Computer Science Review 38 (2020) 100302

Contents lists available at ScienceDirect

Computer Science Review

journal homepage: www.elsevier.com/locate/cosrev

Survey

Offline recognition of handwritten Indic scripts: A state-of-the-art

survey and future perspectives
∗
Reya Sharma , Baijnath Kaushik
Computer Science and Engineering Department, SMVDU, J&K, India

article info a b s t r a c t

Article history: The handwritten script recognition is one of the most interesting and challenging areas of pattern
Received 29 April 2020 recognition due to numerous variations in writing styles. Extensive in-depth research work is reported
Received in revised form 19 June 2020 on the recognition of handwritten text in scripts such as Latin, Chinese, Arabic and Japanese. However,
Accepted 26 August 2020
the work reported on handwritten Indic scripts is still in its infancy, so significant research is required
Available online xxxx
in this field. This paper aims to describe various advancements reported over the last few decades
Keywords: in the field of handwritten Indic scripts recognition by analysing several existing state-of-the-art
Indic scripts studies. This comprehensive survey presents a transparent panorama of various feature extraction and
Handwritten character recognition classification techniques for the offline recognition of handwritten Indic scripts. The most important
Pattern recognition part of this survey is to systematically present the reported works on handwritten Indic scripts
Feature extraction
like Devanagari, Bengali, Gurumukhi, Kannada, Telugu, Gujarati, Oriya, Tamil and Malayalam. After
Classification techniques
exploring the reported works, an analysis is performed based on the findings. Several issues and
challenges related to the recognition of Indic scripts are discussed, which indicates some future
research prospects. Based on the extensive study conducted in this article, it has been contemplated
that there is a need to develop hybrid feature extraction and classification approaches for achieving
the most accurate results. So, a novel framework based on improved particle swarm optimization
(PSO) algorithm to automatically construct optimal convolutional neural network (CNN) architecture
has been proposed with an aim to outperform the existing techniques.
© 2020 Elsevier Inc. All rights reserved.

Contents

1. Introduction......................................................................................................................................................................................................................... 2
1.1. Motivation .............................................................................................................................................................................................................. 3
1.2. Application areas ................................................................................................................................................................................................... 3
1.3. Main contributions of the article......................................................................................................................................................................... 3
1.4. Outline of the article ............................................................................................................................................................................................. 3
2. Background.......................................................................................................................................................................................................................... 4
2.1. Evolution of Indic scripts ...................................................................................................................................................................................... 4
2.2. Writing system....................................................................................................................................................................................................... 4
2.3. Classification of Indic scripts ................................................................................................................................................................................ 4
2.4. Peculiarities and challenges in Indic scripts....................................................................................................................................................... 5
3. Related surveys ................................................................................................................................................................................................................... 6
4. Survey protocol................................................................................................................................................................................................................... 6
4.1. Planning the survey............................................................................................................................................................................................... 6
4.2. Research questionnaire ......................................................................................................................................................................................... 6
4.3. Information sources............................................................................................................................................................................................... 7
4.4. Search criteria ........................................................................................................................................................................................................ 7
4.5. Quality assessment ................................................................................................................................................................................................ 7
4.6. Data extraction....................................................................................................................................................................................................... 8
5. Suvey of Indic scripts datasets ......................................................................................................................................................................................... 8

∗ Corresponding author.
E-mail address: [email protected] (R. Sharma).

https://doi.org/10.1016/j.cosrev.2020.100302
1574-0137/© 2020 Elsevier Inc. All rights reserved.
2 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

5.1. Devanagari standardized datasets ....................................................................................................................................................................... 8

5.2. Bangla standardized datasets ............................................................................................................................................................................... 8
5.3. Telugu standardized datasets ............................................................................................................................................................................... 8
5.4. Tamil standardized datasets ................................................................................................................................................................................. 8
5.5. Kannada standardized datasets ............................................................................................................................................................................ 8
5.6. Oriya standardized datasets ................................................................................................................................................................................. 8
6. State-of-the-art in handwritten Indic scripts OCR ......................................................................................................................................................... 9
6.1. Reported work on Devanagari, Gurumukhi and Bengali scripts...................................................................................................................... 9
6.1.1. Feature extraction .................................................................................................................................................................................. 9
6.1.2. Classification and recognition ............................................................................................................................................................... 11
6.1.3. Some observations on handwritten Devanagari, Gurumukhi and Bengali scripts ......................................................................... 12
6.2. Reported work on Kannada and Telugu scripts................................................................................................................................................. 13
6.2.1. Feature extraction .................................................................................................................................................................................. 13
6.2.2. Classification and recognition ............................................................................................................................................................... 14
6.2.3. Some observations on handwritten Kannada and Telugu
scripts ...................................................................................................................................................................................................... 15
6.3. Reported work on Gujarati, Oriya, Tamil and Malayalam scripts.................................................................................................................... 15
6.3.1. Feature extraction .................................................................................................................................................................................. 15
6.3.2. Classification and recognition ............................................................................................................................................................... 16
6.3.3. Some observations on handwritten Gujarati, Oriya, Tamil and Malayalam scripts....................................................................... 17
7. Analysis of work done on handwritten Indic sripts OCR .............................................................................................................................................. 18
8. Challenges and future perspectives.................................................................................................................................................................................. 19
9. Proposed framework .......................................................................................................................................................................................................... 19
9.1. Representing CNN architectures with swarm initialization ............................................................................................................................. 20
9.2. Evaluate the fitness ............................................................................................................................................................................................... 20
9.3. Velocity computation ............................................................................................................................................................................................ 21
9.4. Update the particle ................................................................................................................................................................................................ 22
10. Experimental analysis ........................................................................................................................................................................................................ 22
10.1. Benchmark datasets............................................................................................................................................................................................... 22
10.2. Experimental setup and overall performance .................................................................................................................................................... 22
10.3. Comparison with the state-of-the-art ................................................................................................................................................................. 23
11. Conclusion ........................................................................................................................................................................................................................... 23
Declaration of competing interest.................................................................................................................................................................................... 23
Appendix A. Quality assessment form ............................................................................................................................................................................. 24
Appendix B. Data items extracted from all papers ........................................................................................................................................................ 24
Appendix C. List of abbreviations..................................................................................................................................................................................... 24
References ........................................................................................................................................................................................................................... 24

pre-processed by applying a series of operations on the input

1. Introduction image to convert it into a more effective form that is suitable
for further processing. Pre-processing constitutes removal of var-
Optical character recognition (OCR) is the electronic conver- ious distortions, orientation issues, noise reduction, slant and
sion of digitized or optically scanned pages of printed or hand- skew correction, binarization, normalization, and several other
written text into a machine-processable format. OCR is one of operations.
the most challenging and interesting fields of pattern recognition In the segmentation stage, the paragraph of text is segmented
having a wide variety of practical applications. The OCR was into lines, then lines into words or ligatures [7]. The segmentation
initially developed by Carley in 1870 with the invention of a approach is categorized into two types on the basis of whether to
retina scanner, which is an image transmission system [1]. segment the words/ligatures or not.
The recognition accuracy of printed documents is a potentially
• Holistic approach or segmentation-free approach
solved problem. However, for handwritten documents, there are
numerous variations in writing styles due to which the task of
• Analytical approach or segmentation-based approach
character recognition for handwritten documents becomes quite The first category includes a segmentation-free approach or
difficult [2–4]. Thus, the handwritten script recognition is con- holistic approach, and the second category has a segmentation-
sidered potentially an active area of research. This process of based approach or analytical approach. In a holistic approach,
character recognition is based on a variety of several distinct the words are not segmented into characters, and the recognition
tools and techniques, which in turn determines the recognition system identifies the ligature as an entire word. However, in an
accuracy of the character recognition system. analytical approach, the words or ligatures are segmented into
The OCR is a computation intensive field that has witnessed individual characters or strokes [8,9]. This analytical approach
significant improvements over these years. These improvements is classified further into two sub-categories, viz. direct segmen-
are mainly achieved, with tremendous advancements in machine tation and indirect segmentation. In the direct segmentation,
learning and computation intensive algorithms. The character a ligature or word is segmented into characters using several
recognition system operates with an aim to replicate human heuristics. Whereas, in indirect segmentation the ligature is seg-
reading ability by maintaining accuracy at a far higher speed. mented into primitives by dividing it into several smaller parts or
The character recognition system works in several stages [5,6] as strokes which might be characters or less than characters, such as
shown in Fig. 1. sub-characters or small-strokes.
The first stage of the recognition process deals with data After segmentation, the next stage is feature extraction, which
acquisition, which transforms handwritten documents into a dig- is one of the most exigent and challenging stages in pattern
italized form. Then in the next step, digitalized documents are recognition problems. It deals with the extraction of pertinent
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 3

Fig. 1. Stages of character recognition process.

