Model Based Recognition- Fuzzy Model 4750 samples 90.65
of Handwritten Hindi Characters Based
[5] A Bilingual OCR for Hindi-Scanned document Based on Principal 200000 characters 96.7
Telugu Documents and itsis Filtered and Component
Applications Binarized. Analysis followed
by support vector
classification
www.ijarsct.co.in
IJARSCT

[6] Online Handwritten CharacterNormalization, SVM, HMM Devanagari Script: 97.27
Recognition of Devanagari andSmoothing, The number of
Telugu Characters usingInterpolation examples used for
Support Vector training is 21780 and
that for testing is
2420. The number of
classes considered for
Devanagari script is
91. Telugu Script:
There are 253 classes
and a total of 33726
samples for training
and 4091 samples for
testing.
[7] Optical Character RecognitionBinarization, Rough Fuzzy HP Labs India Indic 99.8
Systems for Different Noise Removal, Multilayer Handwriting dataset
Languages with Skew Detection Perceptron
Soft Computing (RFMLP)
[8] A Complete Thresholding Tree Classifiers The OCR is tested 93
OCR for printed Hindi Text inmethod used for from various
Devnagari Script Binarization newspapers and
magzines.
[9] Feature extraction based on- Gaussian 2,000 92
moment invariants for Distribution
handwriting recognition
[10] Devnagari Ancient DocumentsAuto correction, MLP, Neural 6152 pre-segmented 88.95
recognition using statiticalBinarization, Local Network (NN), samples of Devanagari
feature extraction techniques / Global CNN, SVM, ancient documents
Thresholding Random Forest
[11] Character Segmentation in TextTraining data SVM - 99.6
line via CNN using weakly
labelled data
via P-N learning
[12] Visualizing and UnderstandingSplitting, Resizing Customised 80000 samples. Out of 94.93
Customized Convolutional CNN(CCNN) 80000 samples,
Neural Network for Recognition written
of Handwritten Marathi in Marathi, 70000
Numerals samples are used for
training and 10000,
for testing.
[13] An OCR system for Telugu - KNN and fringe Estimated to be of the 92
distance template order of 10,000
matching
[14] An efficient Devanagari characterNoise Removal, SVM Total 60 documents 99.54 for
classification in printed andSkew Detection, are considered, where printed
handwritten documents usingNormalization, 60% documents are images and
SVM Gray scaling, used in training and 98.35 for
Binarization 40% were used during handwritten
testing. images
www.ijarsct.co.in
IJARSCT

[15] Performance comparison of - SVM More than 25000 94.1
features on Devanagari handwritten
handprinted dataset Devanagari characters
[16] Recognition of Unconstraint- The combination Experiment with 20 86.5

On-Line Devanagari Characters of multiple different writers with
classifiers that each writer writing 5
focus on either samples of each
local online character in a totally
properties, or unconstrained way
global off-
lineproperties.
[17] Hybrid feature extractionSkeletonization, a Neural network Dongre database 93.4
algorithm for Devanagari script combination of (4000 samples)
structural features
of the character
like number of
endpoints,loops,
and intersection
points is
calculated,
quadratic curve-
fitting model
[18] Comparative study of devnagariGray scaling, PD (Projection ISI Kolkata database 95.19
handwritten character recognitionBinarization Distance), SVM,
using different feature and MQDF (Modified
classifiers Quadratic
Discriminant
Function),
MIL(Mirror Image
Learning),
ED(Euclidean
Distance),NN,
K-NN etc.
[19] A multi-scale deep quad tree- softmax classifier, HPL offline character 95.18
based feature extraction deep learning, database
method for the recognition multi-scale
of isolated handwritten characters convolutional
of popular Indic scripts neural network
(MMCNN)
[20] Combining multiple featureConversion of Classification 4900 samples 92.80
extraction techniques for Handwritten decision obtained
handwritten devnagari characterCharacter to from four Multi
recognition Bitmapped Layer Perceptron
Binary Images, (MLP)
Scaling of the
binary character
images