and unique patterns that reduce the data for recognition and • Publishing houses
thereby enhance the recognition power [10]. The success rate • Reservation counters
of the classifier strongly depends upon the extracted features. • Forensic document analysis
Depending on the extracted features, the final classification and • Post offices
recognition phase is the prime decision-making stage. The pat-
terns are finally recognized on the basis of input features fed to 1.3. Main contributions of the article
the classifier.
• The most important aspect of this article is to present the
1.1. Motivation work reported in previous years (2000–2019) on the recog-
nition of handwritten Indic scripts by exploring more than
The electronic conversion of printed or handwritten text into 100 articles from reputed journals and renowned confer-
a computer-readable form is known as OCR. The problem of ences.
handwritten text recognition is more interesting and fascinating • It comprehensively outlines the peculiarities and challenges
as compared to printed text recognition due to the presence of major Indic scripts and provides a comparative study of
of uneven variations in handwriting style with respect to the our survey with other related surveys and review articles.
writers, content, and time. The handwriting of a person is always • This article illustrates a survey protocol that depicts the
unique, just like our fingerprints, and this uniqueness creates planning of the survey, research questionnaire, information
motivation and interest among the researchers to work in this sources, search criteria, quality assessment process, and data
exigent and challenging field. extraction.
Various miscellaneous languages are used around the world. • This paper scrutinizes various eminent datasets available for
Many languages have disappeared as their usage is limited and research in handwritten Indic script recognition.
due to their presence in rural or geographically inaccessible parts • This article exhaustively surveys numerous feature extrac-
of the globe. So, at this point, it is highly recommended to use tion as well as classification techniques for the recognition
technologies like OCR and natural language processing to stop the of handwritten Indic scripts.
extinction of languages in the world. There are almost 7000 lan- • An analysis is performed on the basis of findings and re-
guages in the world (https://www.ethnologue.com/guides/how- lated work, which highlights various research challenges
many-languages) and handwritten OCR systems are available and future directions that need to be considered for efficient
only for some of them. OCR systems are mostly available for handwritten script recognition.
languages that are of huge importance and strong economic value • Finally, a proposed framework is given with an aim to over-
like Chinese, Latin, Arabic and Japanese. Most of the languages come the drawbacks of existing classification algorithms.
derived from Indic script appear to be at the risk of vanishment An improved PSO algorithm to automatically construct op-
due to the absence of efforts. So, there is an immense need for timal CNN architecture has been proposed as illustrated in
character recognition related research for Indic scripts. With the Algorithm 1.
prevalent use of computers in homes and organizations, auto-
matic paper documents processing is gaining huge importance, 1.4. Outline of the article
which leads us to a paperless environment and the development
of the OCR system for regional languages. So, in this paper, The remainder of this paper is organized into following sec-
we have made a sincere attempt to perform a systematized tions. Section 2 gives the introduction and evolution of major
state-of-the-art survey for problems that have been discussed Indic scripts, along with their peculiarities and challenges. Sec-
above. tion 3 provides the comparison of present survey with other
traditional surveys and review articles. Section 4 discusses the
1.2. Application areas survey protocol including the plan, research questionnaire, infor-
mation sources, search criteria, etc. which provides a cognisance
The handwritten character recognition has numerous real-life assistance to the researchers to perform the survey. Section 5
commercial and practical applications which provide an enthu- illustrates the detailed study of standardized datasets available
siasm to the researchers to explore this field. Some of these for several Indic scripts. Section 6 exhaustively explains and
application areas are listed below: compares the work done on different feature extraction and
• Banks classification techniques for the recognition of handwritten Indic
• Libraries script. The classification techniques are divided into three broad
• Historic data analysis categories, namely neural network based techniques, SVM based
4 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

Fig. 2. Evolution of Indic scripts [11].

for writing these languages [12]. Scripts basically depict the writ-
ing system for the language that indicates sounds and represents
phonetics of the language.

2.1. Evolution of Indic scripts

In Ancient India, Brahmi is the oldest writing script and ma-

jority of Indian scripts are derived from the ancient Brahmi script
through several transformations [13,14]. These transformations
result in the evolution and development of different Indian re-
gional scripts as shown in Fig. 2.

2.2. Writing system

An Alphasyllabary or Abugida writing system is used for In-

dic scripts. It is a segmental writing system that consists of
consonants and vowel arrangements. These consonant-vowel ar-
rangements are written as a single unit, in which consonant letter
acts as a base and vowel symbol is secondary. When vowels
appear on their own or at the starting of a word, then they are
written as inherent separate letters.
Fig. 3. Handwritten vowels belonging to several Indic scripts.

2.3. Classification of Indic scripts

techniques and miscellaneous techniques. Section 7 summarizes The Brahmi script is categorized into two major groups, one
the analysis investigated on the basis of findings and related constituting the north Indian scripts and the other consisting of
work. Section 8 highlights the challenges in Indic script charac- south Indian scripts [15]. A single script may be used for writing
ter recognition and provides future research prospects. A novel more than one language. For example, Devanagari script which is
framework for handwritten Indic script character recognition has the most widely used Indic script [16] is used for writing Hindi,
been proposed in Section 9. The experiment analysis and com- Dogri, Sanskrit, Kashmiri, Marathi, Sindhi, Nepali, and Konkani.
parison with state-of-the-art works are presented in Section 10. After Devanagari, the most widely used script is Bengali and
Finally, Section 11 concludes the paper. it is used for writing Maithili, Bengali, Manipuri, and Assamese
languages. Table 1, presents detailed information including lan-
2. Background guages, category and domain for several Indic scripts.

• Northern Group (Indo-Aryan):

India is a multilingual and multiscript country. There are The northern group resembles much more to the origi-
twenty two official languages in India accepted by the Indian con- nal ancient Brahmi script. This group contains scripts like
Devanagari, Gurmukhi, Bengali, Gujarati, Oriya, Manipuri.
stitution, namely Hindi, Sanskrit, Dogri, Urdu, Punjabi, Kashmiri,
• Southern Group (Dravidian):
Bengali, Odia, Kannada, Tamil, Nepali, Gujarati, Marathi, Bodo,
The southern group is originated from ancient Brahmi script
Konkani, Malayalam, Assamese, Santali (Santhali), Manipuri, Tel-
with several transformations and hence, it looks completely
ugu, Maithili and Sindhi. Twelve major scripts, namely, Devana-
different from north Indian group. This group contains
gari, Gurmukhi, Bengali, Urdu (Perso-Arabic), Kannada, Gujarati, scripts like Kannada, Telugu, Tamil, and Malayalam.
Tamil, Malayalam, Assamese, Telugu, Oriya and English are used
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 5

Fig. 4. Handwritten consonants belonging to several Indic scripts.

Table 1 Table 2
Description and related information of Indic scripts. Comparative analysis of our survey with other related surveys.
Script Languages Script category Domain Topics Our survey [16] [17] [18] [11] [19] [12]
Devanagari Hindi, Sanskrit, Indo-Aryan/ Many States (2020)
Rajasthani, etc. Northern-group Motivation for Research ✓
Gurumukhi Punjabi Indo-Aryan/ Punjab Application areas ✓ ✓ ✓ ✓ ✓ ✓
Northern-group Sources of information ✓
Bengali Bengali, Assamese Indo-Aryan/ Tripura, West Research questions ✓
Northern-group Bengal, Assam Evolution of Indic scripts ✓ ✓
Gujarati Gujarati Indo-Aryan/ Gujarat Peculiarities and challenges ✓ ✓ ✓
Northern-group in Indic scripts
Oriya Odia Indo-Aryan/ Orissa Survey of feature extraction ✓ ✓ ✓ ✓ ✓ ✓ ✓
Northern-group techniques
Takri Dogri, Chambeali Indo-Aryan/ J&K, Himachal Survey of classification ✓ ✓ ✓ ✓ ✓ ✓ ✓
Northern-group Pradesh methods
Comparative analysis based ✓ ✓ ✓ ✓ ✓
Manipuri Meithei Indo-Aryan/ Manipur on accuracy
Northern-group Synthesis analysis based on ✓
Kannada Kannada Dravidian/ Karnataka findings
Southern-group Challenges and future ✓ ✓ ✓ ✓ ✓ ✓
Telugu Telugu Dravidian/ Andhra Pradesh directions
Southern-group Proposed framework ✓
Tamil Tamil Dravidian/ Tamil Nadu
Southern-group
Malayalam Malayalam Dravidian/ Kerala
Southern-group • In addition to compound characters, there are also modi-
fied characters in most Indian scripts. The vowel remains
in its original shape when it appears at the beginning of
2.4. Peculiarities and challenges in Indic scripts the word, however when a vowel follows a consonant it
modifies its shape and thus known as modified character or
• The character set in most Indic scripts consists of basic diacritics [18]. The modified character may be placed on the
characters and compound characters. The basic characters top, bottom, left or right of the consonant. The character set
are simply the collection of consonants and vowels, however of major Indic scripts is shown in Figs. 3 and 4.
compound characters are formed by the combination of two • Most Indian scripts are written from left to right [22]. How-
or more basic characters [20,21]. The compound characters ever, some modified characters may not follow this left to
are usually more complex in shape than the basic character. right writing sequence. In Indic scripts there is no concept
However, some Indic scripts like Gurumukhi and Tamil do of upper case and lower-case characters. Unlike other Indic
not have compound characters. scripts the alphabetic sequence in Urdu script is from right
6 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

Table 3
Research questionnaire with motivation.
S. No. Research questions Motivation
RQ1 What is handwritten OCR? To know about suitable approaches for handwritten text recognition.
RQ2 Application areas where handwritten OCR systems are most To know about the advantages and usefulness in real-life particle and
suited? commercial applications to make handwritten OCR systems more demanding.
RQ3 What are the different approaches used for building To explore and study techniques as well as literature available for
handwritten OCR model for Indic scripts? handwritten OCR systems for Indic script.
RQ4 Identify the number of Indic scripts in which handwritten OCR To study and identify that for many Indic scripts, handwritten OCR systems
systems are successful? have been developed successfully.
RQ5 Availability of standardized datasets for the research? It provide a detailed study of benchmark and standardized datasets available
for Indic scripts.
RQ6 What are the best feature extraction techniques and It urges to exhaustively explore and study various feature extraction
classification algorithms for handwritten Indic script OCR? techniques and classification algorithms for handwritten Indic script
recognition.
RQ7 Work done in major Indic scripts? It motivates towards a comparative study for major Indic scripts (Devanagari,
Gurumukhi, Bengali, Kannada, Telugu, Gujarati, Oriya, Tamil and Malayalam)
and reports the observations made on reviewed work.
RQ8 Does the accuracy of proposed system depend upon the size Experimental analysis has to be conducted in order to evaluate and compare
of dataset? the recognition accuracies of handwritten characters with specific algorithm.
RQ9 Does the accuracy increase with the use of hybrid technique To study and investigate the experimental studies using novel hybrid
for feature extraction and classification? techniques and analyse the comparison on the basis of recognition accuracies.

to left. The Urdu script is related to Perso Arabic group of

scripts [23]. Like other Indian scripts, the concept of upper
case and lower-case letters is also absent in Urdu script, but
this script is much more calligraphic as compared to other
Indic scripts.
Fig. 5. Zones in Indic script (Devanagari) with shirorekha/headline.
• One of the interesting characteristics of Indic scripts is that
characters have a horizontal line called Matra or Shirorekha
at the upper part [24]. The presence of this horizontal line
at each character results in the formation of a word with
a fully connected bar at the top. Thus, each word consists
of characters written together without any gap so that the
horizontal bar at the top is unbroken.
• Indic scripts may be categorized into two classes depending Fig. 6. Zones in Indic script (Gujarati) without shirorekha/headline.
upon the presence or absence of Matra. In Indic scripts
namely, Devanagari, Gurumukhi and Bengali, the characters
have Shirorekha or Matra on the top. However, Indic scripts state-of-the-art survey to analyse and consolidate the existing
like Gujarati, Oriya, Kannada, Telugu, Tamil, Malayalam and studies. Table 2 summarizes the comparative study of our survey
Assamese the characters do not have Shirorekha or Matra. with other related surveys and review articles.
The task of segmentation becomes quite difficult in Indic
scripts due to the existence of Matra [25,26] because it 4. Survey protocol
becomes necessary to segment characters from word in
order to get recognized by the OCR. For this segmentation, Survey Protocol depicts the comprehensive layout to investi-
an additional task of eliminating the headline from word has gate the plan related to various tasks like inventory and monitor-
to be done first so as to obtain individual characters. ing. The survey protocol also describes the guidelines and set of
• In all Indic scripts, the text line is partitioned into three rules for carrying out a systematic survey on the existing state-of-
different zones namely upper zone, middle zone and lower the-art studies on handwritten Indic scripts recognition. It gives
zone. The upper zone corresponds to the portion lying above cognisance and guidance to the novice researchers working in
the headline or shirorekha, the middle zone represents the this domain by providing sufficient details about the nature of
portion lying between the baseline and the headline i.e. the the OCR system for handwritten Indic scripts. It includes various
main body of character and finally, the lower zone corre- sub-sections like planning, research questionnaire, information
sponds to the portion lying below the baseline [27]. In some sources, search criteria, quality assessment, and finally the data
Indic scripts, where the headline is absent (like Gujarati, extraction phase.
Tamil, Oriya, Kannada, Malayalam) the upper zone is repre-
sented by portion lying above the meanline and similarly, 4.1. Planning the survey
the middle zone is represented by portion lying between
the meanline and the baseline. The baseline and meanline The planning of a systematic survey initiates with identify-
are imaginary lines, where the majority of lower-most and ing the purpose of studying handwritten OCR for Indic scripts.
upper-most points of characters lie in the text line. Fig. 5 Stages involved in this survey include building a survey frame-
represents the zones in Indic scripts with shirorekha or work, analysing research papers obtained from various online
headline and Fig. 6 represents the zones in Indic scripts libraries, interpreting the survey results, and finally identifying
without shirorekha i.e. on the basis of mean line. the research challenges and future directions.

3. Related surveys 4.2. Research questionnaire

The research on handwritten Indic script character recognition Research questionnaire form the basic structure of a research
is still at its infancy, and this leads to the necessity of methodical paper, project, survey, or study. They provide a clear focus and
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 7

Fig. 7. Demonstration of search criteria to obtain relevant articles.

Table 4 publication year, dataset, feature extraction techniques and clas-

Information sources used for the survey. sifiers. Several keywords like handwritten character recognition,
Sources Indic script feature extraction, Indic script classification, etc. are
Association for Computing Machinery (www.acm.org) used for the initial selection. Almost 1600 articles are obtained
arXiv Cornell University (www.arxiv.org) after exploring the web. Then further refinement is done by
Elsevier (www.elsevier.com) removing the articles on the basis of title, abstract and number
Google Scholar (www.scholar.google.com)
of citations. Final exclusion is made by reading the articles, and
IEEE Xplore (www.ieee.org)
InderScience (www.inderscience.com) a total of 126 relevant articles, are chosen for this comprehen-
ScienceDirect (www.sciencedirect.com) sive survey. The demonstration of search criteria for obtaining
Springer (www.springer.com) relevant articles is shown in Fig. 7.
Taylor and Francis (taylorandfrancis.com)
Wiley online library (www.onlinelibrary.wiley.com) 4.5. Quality assessment
Books and other online sources
The objective of quality assessment phase is to determine
which research article has to be included in the survey and which
purpose to the study. The main objective behind the formulation has to be removed on the basis of qualitative analysis. After
of research questions is to guide the reader about what this obtaining the relevant articles by following the above mentioned
methodical survey intends to answer. The research questionnaire search criteria, the next step is to check the quality of selected
is drafted in such a way that the field of OCR is completely articles through a quality assessment form (Appendix A). In or-
covered. Table 3 elucidates the research questions corresponding der to perform this quality assessment, we need some standard
protocols. According to the quality assessment protocol, our study
to handwritten OCR for Indic scripts.
must include:
4.3. Information sources • High quality research articles.
• Standardized and benchmark datasets.
Research papers belonging to Scopus journals, Science Cita-
• Comprehensive review, state-of-the-art survey and experi-
tion Index (SCI) journals and flagship conference proceedings are
mental studies related articles.
considered for the study. Apart from these various books, thesis
• Confined to the research topic.
and other online sources and materials are also used to do the
• Approach and techniques used for handwritten Indic script
survey work more widely and comprehensively. These sources of
recognition.
information are summarized in Table 4.
• The study must clearly mention the type and method of
4.4. Search criteria character recognition.
• Enough data to perform the comparative analysis.
Search criteria begin with selecting the papers related to char- • Important parameters (like recognition accuracy) for com-
acter recognition. The research articles are categorized based on paring the results obtained from different studies.
8 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