www.ijarsct.co.in
IJARSCT

[21] Classification Of GradientBinarization Multilayer 300 samples of 99.10 and
Change Features Using MLP for perceptron (MLP) handwritten characters 94.15
Handwritten Character neural networks recognition
Recognition rates on
training and
test sets
respectively
[22] Hindi handwritten characterHistogram based Regular 5000 samples 82.0
recognition using multiple global binarising expressions of
classifiers algorithm, strokes
Removal of
isolated dot near
header line
[23] A system for off-line ArabicBase Line Variants of Bayes IFN/ENIT Tunisian 90.02
handwritten word recognitionestimation classifier city names dataset
based on Bayesian approach ground- truth [31]
[24] Handwritten Hindi characterNormalization, KNN In-house dataset 90
recognition using curvelet noise containing 200 images
transform removal and gray of character set (each
scale conversion. image contains all
Hindi characters)
[25] Handwritten DevanagariBinarization, noise Convolutional Devanagari characters 91.23% for
Character Recognition using removal Neural Network and Devanagari Devanagari
Convolutional Neural Network numerals. characters
and 100% for
Devanagari
numerals.
[26] Devanagari HandwrittenXnConvert batch Deep 5800 isolated images 94.84%
Character Recognition using fine-processing, data Convolutional of 58 unique character testing
tuned Deep Convolutional Neuralaugmentation and Neural classes: 12 vowels, 36 accuracy with
Network on trivial dataset some Ne consonants and 10 training loss
regularization twork (DCNN) numerals. of 0.18
techniques like
Dropout and Batch
Normalization.
[27] Deep Learning Approach for Remove noise, Deep Learning Large set of 91.81
Devanagari Script Recognition converted to Approach handwritten
binary image, numerical, character,
resized to fixed vowel modifiers and
size of 30×40 and compound characters.
then convert to
gray scale image
using mask
operation, it blurs
the edges of the
images.

www.ijarsct.co.in
IJARSCT

[28] Recognition of handwrittenBinarization, ANN Classifier Fifty handwritten Recognition
Devanagari characters usingNoise Elimination characters from 10 of 50
machine learning approach. people resulting 500 individual
characters are used. handwritten
characters is
92% and for
handwritten
sentences,
accuracy
obtained is
88.25%.
[29] Machine recognition ofDigitized, cleaned, - A small set of data 90
Devanāgarī script thinned, consisting of
segmented (to Resembling characters
extract composite and upper and lower
characters),and signs
labeled
[32] Arabic handwriting recognitionBinarization, CNN Two dataset used. 97% and 88%
system using convolutionalNoise Elimination, 1. Hijja Dataset- a on the AHCD
neural network Tilted scanned new dataset of Arabic dataset and
papers were letters written the Hijja
manually rotated. exclusively by dataset,
children aged 7–12. respectively
The dataset contains
47,434 characters
written by 591
participants. 2.Arabic
Handwritten Character
Dataset (AHCD)
dataset
[34] Recognition of Off- LineDigitized, cleaned Quadratic 11270 samples of For
Handwritten Devnagari classifier-based Devanagari characters Devanagari
Characters Using Quadratic system (vowels as well as characters
Classifier consonants) from 80.36% and
different individuals for numerals
and digitalized them. 98.86%
[35] Handwritten Hindi CharacterSegmented first K-Nearest 200 in house images 90
Recognition Usingthen curvelet Neighbor classifier of character set (each
Curvelet Transform. features are using Curvelet image contains all
obtained by Transform Hindi characters)
calculating
statistics of thick
and thin images by
applying curvelet
transform.

www.ijarsct.co.in
IJARSCT

[36] Transfer Learning using CNN forFinetuning Pre trained model Handwritten 99
Handwritten Devanagari for Deep Devanagari characters
Character Recognition Convolution has 46 classes with
Neural Network 2000 images of each
class. We partitioned
the dataset of 92,000
images into a training
set of 78,200 images
(0.85) and a testing set
of 13,800 (0.15).
[37] Recognition of Filtering of the SVM classifier 47 handwritten 70.58
handwritten devanagariscanned image, Devanagari characters
characters using linear Grey scale are used of ISI dataset.
discriminant analysis conversion, Image size is 270x270
Dilation pixels each. Total
4700 images are used
for training (100
images per character).
[41] Offline handwritten DevanagariResizing of Double CNN Hindi consonants and 90.99
modified character recognitionimages and Matras dataset
using convolutional neuralconversion of
network samples from
colour to
greyscale.
[42] Devanagari HandwrittenGrey Scale Deep Convolution Self-made 34604 99.65
Character Recognition usingconversion, Edge Neural network handwritten images
Convolutional Neuraldetection, Noise used for Devanagari
Networks Removal script with no header
line (Shirorekha) over
them
4.2 Comparative Analysis