4.6. Data extraction • DHCD: Devanagari Handwritten Character Dataset (DHCD)

[33] is a UCI machine learning repository consisting of 46
The quality assessment process acts as a base for the data ex- classes of handwritten Devanagari characters images with
traction phase. The data extraction criteria illustrate that find sig- each class having 2000 examples. Therefore, the total num-
nificant and relevant data items (Appendix B) from the selected bers of instances in DHCD are 92,000 character images.
research articles which are to be studied, compared, analysed and
interpreted. Since the extraction of all necessary data items from
various studies was a complicated task. Due to this difficulty, 5.2. Bangla standardized datasets
several authors were contacted whenever required in order to
find out the in-depth details of the research. The following steps • CMATER: The CMATERdb 3.1.1 [34] is a Bangla numeral
were performed to carry out the data extraction process in our database, it contains 4000 Bangla digit images specified for
survey: training and 2000 digit images specified for testing. The
• Firstly, all the articles were surveyed, and then relevant data CMATERdb 3.1.2 [35] is a Bangla basic character dataset, it
were extracted from the initial studies. consists of 12,000 training samples and 3000 test samples.
• Then the consistency of extracted data was cross-checked The CMATERdb 3.1.3 [36] is a Bangla compound character
and verified by another author.
dataset, it consists of 34,229 training samples and 8468 test
• If any conflicts appeared during the cross verification, then
samples. The CMATERdb 3.1.4 [37] is a Bangla modified char-
the issues were resolved by calling a meeting.
acter dataset, containing of 1637 training samples and 407
The significant data items extracted in this survey are: test samples. The CMATERdb 2.1.3 [29] dataset consists of
18,931 Bangla word images and CMATERdb 1.1 [38] dataset
• Bibliographic data including author, year of publication, title consists of handwritten Bangla text lines extracted from 100
and source was collected. document images.
• Journal name and conference proceeding title were clearly • ISI-HBND: The ISI Handwritten Bangla Numeral Database
mentioned.
(ISI-HBND) [39] developed at Indian Statistical Institute, con-
• Aim of the study.
sists of 23,392 handwritten Bangla numeral samples written
• Type of Indic script.
by 1106 different writers.
• Dataset size.
• Feature extraction technique, classification method and
recognition accuracy. 5.3. Telugu standardized datasets
• Test errors and identification rate attained from experimen-
tal analysis. • CMATER: The CMATERdb 3.4.1 [28] is a Telugu numeral
• Carefully examine the research challenges obtained by anal- dataset, it contains 4000 training samples and 2000 test
ysis various studies. samples.
• Future directions are discussed. • HPL: In HPL Offline Handwritten Telugu Isolated Charac-
ter database (HPL-iso-telugu-char) [40], the training dataset
5. Suvey of Indic scripts datasets
contains 34,325 handwritten character images and the test-
Database plays a very significant role in the evaluation of a ing dataset contains 10,829 handwritten character images.
new application model. The training and testing of a robust char-
acter recognition system require a sufficiently large and exhaus- 5.4. Tamil standardized datasets
tive dataset. In recent years, a number of handwritten datasets
have been created for several Indic scripts. This section provides
• HPL: The HPL Offline Handwritten Tamil Isolated Charac-
a detailed study of various benchmark and standardized datasets
ter database (HPL-iso-tamil-char) [41] has 70,000 handwrit-
for Indic scripts.
ten character images for training and 12,000 handwritten
5.1. Devanagari standardized datasets character images specified for testing.

• CMATER: Center for Microprocessor Applications for Train- 5.5. Kannada standardized datasets
ing Education and Research (CMATER) in Jadavpur Univer-
sity, Kolkata, India has created various database repository
for major Indic scripts OCR. The CMATERdb 3.2.1 [28] is a
• KHTD: The Kannada Handwritten Text Dataset (KHTD) [42]
Devanagari numeral database, it consists of 2000 training consists of 4298 text lines and 26,115 words, gathered from
samples and 1000 test samples. The CMATERdb 2.2.3 [29] 204 different handwritten documents written by almost 51
dataset consists of 15,528 Devanagari word images. The native Kannada writers.
CMATERdb 1.4 [30] dataset has 150 document images with
handwritten Devanagari text lines. 5.6. Oriya standardized datasets
• ISI-HDND: The ISI Handwritten Devanagari Numeral
Database (ISI-HDND) [31], consists of 22,556 handwritten
• ISI-HOND: The ISI Handwritten Oriya Numeral Database (ISI-
Devanagari numeral samples written by 1049 different writ-
HOND) [39] consists of 5970 handwritten Oriya numerals
ers. These samples are scanned optically from three types
of handwritten documents like job application forms, postal collected from 356 writers and prepared by using 300 dpi
mail and special set of forms designed by the collectors. flatbed HP scanner.
• HPL: The HPL Offline Handwritten Devanagari Isolated Char- • IITBBS: A dataset named ‘‘IITBBS’’ [43] developed at IIT
acter database (HPL-iso-dev-char) [32], consists of 18,192 Bhubaneswar consists of 35,000 handwritten Oriya charac-
handwritten character images specified for training and ter images written by 500 native Oriya writers. The images
3894 handwritten character images specified for testing are collected by using optical scanner with a resolution of
written by more than 100 native. speakers of Devanagari. 300 and 600 dpi.
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 9

Table 5
Detailed description of standardized handwritten datasets for various Indic scripts.
Dataset name (Reference) Script Dataset type Dataset size
ISI-HDND [31] Devanagari Numerals 22,556
CMATERdb 3.2.1 [28] Devanagari Numerals 3000
HPL-iso-dev-char [32] Devanagari Characters 29,970
DHCD [33] Devanagari Characters 92,000
CMATERdb 2.2.3 [29] Devanagari Words 15,528
CMATERdb 1.4 [30] Devanagari Text lines 150 document images
CMATERdb 3.1.1 [34] Bengali Numerals 6000
ISI-HBND [39] Bengali Numerals 23,392
CMATERdb 3.1.2 [35] Bengali Basic characters 15,000
CMATERdb 3.1.3 [36] Bengali Compound characters 42,697
CMATERdb 3.1.4 [37] Bengali Modified characters 2044
CMATERdb 2.1.3 [29] Bengali Words 18,931
CMATERdb 1.1 [38] Bengali Text lines 100 document images
CMATERdb 3.4.1 [28] Telugu Numerals 6000
HPL-iso-telugu-char [40] Telugu Characters 45,154
HPL-iso-tamil-char [41] Tamil Characters 82,000
KHTD [42] Kannada Text lines and words 4298 text-lines, 26,115 words
ISI-HOND [39] Oriya Numerals 5970
IITBBS [43] Oriya Characters 35,000

6.1. Reported work on Devanagari, Gurumukhi and Bengali scripts

6.1.1. Feature extraction

The feature extraction technique transforms large and redun-
dant input into a reduced collection of parameters called a feature
vector. Feature vectors hold great significance as they efficiently
and unambiguously recognize the patterns and thereby improve
the recognition rate of OCR [44,45]. Features are broadly classified
into two categories, namely statistical and structural features.
• Statistical Features: Statistical features are based on pixels
distribution in an image [46]. Zoning, moments, projection
profiles, histograms, crossing and distances are some of
the popular statistical feature extraction techniques [47].
Statistical features can be used for global or local feature
extraction depending upon the character image [48]. Global
Fig. 8. Statistical distribution for the availability of standard/benchmark datasets features are extracted from the entire character image. Local
for Indic scripts.
features, on the other hand, are obtained from the local
zones or neighbourhood of the character image. Statistical
features are also very useful for the dimension reduction
In addition to the datasets mentioned above, there are also process of the feature set.
inhouse datasets created by the researchers. Table 5 summa- • Structural Features: Structural features are related to the
rizes the standardized and benchmark datasets available for Indic topological and geometric structure of characters like inter-
scripts. The use of standardized datasets is essentially recom- sections, loops, branches, start points, extreme points, num-
mended, as they provide an efficient working environment for ber of endpoints, vertical lines, horizontal lines and many
comparing different research works. So it is worth mentioning more [49]. This section covers all the pertinent features ex-
that successful research needs more efforts for creating large, traction techniques for handwritten Devanagari, Gurumukhi
robust datasets for each of the Indic scripts. Fig. 8 depicts the and Bangla script.
availability of standardized datasets for various Indic scripts, anal-
ysis has been performed on the basis references used in this Among the earlier works, Sharma et al. [50] proposed a hand-
survey. written Devanagari character recognition scheme based on a di-
rectional chain code histogram. They obtained the features by
6. State-of-the-art in handwritten Indic scripts OCR observing the directional chain code information for each contour
point of characters. For each character, the bounding box was
In this work, the Indic scripts are categorized into three broad partitioned into 7 × 7 blocks and the histogram of chain code
categories or groups based on their evolution and resemblance. was computed for each block. On the basis of this chain code
histogram, 64 dimensional features were obtained and fed to the
• The first group comprises of Devanagari, Gurumukhi and recognition phase.
Bengali scripts, since all these scripts have shirorekha or Arora et al. [51] have extracted four different types of features:
matra on top of each character. shadow features, straight line fitting, chain code histogram and
• The second group is composed of Kannada and Telugu intersection points for the classification of handwritten Devana-
scripts, since the character sets of these two scripts are very gari characters. The shadow features were extracted globally
similar. The ancient Kannada script has been morphed into for the entire character image. On the other hand, line fitting
Telugu and Kannada scripts by 1500 AD, through several features, chain code histogram and intersection points were com-
transformations. puted by segmenting the character image into different blocks.
• Finally, the third group has Gujarati, Oriya, Tamil and Malay- Robert filter was used in [67] to obtain the gradient features,
alam scripts. and then a bi-quadratic interpolation technique was applied to
10 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

Table 6
Description of feature extraction techniques for Devanagari, Gurumukhi and Bengali scripts.
Features Category Author [Reference]
Shadow, CC, Intersection and Line fitting Structural features Arora et al. [51]
Longest run, Octant centroid, Modified shadow Statistical and structural features Basu et al. [52]
Longest run, Quad tree and Shadow Statistical and structural features Das et al. [53]
Curvelet coefficients Structural features Singh et al. [54]
Legendre and Zernike moment Statistical features Kale et al. [55]
Power curve fitting based features Statistical features Kumar et al. [56]
Topological Structural features Bag et al. [57]
Zone based centroid, Distance profile, BDD and CC Statistical and structural features Singh and Maring [58]
Pixel density features, Structural Statistical and structural features Shelke and Apte [59]
Scale invariant features transform Statistical features Surinta et al. [60]
Gradient and Curvature based composite feature Statistical and structural features Aggarwal and Singh [61]
Discrete cosine transform based features Statistical features Kumar et al. [62]
HOG, Projection profile Statistical features Yadav and Purwar [63]
LBP, Directional, Regional features Statistical and structural features Kumar and Gupta [64]
Automatic extraction using BornoNet Non-explicit features Rabby et al. [65]
Supervised Layerwise-DCNN Non-explicit features Jangid and Srivastava [66]

compute the curvature features. In [54] curvelet transform tech- The results indicate that GABN based feature extraction provides
nique was used for feature extraction. Curvelet coefficients were more promising results than GABM for handwritten Gurumukhi
computed for thick and thin character images. The fusion of these character recognition.
curvelet coefficients computed by inverse Fast Fourier Transform Kumar et al. [56] proposed two curve fitting based feature
acts as features. extraction methods, i.e., power curve fitting and parabola curve
Singh and Lehri [68] presented a pixel based feature extraction fitting. In both power and parabola curve fitting methods, the
technique in which the feature vector was composed of pixel val- thinned character image was segmented into z zones. Then the
ues obtained from the normalized thinned character image. Kale power curve and parabola curve were fitted to the foreground
et al. proposed an orthogonal feature extraction technique in [55], pixels in each zone using the least square method. In this work,
based on Legendre moment and Zernike moment. Shelke and the power curve fitting method provides the highest accuracy for
Apte [59] recognized characters using both structural features Gurumukhi characters.
and normalized pixel density features. Singh et al. [58] proposed Aggarwal et al. [61] used gradient and curvature based fea-
feature extraction technique based on four different kinds of tures extraction approach. These two sets of features were then
features, namely, zone based centroid, chain code, distance profile combined which result in the formation of a composite feature
and background directional distribution features. vector. The composite feature vector (CGCFV) was computed by
The research work reported in [63] was based on two kinds of the cross product of gradient directions (feature vector) and cur-
features: Projection profile histogram and Histogram of oriented vature levels (feature vector). The research work reported in [62]
gradient (HOG). The projection profile histogram was obtained by was based on the collection of different feature extraction tech-
calculating the number of background pixels along the horizontal niques. These techniques include discrete cosine transformation
and vertical directions in the entire character image. The HOG (DCT), discrete wavelet transformation (DWT), fan beam trans-
features were extracted using shape description by computing formation (FBT), and fast Fourier transformation (FFT). Among all
gradient in local segments of the images. these techniques the DCT provides the most promising results for
In [69], Neocognitron neural network was proposed for ex- feature extraction.
tracting features from handwritten Gurumukhi characters. The Kumar and Gupta [64] have extracted three kinds of features,
neocognitron neural network performs feature extraction by go- namely directional features, local binary pattern (LBP) and re-
ing through successive stages. In each stage, the recognition sys- gional features. For each character image, a sum of 117 features
tem extracts relevant features from the preceding stage and thus were extracted with 54 directional features, 59 LBP features, and
a compressed representation was formed from those extracted 4 regional features.
features for further recognition. Among earlier works on handwritten Bengali script, Basu
Siddharth et al. [70] extracted zoning density based features et al. [52] relied strongly on topologically features, namely, octant
(ZD) and background directional distribution features (BDD). In centroid features, longest run features and modified shadow
this work, they extracted 16 zonal densities features for each features for handwritten Bengali script. Surinta et al. [60] have
image and 8 BDD features were computed for each of the 16 proposed two kinds of local gradient feature descriptors, viz.,
zones. histogram of oriented gradient features (HOG) and scale invari-
Sinha et al. [71] proposed a zone-based technique based on ant features transform (SIFT) for handwritten Bengali characters
the combination of the image centroid zone and zone centroid recognition.
zone. In zone centroid zone method, the image was segmented In [73], the extracted features were categorized into two types
into z equal zones and then each zone centroid and the average i.e. global and local features. A total of 175 global features con-
distance between zone centroid and each pixel in the zone was sidered in the research work, among them 155 were convex hull
computed. In image centroid zone method, the character image features and the remaining were centre of gravity based quad
was segmented into z equal zones and then the centroid of image tree longest run features. The longest run features contributing
and the average distance between image centroid and each zone to local feature set were then obtained from each local or sub
was computed. region depicted by node in the quad tree, across four directions,
Gabor filter based feature extraction was reported in [72], in vertical, horizontal and two diagonals.
this work two variants of Gabor features GABM and GABN were Pal et al. [74] used gradient features for classifying handwrit-
used. In GABM based technique the Gabor filter outputs were ten Bengali compound characters. Firstly, size normalization was
processed to compute the energy magnitudes, which serves as done by applying 2 × 2 mean filtering on the grey scale char-
a feature vector. However, in GABN based technique the outputs acter image. The gradient image was then obtained by applying
were not processed any further and used as extracted features. Roberts filter on the normalized image. Initially, the direction of
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 11

gradient was quantized along 32 directions, therefore histograms like, feed-forward neural network, feed-backward neural net-
are computed for these 32 directions. Finally, Gaussian filter was work, self-organizing neural network, recurrent neural network
used to down sample these directional frequencies so as to obtain and many more.
392-dimensional feature vector. Among the initial works reported on handwritten Devanagari
Das et al. [53] used feature set composed of three kinds of script, Arora et al. [51] developed a recognition technique based
features, viz., shadow features, longest run features and centre on the ensemble of four MLP classifiers. Each MLP classifier was
of gravity based quad tree features. Topological features based trained with conventional backpropagation algorithm and sig-
recognition scheme was proposed in [57]. Decomposition rules moid activation function, but with different features, including 16
were formulated in this work that breaks compound characters shadow features, 200 chain-code histogram features, 32 intersec-
into subsequent simple shape components. This decomposition tion features and 48 line-fitting features. The momentum term
helps in improving the efficiency of features used and thereby and learning rate were set as 0.7 and 0.8 respectively. Finally,
provides better recognition performance. a majority voting scheme was used to aggregate the results
Pramanik and Bag [75] used chain-code histogram based fea- obtained from all four MLP classifiers.
ture set for the recognition of handwritten Bengali characters. The recognition system outlined in [68] trains two layer back-
Firstly, the smallest rectangular frame was computed for each propagation neural network (BPNN) with gradient descent delta
character image and then each image was divided into 7 × 7 learning to identify handwritten Devanagari characters. The
blocks. Chain-code histogram was computed for each block and method reported in [82] recognizes handwritten Devanagari vow-
each histogram can have one of 8 possible values, viz., 0, 1, 2, 3, els using feed-forward MLP due to its well-known generalization
4, 5, 6, 7. Thus, 7 × 7 × 8 = 392 dimensional feature vector was and edification abilities. The FFMLP was trained using static
obtained. back-propagation with sigmoid and tanh activation functions.
The method used in [76] has extracted shadow features, tran- Khanduja et al. [83] proposed a gradient descent rule-based NN to
sition features, diagonal features, directional features, zoning fea- minimize the squared error between output values and network
tures, centroid features, intersection point features, power curve- target values. This network architecture has two hidden layers
fitting features, and parabola curve-fitting features. In this work, with the first layer having 70 neurons, and the second layer hav-
the authors have also proposed two significant feature extraction ing 40 neurons. The recognition system outlined in [84] trained
approaches, namely, modified division point and peak extent using deep convolutional neural network (DCNN) with RMS Prop
based feature extraction approaches. optimizer to identify handwritten Devanagari characters.
Table 6 elucidates the feature extraction techniques for hand- Jain et al. [69] proposed a neocognitron ANN to identify hand-
written Devanagari, Gurumukhi and Bengali scripts. written Gurumukhi characters. Neocognitron ANN is a well-
known classifier in pattern recognition problems for its significant
6.1.2. Classification and recognition performance and fast processing time. In this work, neocog-
Classification is the process of categorizing feature vectors nitron ANN firstly extracts features from handwritten images
into distinct groups or classes depending upon their similarities. and then classify them. The classification is performed in series
Classification techniques works by dividing the entire dataset into of functionally equivalent stages. At every stage, a compressed
two parts: training set and testing set. Classification is a broad representation is formed by extracting relevant features from
research area in itself and also a significant phase of character the output of the previous stage. This compressed representation
recognition. There are extensively large numbers of classifiers stores the spatial location of derived features and acts as an input
available in the field of pattern recognition [77,78], so we need to the succeeding stage. Thus, the classification was done by
to restraint our scope, thus the classifiers having more relevance subsequently extracting and compressing features until the input
to this field are discussed here. character image gets reduced into a vector representation.
The classification techniques are broadly categorized into Kumar et al. [64] presented a deep neural network (DNN)
three classes: NN based approaches, SVM based approaches and based classification technique for handwritten Gurumukhi char-
miscellaneous techniques. These techniques have been analysed acters. In contrast to traditional neural networks, deep learning
in the following subsections. neural networks have a significantly large number of hidden
A. Neural Network based techniques layers. The proposed DNN has 117 hidden layers, which also
Artificial Neural network (ANN) based techniques are poten- represent the number of extracted features. For training the clas-
tially used for the classification and recognition of handwritten sifier, two autoencoder layers were used followed by a single
characters [79,80]. Due to its parallel architecture, ANN performs soft-max layer. The above system reported an accuracy of 99.30%
the computations at significantly higher frequency as compared for 2700 handwritten Gurumukhi characters. Basu et al. [52] used
to traditional techniques. These classifiers are built with an aim a back-propagation based MLP (BPMLP) for the recognition of
to simulate human brains. Neurons are the basic adaptive units handwritten Bengali basic characters with a momentum value of
in ANN with associated connections and corresponding synaptic 0.7 and a learning rate of 0.8.
weights. These classifiers are mainly composed of three layers, Pramanik and Bag [75] presented a BPMLP for the classifi-
namely, input, hidden and output layer. cation of handwritten Bengali compound characters. The back-
Jangid et al. [66] proposed a layer-wise trained Deep Convo- propagation learning algorithm was based on an iterative gra-
lutional Neural Network (DCNN) architecture for the recognition dient approach which reduces the error between actual output
of handwritten Devanagari character. The proposed model was and desired output while training MLP. Using BPMLP, the authors
tested on ISIDCHAR database with 56,477 handwritten character claim a recognition accuracy of 88.74% for handwritten Bengali
samples and claim a recognition accuracy of 98%. compound characters. Keserwani et al. [85] designed an efficient
On the basis of number of hidden layers, the neural network unified-CNN model based on the Adadelta optimizer to recognize
(NN) may be classified into different categories [81], for exam- handwritten Bengali characters. In this work, the proposed model
ple, NN with a single hidden layer is called perceptron whereas reported 98.56% accuracy and has a significantly lesser number
NN with multiple hidden layers is called Multi-Layer Percep- of parameters as compared to the conventional deep learning
tron (MLP). The NNs are further classified into different kinds architectures.
12 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