In this survey, approximately 80 papers are collected and 42 papers were taken for technical analysis. Each of the
papers is referred to as the category of deep learning. The papers are selected from different kinds of journals such as
Elsevier, Springer, IEEE, Conference, and others. The pie chart representation of papers selected from different kinds of
journals is described in Fig. 7. From this, 7% of papers are selected from Elsevier, 21% of papers are selected from
springer, 48% of papers are selected from IEEE, and 7% deep learning method papers are selected from the ACM
journals. Finally, the highest number of papers are chosen from the IEEE journal, and the papers are collected from the
character recognition and deep learning domain
The character recognition papers which focus on deep learning are selected from the year 2000 to 2021. In this survey,
we used approximately 42 papers for technical review. There are ten papers are chosen before the year 2010, four
papers from 2010-2015, seven papers from 2015-2018. The highest number of papers are selected in between 2019-
2021 that is approximately twelve papers are taken. The number of papers chosen from year is noted in Fig 8.

www.ijarsct.co.in
IJARSCT

7%
5%
7% IEEE
Springer
48% Elsevier
12%
ACM
Conference
Others
21%
Fig 7. Representation of Character Recognition papers chosen from different journals.
Number of Papers Vs
14 Years
12
Number of Papers
10
0
< 2010 2010 -2015 2015 -2018 2019 -2021
Years
Fig 8. Analysis of Character Recognition papers published every year.
16 15
14
Percentage Of Papers
12
10
7
8 6 6
6 4
3 3
4 2
2
0
SVM CNN Fuzzy KNN Bayes PCA Gaussian Other
Models
Deep Learning Models
Fig 9. Analysis of character recognition papers based on deep learning methods.

www.ijarsct.co.in
IJARSCT

In this paper, the Devanagari character recognition based on deep learning methods is categorized into eight classes
such as Convolutional Neural Network (CNN), Support Vector Machine (SVM), Fuzzy Model, K-Nearest Neighbor
(KNN), Bayes Theorem, Principal Component Analysis (PCA), Gaussian Distance, Others. We have selected 7% of
papers from SVM, 15% of papers from CNN, 3% of papers based on Fuzzy models, 6% papers from KNN learning, 3%
of papers based on Bayes learning, 2% of papers based on PCA learning and remaining papers are based on the concept
of common deep learning. Several papers belong to the recurrent neural network domain. The number of papers
published based on the deep learning types is represented in Fig 9.
V. CONCLUSION
Character Recognition is one of the most common applications in image processing. Due to complexities of Indian
languages, it has been recognized as one of the challenging researches is in the field of computer vision and pattern
recognition. But still, a lot of research is being done on large datasets of these languages to handle the complexities and
other issues.
This paper carries out a study on handwritten character recognition using deep learning with the help of various similar
kinds of papers. This paper also represents a survey of preprocessing techniques, various classifiers used and
recognition techniques for handwritten Devanagari character recognition. Deep learning techniques are commonly
performed for character recognition due to high tolerance and less errors. This survey paper helps researches and
developers to understand various techniques and the way they are implemented for recognition. The image is scanned
firstly, and then data is preprocessed. The preprocessing involves various techniques such as Binarization, removal of
noise and Normalization. After that, features are extracted from the preprocessed data so that the relevant data is further
used to train the model. Various classification techniques are applied and the best approach is considered based on
accuracy. Various models which uses multilayer perceptron compares input image with the trained set to get high
accuracy. This paper has focused on various approaches effective algorithm. The study concludes that SVM classifier as
well as CNN classifier both provides better results with an accuracy of 99.6% and 98.47% respectively. The study
points out that the work in Devanagari scripts still in progress, so further a lot can be done in this domain.
VI. FUTURE WORK