B. SVM based techniques Discriminant Function (MQDF) uses 64-dimensional chain-code

Support Vector Machine (SVM) is a hyperplane classifier pro- features to recognize 11,270 handwritten characters with an ac-
posed by Vapnik [86]. On the basis of Vapnik statistical theory, the curacy of 80.36%.
SVM classifier looks for optimal hyperplane in order to maximize Pal et al. [74] developed a MQDF classifier to identify hand-
the margin or distance between two classes in the feature space. written Bengali compound characters. In this work, the quadratic
This maximal margin hyperplane constructed by using kernel classifier used 392 dimensional gradient features for the recog-
function is called a support vector. SVM solves two-class or binary nition of 20,543 handwritten samples and achieved recognition
classification problem, a multi-class classification problem can accuracy of 85.9% with five-fold cross-validation scheme.
also be solved using SVM by combining several binary SVMs
(2) K Nearest Neighbour Classifier:
using different strategies like winner-takes-all, max-win strategy,
The k Nearest Neighbour (kNN) classifier identifies an un-
etc. For multiclass classification problems, the common existing
known sample on the basis of known neighbour’s classification.
approaches are one-against-all, one-against-one, directed acyclic
kNN has different variants depending upon the number of neigh-
graph, unbalanced decision tree and binary tree of SVM, etc. SVM
bours used in the classification [88], for example, 1NN indicates
mainly works with three kinds of kernels namely, radial basis
consideration of only one neighbour, whereas, kNN indicates the
function (RBF) kernel, linear kernel and polynomial kernel.
consideration of k neighbours. On the basis of parameters like
Kale et al. [55] used a multi-class SVM classifier based on RBF
distance method and training samples, kNN classifier can also be
kernel for the recognition of handwritten Devanagari characters.
categorized as fine or coarse kNN.
This work reported satisfactory recognition accuracy of 98.51% for
Singh et al. [54] used kNN classifier for the identification of
handwritten basic characters and 98.30% for handwritten com-
handwritten Devanagari characters. In this work, 1024 dimen-
pound characters. Yadav and Purwar [63] proposed a quadratic sional curvelet features were extracted and later reduced to 190-
SVM classifier to identify handwritten Devanagari characters. The dimensional feature vector using PCA. Then the kNN classifier was
proposed quadratic SVM classifier provides an accuracy of 96.60% used with 190-dimensional curvelet features to identify 31,860
for handwritten characters using 10-fold cross-validation scheme. character samples and claims recognition accuracy of 93.8%.
Another technique reported in [70] used RBF kernel based Kumar et al. [56] proposed a kNN classifier using power curve
multi class SVM for the classification of handwritten Gurumukhi fitting features to recognize handwritten Gurumukhi characters.
characters. While evaluating 7000 handwritten characters, a For experimental evaluations, the value of k was considered to
recognition accuracy of 95.04% was achieved using 5-fold cross be k = 1, 3, 5, 7 and the best recognition accuracy was obtained
validation. SVM classifier using RBF kernel was also discussed when k = 5. Therefore, for 3500 handwritten characters the best
in [72]. In this work, Gabor features are fed to the SVM classi- accuracy of 98.10% was obtained using 5NN classifier with power
fier to identify handwritten Gurumukhi characters. The authors curve fitting features.
claimed a recognition accuracy of 94.29% for 7000 samples of
handwritten character, with the kernel parameter ranging from (3) Mirror Image Learning:
0.64 to 1.28. Pal et al. [67] employed Mirror Image Learning (MIL) classifier
The method reported in [62] presented both polynomial and to identify handwritten Devanagari characters. MIL technique is a
linear kernel based SVM classifier with 5-fold cross-validation to corrective learning algorithm in which mirror image is generated
classify handwritten Gurumukhi characters. In this work, while for pattern belonging to a pair of confusing classes so as to
evaluating 10,500 handwritten characters, DCT2 features were enlarge the size of learning sample for other class. In this work,
fed to the SVM classifier. After experimental analysis, recognition an accuracy of 95.19% has been achieved for 36,127 handwritten
accuracies of 95.80% and 82.20% were achieved using SVM with characters using MIL classifier with curvature features.
linear kernel and degree 2 polynomial kernel respectively. (4) Template Matching:
Surinta et al. [60] proposed a RBF kernel based SVM classifier Bag et al. [57] used a Template Matching (TM) approach for
with 10-fold cross-validation to recognize handwritten Bengali the recognition of handwritten Bengali compound characters. In
basic characters. In this approach, grid search along with loga- this approach, a standard feature template has been constructed
rithmic scale was used in order to explore 2-D parameter space for each compound character. The template represents spatial
of SVM. This multiclass classification problem was based on one- layout of corresponding shape primitive of character. This repos-
against-all strategy. Das et al. [53] used SVM classifier based on itory of standard template was used for matching and identifying
RBF kernel with a 3-fold cross-validation scheme in order to handwritten compound characters. The match score of feature
identify handwritten Bengali compound characters. template with each available standard template was computed
The SVM classifier proposed in [87] has been trained using SIFT for labelling the given character image. Finally, the character
and Gabour filter features for the classification of handwritten De- image was labelled using standard template with the highest
vanagari characters. The model claims an overall recognition ac- match score. While evaluating 19,800 handwritten characters,
curacy of 91.39% using polynomial SVM classifier with a ten-fold a recognition accuracy of 86.74% was obtained using TM with
cross-validation technique. topological features.
C. Miscellaneous techniques 6.1.3. Some observations on handwritten Devanagari, Gurumukhi
The miscellaneous techniques are composed of those methods and Bengali scripts
that do not fall under two above mentioned categories. These Table 7 summarizes the performance details of several classi-
techniques include several approaches like Quadratic Discrim- fiers for the recognition of handwritten Devanagari, Gurumukhi
inant Function (QDF), K Nearest Neighbour (KNN) algorithm, and Bengali scripts. The best recognition accuracy reported so far
Template Matching (TM), Mirror Image Learning (MIL), Hidden for handwritten Devanagari script is 98.51% [55]. This recognition
Markov Model (HMM), Decision Tree (DT) classifier and many accuracy is obtained by feeding SVM classifier with statistical
more. The work reported on techniques is summarized below. features like Legendre and Zernike moment extracted from a
dataset of 12,000 handwritten basic character images. As shown
(1) Quadratic Classifier: in Table 8, the maximum accuracy achieved for Gurumukhi script
A quadratic classifier with 5-fold cross-validation was pre- is 99.30% [64] by training deep-NN with directional, LBP and
sented in [50] for the recognition of handwritten Devanagari regional features. Table 9 compiles the experimental results ob-
characters. The quadratic classifier based on Modified Quadratic tained for handwritten Bengali script. The best accuracy of 98.12%
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 13