In future, the system has a great scope of research in the area of Devanagari word as well as sentence recognition. Also,
a large and complex dataset can be considered. Recognition of a full handwritten document through an OCR can be
done. Still a hybrid model could be proved effective and optimized from which highly accurate results can be obtained.
The system could be modified with other orthogonal moment features set Also, more work could be done on half
characters as well as missing part of Hindi character recognition. The model can also predict errors of the input. Various
other Indian scripts and languages can also be considered to make a generic system.
REFERENCES
[1]. Shailesh Acharya, Ashok Kumar Pant, Prashnna Kumar Gyawali. Deep Learning Based Large Scale
Handwritten Devanagari Character Recognition, 2015 9th International Conference on Software, Knowledge,
Information Management and Applications (SKIMA).
[2]. C. V. Lakshmi, R. Jain, and C. Patvardhan. Handwritten Devnagari numerals recognition with higher
accuracy, in Proc. Int. onfut.Intell. Multimedia Appl., 2007.
[3]. Shanthi N and Duraiswami K. A Novel SVM - based Handwritten Tamil character recognition system,
Springer Pattern Analysis & Applications,Vol-13, No. 2, 173-180,2010.
[4]. M. Hanmandlu, O. V. R. Murthy, and V. K. Madasu. Fuzzy Model based recognition of handwritten Hindi
characters, in Proc. Int. Conf. Digital Image Comput. Tech. Appl., 2007.
[5]. C. V. Jawahar, M. N. S. S. K. Pavan Kumar, S. S. Ravi Kiran. A Bilingual OCR for HindiTelugu Documents
and its Applications, Proc. of the 11th ICPR, vol. II, pp. 200-203, 1992.
[6]. C. Chandra Sekhar. Online Handwritten Character Recognition of Devanagari and Telugu Characters using
Support Vector Machines, Tenth International Workshop on Frontiers in Handwriting Recognition, Universit
éde Rennes 1, Oct 2006, La Baule (France). inria00104402.

www.ijarsct.co.in
IJARSCT

[7]. A. Chaudhuri. Optical Character Recognition Systems for Different Languages with Soft Computing, Studies
in Fuzziness and Soft Computing 352, DOI 10.1007/978-3-319-50252-6_8
[8]. Veena Bansala and R M K Sinha. A Complete OCR for printed Hindi Text in Devnagari Script ,IEEE 2002.
[9]. R. J. Ramteke and S. C. Mehrotra. Feature extraction based on moment invariants for handwriting
recognition, in Proc. IEEE Conf. Cybern. Intell. Syst., 2006
[10]. Sonika Narang, M K Jindal and Munish Kumar. Devnagari ancient documents recognition using statistical
feature extraction techniques , MS received 19 July 2018; revised 14 November 2018; accepted 4 March
2019; published online 13 May 2019.
[11]. Xiaohe Li, Xingming Zhang, Bin Yang, Siyu Xia. Character Segmentation in Text Line via Convolutional
Neural Network, 2017 4th International Conference on Systems and Informatics (ICSAI).
[12]. D. T. Mane , U. V. Kulkarni. Visualizing and Understanding Customized Convolutional Neural Network for
Recognition of Handwritten Marathi Numerals , IEEE Electron Device Lett., vol. 20, pp. 569–571, Nov.
1999.
[13]. A. Negi, C. Bhagvathi, and B. Krishna. An OCR system for telugu. In Proc. of ICDAR, 2001.
[14]. Shalini Puria, Satya Prakash Singh, An efficient Devanagari character classification in printed and
handwritten documents using SVM, International Conference on Pervasive Computing Advances and
Applications – PerCAA 2019.
[15]. S. Kumar, Performance comparison of features on Devanagari handprinted dataset, Int. J. Recent Trends, vol.
1, 2009.
[16]. S. D. Connel, R.M.K. Sinha and Anil K. Jain, Recognition of Unconstraint On-Line Devanagari Characters,
Proceedings of the International Conference on Pattern Recognition, 2000.
[17]. Khandja D., Nain N., and Panwara S.: Hybrid feature extraction algorithm for Devanagari script, ACM
Trans. Asian Low Resour. Lang. Inf. Process., 2015, 15, (1), p. 2:1– 2:10
[18]. Pal U., Wakabayashi T., and Kimura F.: Comparative study of devnagari handwritten character recognition
using different feature and classifiers. Int. Conf. Document Analysis and Recognition, Barcelona, Spain, July
2009, pp. 1111– 1115
[19]. Sarkhel R., Das N., and Das A. et al: A multi‐scale deep quad tree based feature extraction method for the
recognition of isolated handwritten characters of popular Indic scripts, Pattern Recognit., 2017, 71, pp. 78–
93
[20]. Arora S., Bhattacharjee D., and Nasipuri M. et al: Combining multiple feature extraction techniques for
handwritten devnagari character recognition. Third int. Conf. Industrial and Information Systems, Kharagpur,
India, December 2008, pp. 1– 6
[21]. S. Arora, D. Bhattacharjee, M. Nasipuri, L. Malik, Classification Of Gradient Change Features Using MLP
for Handwritten Character Recognition, Emerging Applications of Information Technology (EAIT), Kolkata,
India, 2006
[22]. Yadav M., and Purwar R.: Hindi handwritten character recognition using multiple classifiers. 7th Int. Conf.
Cloud Computing, Data Science & Engineering – Confluence, Noida, India, Jan 2017, pp. 149– 154
[23]. Khémiri, A., Echi, A. K., Belaïd, A., et al.: A system for off-line Arabic handwritten word recognition based
on Bayesian approach. Int. Conf. Frontiers in Handwriting Recognition (ICFHR), Shenzhen, China, October
2016, pp. 560–565
[24]. Verma G. K., Prasad S., and Kumar P.: Handwritten Hindi character recognition using curvelet transform.
Int. Conf. Information Systems for Indian Language, Patiala, India, 2011, pp. 224– 227
[25]. A. Mohite and S. Shelke, Handwritten Devanagari Character Recognition using Convolutional Neural
Network, 2018 4th International Conference for Convergence in Technology (I2CT), Mangalore, India, 2018,
pp. 1-4, doi: 10.1109/I2CT42659.2018.9057991.
[26]. Deore, S.P., Pravin, A. Devanagari Handwritten Character Recognition using fine-tuned Deep Convolutional
Neural Network on trivial dataset. Sādhanā 45, 243 (2020). https://doi.org/10.1007/s12046-020-01484-1
[27]. S. Prabhanjan and R. Dinesh, Deep learning approach for devanagari script recognition, Int. J. Image Graph.
17(3) (2017) 1750016. https://doi.org/10.1142/S0219467817500164