Table 7
Recognition accuracies for handwritten Devanagari characters.
Methodology Dataset size Feature extraction Classification technique Recognition accuracy (%)
Kale et al. [55] 27,000 Legendre and Zernike moment Support Vector Machine 98.51(Basic),
98.30(Compound)
Jangid and Srivastava [66] 56,477 Automatic SL-DCNN 98.00
Singh and Maring [58] 20,000 Zone based centroid, CC, Distance Profile Support Vector Machine 97.61
and BDD
Shelke and Apte [59] 40,000 Pixel density features, Structural Fuzzy system and FFNN 96.95
Yadav and Purwar [63] 4428 Histogram of oriented gradient, Projection Quadratic SVM 96.60
profile
Jangid et al. [84] 36,172 Automatic DCNN with RMS Prop 96.00
Pal et al. [67] 36,172 Curvature, Gradient Mirror Image Learning 95.19(Curvature),
94.94(Gradient)
Sarkhel et al. [89] 22,086 Multiscale-multicolumn CNN Support Vector Machine 95.18
Singh et al. [54] 31,860 Curvelet based features K-Nearest Neighbour 93.80
Singh and Lehri [68] 1000 Pixel based features Backpropagation NN 93.00
Arora et al. [51] 4900 Shadow, CC histogram, Intersection features Feed-foward NN 92.80
and Line fitting
Narang et al. [87] 5484 SIFT, Gabor filter, PCA Poly-SVM 91.39
Sharma et al. [50] 11,270 Direction CC histogram Modified QDF 80.36

Table 8
Recognition accuracies for handwritten Gurumukhi characters.
Methodology Dataset size Feature extraction Classification technique Recognition accuracy (%)
Kumar and Gupta [64] 2700 Directional, LBP, Regional features Deep Neural Network 99.30
Aggarwal and Singh [61] 7000 Curvature and Gradient Support Vector Machine 98.56
Kumar et al. [56] 3500 Power curve fitting K-Nearest Neighbour 98.10
Kumar et al. [62] 10,500 Discrete cosine transform SVM with linear kernel 95.80
Sinha et al. [71] 7000 Zone based features SVM, KNN 95.11, 90.64
Siddharth et al. [70] 7000 Zoning density and BDD Support Vector Machine 95.04
Singh et al. [72] 7000 Gabor features SVM with RBF Kernel 94.29
Jain and Sharma [69] 15,000 Automatic features Neocognitron NN 92.78
Garg et al. [76] 8960 Peak extent, MDP, PCA Linear SVM+polySVM+kNN 92.30
Kaur and Rani [90] 2450 Zoning features Convolutional NN 92.08

Table 9
Recognition accuracies for handwritten Bengali basic and compound characters.
Methodology Dataset size Dataset type Feature extraction Classification technique Recognition accuracy (%)
Sarkhel et al. [89] 42,697 Compound Multiscale-multicolumn CNN Support Vector Machine 98.12
Keserwani et al. [85] 41,536 Compound Automatic Unified-CNN 98.12
Roy et al. [91] 42,959 Compound Automatic SL-DCNN 90.33
Pramanik and Bag [75] 10,240 Compound Chain code histogram Multi-layer perceptron 88.74
Bag et al. [57] 19,800 Compound Topological Template matching 86.74
Pal et al. [74] 20,543 Compound Gradient MQDF 85.90
Das et al. [53] 19,765 Compound Shadow, Quad tree, LR Support Vector Machine 80.51
Keserwani et al. [85] 15,000 Basic Automatic Unified-CNN 98.56
Rabby et al. [65] 15,000 Basic Automatic extraction BornoNet 98.00
Sarkhel et al. [73] 15,000 Basic CG based Quad tree LR, Convex hull Support Vector Machine 87.28
Sarkhel et al. [92] 15,000 Basic LR, Enhanced harmony search SVM with RBF kernel 86.53
Roy et al. [93] 15,000 Basic ABCO, Gradient features Support Vector Machine 86.40
Gupta et al. [94] 15,000 Basic HOG, Convex hull, LR, Harmony search Support Vector Machine 86.10
Surinta et al. [60] 5527 Basic Scale invariant feature transform Support Vector Machine 85.00
Basu et al. [52] 10,800 Basic Modified shadow, Octant centroid, LR MLP based two stage classifier 80.58

is achieved by the SVM classifier with the extraction of non- Initially, seven invariant moments has been derived for each char-
explicit MMCNN features from a standardized CMATER db 3.1.3 acter image but this process provides very poor recognition rate.
dataset [89]. Additionally, the unified-CNN model [85] trained us- Thus, the character image was firstly segmented into four zones,
ing the Adadelta optimizer with 0.95 decay rate and 1.00 learning viz., upper left, upper right, lower left and lower right. Then,
rate also achieved 98.12% recognition accuracy on CMATER db invariant moments were evaluated for each zone resulting in total
3.1.3 dataset. of 7 × 4 = 28 features. Dhandra et al. [97] used the directional
6.2. Reported work on Kannada and Telugu scripts spatial features to recognize handwritten characters. They ex-
tracted directional spatial features like number of strokes, stroke
6.2.1. Feature extraction length and stroke density as potential features for handwritten
Among earlier works on Handwritten Kannada and Telugu Kannada character recognition.
script, Pal et al. [95] has proposed a directional feature based The method used in [98] has extracted shape based features
technique. In this work, they segmented each character image viz., chain codes and invariant Fourier descriptors to identify
into equal blocks and the directional features were evaluated for character in handwritten Kannada script. Authors in [101] also
each block. Then The Gaussian filter was used to downsample relied on shape based features, namely, wavelet packets and nor-
these blocks and the directional features obtained from these malized chain codes. In this work, wavelet decomposition coeffi-
downsampled blocks are then fed to the classifier. cients and normalized chain codes were extracted as 22 dimen-
Sangame et al. [96] extracted moment invariant features from sional feature vector from normalized binary character images of
zoned images for handwritten Kannada character recognition. size 40 × 40.
14 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

Table 10
Description of Feature extraction techniques for handwritten Kannada and Telugu scripts.
Features Category Reference
Moment invariant features Statistical features Sangame et al. [96]
Fourier descriptors and chain codes Statistical and structural features Rajput and Horakeri [98]
Zone based and pixel density features Statistical features Mukarambi et al. [99]
Positional features Statistical features Vaidya and Bombade [100]
Wavelet decomposition and normalized chain code Statistical and structural features Dhandra and Mukarambi
[101]
Zoning method Statistical features Sastry et al. [102]
MMCNN Non-explicit features Sarkhel et al. [89]
Block wise pixel count Statistical features Lakshmi [103]

Table 11
Recognition accuracies for handwritten Kannada script.
Methodology Dataset size Feature extraction Classification technique Recognition accuracy (%)
Karthik et al. [104] 18,800 Distributed average of gradients Deep belief networks 97.04
Dhandra and Mukarambi [101] 1400 Normalized chain code and wavelet decomposition K-Nearest Neighbour 95.07
Rajput and Horakeri [98] 6500 Fourier descriptors and chain codes SVM 93.92
Rani et al. [105] 5200 Automatic Alex net 92.00
Pasha and Padma [106] 4800 Structural features and Wavelet transform ANN 91.00
Pal et al. [95] 10,779 Directional features Quadratic classifier 90.34
Dhandra et al. [97] 1400 Spatial features K-Nearest Neighbour 90.10
Angadi and Angadi [107] 2490 Structural features SVM 89.84
Vaidya and Bombade [100] 7350 Positional features GRNN 85.62
Sangame et al. [96] 1625 Moment invariant features K-Nearest Neighbour 85.53
Mukarambi et al. [99] 2800 Zone based and pixel density features SVM 73.33