www.ijarsct.co.in
IJARSCT

[28]. Pankaj Kale, Arti V. Bang, and Devashree Joshi, Recognition Of Handwritten Devanagari Characters Using
Machine Learning Approach, International Journal of Industrial Electronics and Electrical Engineering,
ISSN: 2347-6982, Volume-3, Issue-9, pp. 48- 51, Sept.-2015.
[29]. R.M.K. Sinha and H.N. Mahabala, Machine recognition of Devanāgarī script, IEEE Transactions on
Systems, Man and Cybernetics, 9(8), 1979, pp. 435-441.
[30]. Support Vector Machines for Binary Classification, Available at:
https://www.mathworks.com/help/stats/support-vector-machines-for-binary-classification.html
[31]. Pechwitz, M., Maddouri, S. S., Märgner, V., et al.: ‘IFN/ENIT – database of handwritten Arabic words’.
Proc. 7th Colloque Int. Francophone sur l'Ecrit et le Document, Hammamet, Tunisia, October 2002, pp. 129–
136
[32]. Altwaijry, N., Al-Turaiki, I. Arabic handwriting recognition system using convolutional neural network.
Neural Comput & Applic (2020). https://doi.org/10.1007/s00521-020-05070-8
[33]. Bharath A and Madhvanath S 2009 Online handwriting recognition for Indic scripts. In: Guide to OCR for
Indic scripts, pp. 209–234 [34]. Sharma N., Pal U., Kimura F., Pal S. (2006) Recognition of Off-Line
Handwritten Devnagari Characters Using Quadratic Classifier. In: Kalra P.K., Peleg S. (eds) Computer
Vision, Graphics and Image Processing. Lecture Notes in Computer Science, vol 4338. Springer, Berlin,
Heidelberg. https://doi.org/10.1007/11949619_72
[34]. Verma G.K., Prasad S., Kumar P. (2011) Handwritten Hindi Character Recognition Using Curvelet
Transform. In: Singh C., Singh Lehal G., Sengupta J., Sharma D.V., Goyal V. (eds) Information Systems for
Indian Languages. ICISIL 2011. Communications in Computer and Information Science, vol 139. Springer,
Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-19403-0_37
[35]. N. Aneja and S. Aneja, "Transfer Learning using CNN for Handwritten Devanagari Character Recognition,"
2019 1st International Conference on Advances in Information Technology (ICAIT), Chikmagalur, India,
2019, pp. 293-296, doi: 10.1109/ICAIT47043.2019.8987286.
[36]. S. Shitole and S. Jadhav, "Recognition of handwritten devanagari characters using linear discriminant
analysis," 2018 2nd International Conference on Inventive Systems and Control (ICISC), Coimbatore, 2018,
pp. 100-103, doi: 10.1109/ICISC.2018.8398991.
[37]. Veena Bansal, R.M.K. Sinha, Segmentation of touching and fused Devanagari characters, Pattern
Recognition, Volume 35, Issue 4, 2002, Pages 875-893,ISSN 0031-3203, https://doi.org/10.1016/S0031-
3203(01)00081-4
[38]. Garg N.K., Kaur L., Jindal M.K. (2011) The Segmentation of Half Characters in Handwritten Hindi Text. In:
Singh C., Singh Lehal G., Sengupta J., Sharma D.V., Goyal V. (eds) Information Systems for Indian
Languages. ICISIL 2011. Communications in Computer and Information Science, vol 139. Springer, Berlin,
Heidelberg. https://doi.org/10.1007/978-3-642-19403-0_8
[39]. B. Dessai and A. Patil, "A Deep Learning Approach for Optical Character Recognition of Handwritten
Devanagari Script," 2019 2nd International Conference on Intelligent Computing, Instrumentation and
Control Technologies (ICICICT), Kannur, India, 2019, pp. 1160-1165, doi:
10.1109/ICICICT46008.2019.8993342.
[40]. Bisht, M., Gupta, R. Offline handwritten Devanagari modified character recognition using convolutional
neural network. Sādhanā 46, 20 (2021). https://doi.org/10.1007/s12046-020-01532-w
[41]. Y. Gurav, P. Bhagat, R. Jadhav and S. Sinha, "Devanagari Handwritten Character Recognition using
Convolutional Neural Networks," 2020 International Conference on Electrical, Communication, and
Computer Engineering (ICECCE), Istanbul, Turkey, 2020, pp. 1-6, doi:
10.1109/ICECCE49384.2020.9179193.

www.ijarsct.co.in

Paper 8863

Uploaded by

Copyright:

Available Formats

Paper 8863

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Paper 8863

Uploaded by

Copyright:

Available Formats

ISSN (Online) 2581-9429

Volume 2, Issue 4, June 2020

Review for Handwritten Devanagari Character

Keywords: Devanagari Script; Optical Character Recognition; Segmentation; Convolutional Neural

II. SOME CONCEPTS OF CHARACTER RECOGNITION

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 71

Volume 2, Issue 4, June 2020

Fig 2. Vowels with Corresponding Modifiers of Devanagari Script

Fig 3. Half Form of Consonants of Devanagari Script

2.2 Character Recognition

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 72

Volume 2, Issue 4, June 2020

2.3 Optical Character Recognition

2.4 Working Principle of OCR

Fig 4. Flow control for Hindi OCR System

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 73

Volume 2, Issue 4, June 2020

E. Feature Extraction, Detection and Error Correction

2.5 Types of Classifiers

Volume 2, Issue 4, June 2020

2.6 Architecture of CNN

Fig 5. Architecture Diagram of CNN

A. Support Vector Machine (SVM)

Fig 6. SVM Algorithm

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 75

Volume 2, Issue 4, June 2020

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 76

Volume 2, Issue 4, June 2020

IV. ANALYSIS AND DISCUSSION

4.1 Comprehensive Study

Volume 2, Issue 4, June 2020

Volume 2, Issue 4, June 2020

[16] Recognition of Unconstraint- The combination Experiment with 20 86.5

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 79

Volume 2, Issue 4, June 2020

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 80

Volume 2, Issue 4, June 2020

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 81

Volume 2, Issue 4, June 2020

4.2 Comparative Analysis

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 82

Volume 2, Issue 4, June 2020

Fig 7. Representation of Character Recognition papers chosen from different journals.

Fig 8. Analysis of Character Recognition papers published every year.

Fig 9. Analysis of character recognition papers based on deep learning methods.

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 83

Volume 2, Issue 4, June 2020

VI. FUTURE WORK

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 84

Volume 2, Issue 4, June 2020

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 85

Volume 2, Issue 4, June 2020

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8863 86

You might also like