Mukarambi et al. [99] proposed zoned based feature extraction features include dimensional features, small glyphs positional
approach. All the normalized 32 × 32 dimensional character features and zone based features. Sarkhel et al. [89] presented a
images were segmented into 64 non overlapping zones. Thus non-explicit feature extraction techniques. In this work, a Multi-
64 zone based features were obtained by computing the pixel column Multiscale Convolutional Neural Network (MMCNN) ap-
density for each zone. Vaidya et al. [100] developed a feature proach was developed. The multiscale convolutional sampling
extraction technique based on positional information of each method helps in extracting more invariant and robust features
pixel in the handwritten character image. Firstly, the avg_matrix from character images and the multicolumn architecture im-
is computed by adding all the character image matrices into proves the effectiveness of the system.
a single matrix and then this resultant matrix was divided by Lakshmi used [103] two feature extraction techniques, namely,
the total number of image matrices. The avg_matrix was then cell pixel count feature extraction and histogram profile. Among
subtracted from each character image matrix so as to obtain these two techniques cell pixel count feature extraction provides
unique features based on positional information of each pixel in better performance to identify characters.
the character image. Angadi et al. [107] used structural features Table 10 summarizes the feature extraction techniques for
for the identification of handwritten Kannada characters. These handwritten Kannada and Telugu scripts.
features include structural information of character images like 6.2.2. Classification and recognition
eccentricity, orientation, perimeter, filled area and convex area. A. Neural Network based techniques
The work reported in [106] was based on the extraction of Vaidya [100] proposed a generalized regression neural net-
wavelet features and structural features, namely, aspect ratio, work classifier (GRNN) for the recognition of handwritten Kan-
quadrant density, corner detection, width and correlation to rec- nada characters. The architecture of GRNN classifier was com-
ognize handwritten Kannada characters. A distributed average of posed of pattern layer and summation layer. The GRNN recogni-
gradient-based feature extraction technique (DAG) has been pro- tion model was based on radial basis network and trained using
posed in [104] to recognize handwritten Kannada characters. The positional features. Experimental evaluation on 7350 character
DAG features are inspired by the conventional speeded up robust samples reported recognition accuracy of 85.62%.
features (SURF) descriptors, however, the numbers of windows in A Feed-Forward ANN classifier has been presented in [106]
DAG, are kept fixed to 4. to recognize handwritten Kannada characters. The FFANN trained
Sastry et al. [108] proposed a 3 D feature extraction technique using wavelet and structural features like aspect ratio, quadrant
for the classification of handwritten Telugu characters. These 3 density, corner detection, width, and correlation. This classifier
D features were based on the X, Y and Z coordinates of pixels for evaluated 4800 samples of handwritten characters and achieved
each character image, where the attribute Z indicates the depth of an accuracy of 91.00%. Ragha et al. [110] used back propagation
indentation at the pixel. This depth is very important attribute for MLP classifier to identify handwritten Kannada characters. In
character recognition and is directly proportional to the pressure this work, they trained BPMLP using statistical and moment fea-
applied by pen at the pixels. tures extracted from cut images, directional images and original
Zone based features were extracted in [102], the character images.
image was firstly partitioned into number of non-overlapping Angadi et al. [111] presented a CNN based recognition model
zones or segments. The statistical features were computed for for handwritten Telugu characters. In this work, the CNN ar-
each pixel contained in the segment of character image. These chitecture comprises of four convolutional layers, followed by a
pixel intensities were then summed to obtain feature for each max-pooling layer and two fully-connected layers. Furthermore,
segment of the image. Then the features of all segments in the the generalization techniques, such as data augmentation and
character image form the feature vector. dropout, are also applied. SGD optimizer and categorical cross-
Manisha et al. [109] proposed a hybrid feature extraction entropy loss are used for training and evaluating the model,
method to recognize handwritten Telugu characters. These hybrid respectively.
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 15

B. SVM based techniques classifier with the extraction of non-explicit MMCNN features
Rajput et al. [98] presented a Gaussian kernel based SVM from standardized HPL-iso-telugu-char dataset.
classifier with 5-fold cross-validation to recognize handwritten 6.3. Reported work on Gujarati, Oriya, Tamil and Malayalam scripts
Kannada characters. In this work, normalized chain-code and
Fourier descriptor features were fed to the multiclass SVM clas- 6.3.1. Feature extraction
sifier using the one-versus-rest class model. While evaluating Prasad et al. [112] used the structural features or pixel based
6500 handwritten characters, they claimed recognition accuracy features to identify handwritten Gujarati characters, where indi-
of 93.92%. Mukarambi et al. [99] developed an SVM classifier vidual image pixels were treated as features. Patel et al. [113]
have extracted a combination of statistical and structural features.
based on zoning and pixel density features with a two-fold cross-
The structural features were used as primary features and statisti-
validation scheme for the classification of handwritten Kannada
cal features were used as secondary features. The primary feature
characters.
set consists of number of objects in image, number of objects in
In [107] the authors presented a one-versus-all approach
image’s upper part, number of objects in image’s lower part and
based multiclass SVM, trained using topological and structural number of holes in the character image. The secondary feature
features to identify handwritten Kannada characters. SVM classi- set consists of averaging features, moment features and centroid
fier is one of the prominent classifiers in the pattern recognition distance features.
field. A hybrid feature extraction technique was used in [114,115],
The non-explicit multicolumn multiscale CNN based features based on the combination of statistical and structural features.
were fed to SVM classifier in [102] for recognition of handwritten The proposed technique extracts four features, namely, pattern
Telugu characters. This work reported an accuracy of 95.76% with descriptor, Gabor phase XNOR pattern features (GPXNP), autocor-
a five-fold cross-validation scheme. relation and contour direction probability distribution function
features (CDPDF). The pattern descriptor represents the structural
C. Miscellaneous techniques features whereas GPXNP, autocorrelation and CDPDF represents
statistical features. Among these features only pattern descriptor
(1) Quadratic Discriminant Function: and GPXNP are considered to be relevant due to their higher
Pal et al. [95] used a quadratic classifier for the recognition power values and the fuzzy hedges values are also maximum for
of handwritten Kannada and Telugu characters. In this work, each class. The remaining two features, viz., autocorrelation and
400-dimensional directional features were fed to the quadratic CDPDF are considered to be irrelevant on the basis of their fuzzy
classifier to recognize 10,779 Kannada characters and 10,872 hedge values and hence rejected.
Telugu characters. They claimed recognition accuracy of 90.34% The structural features also plays significant role in the re-
and 90.90% respectively. search reported in [116]. The method includes structural features
like number of closed loops, connected components or discon-
(2) K Nearest Neighbour Classifier: nected components and number of end points. In [117], the
Sangame et al. [96] applied the kNN classifier with Euclidean direction and strength of gradient was obtained by applying
distance criterion to classify handwritten Kannada characters. Robert’s filter on Oriya character image. Then the biquadratic in-
Dhandra et al. [97] used the kNN classifier with 4-fold cross- terpolation technique was used for computing curvature features,
validation to recognize handwritten Kannada characters. While quantized in three levels according to linear, concave and convex
experimental evaluation a recognition accuracy of 90.1% was regions. Finally, the dimension of obtained feature vector was
achieved using spatial features and kNN classifier with k = 1. reduced by applying principal component analysis (PCA) and fed
Authors in [101] presented a kNN classifier with normal- to the classifier.
ized chain code and wavelet decomposition features to identify Pal et al. [117] used the Fisher ratio (F-ratio) and gradient
handwritten Kannada vowels. In this work, they performed the based technique for the recognition of handwritten Oriya char-
experimental evaluations using varying k values with k = 1, 3, acters. Initially, they obtained 400-dimensional gradient feature
5 and optimal results were obtained when k = 3. The average vector by applying Roberts and Gaussian filter. Then the feature
recognition accuracy achieved for handwritten Kannada charac- vector was modified by using feature weighting scheme based
on F-ratio. This feature weighting technique identifies the similar
ters was 95.07%. Sastry et al. [102] used the kNN classifier with
shaped characters much easily by reducing the features belonging
zone-based pixel intensity features for identifying handwritten
to identical parts of similar shaped characters and enhancing
Telugu characters.
the features belonging to distinguishable parts of similar shaped
(3) Decision Tree Classifier: characters.
Sastry et al. [108] proposed a Decision Tree (DT) classifier Padhi [118] extracted several features like average angle based
on image centroid, average angle based on zone centroid, average
trained using 3 D features to recognize handwritten Telugu char-
distance based on zone, centroid distance based on zone and
acters. The DT was constructed using SEE5 algorithm, which
standard deviation based features. Dash et al. [119] proposed a
claims a recognition accuracy of 93.10% for handwritten samples.
binary external symmetry axis constellation (BESAC) feature ex-
6.2.3. Some observations on handwritten Kannada and Telugu
traction technique. This method works by generating two binary
scripts
coded histograms of orientations which were then concatenated
Table 11 compiles the experimental results obtained from
to obtain proposed BESAC feature.
several classifiers for the recognition of handwritten Kannada In [120], the character image was partitioned into 3 × 3 zones
scripts. The maximum recognition accuracy obtained so far for and the centroid is evaluated for each zone. Then the vertical and
the handwritten Kannada script is 97.04% [104]. This recognition horizontal Euclidean distance was computed for the nearest pixel
accuracy is obtained by feeding Deep Belief Networks classifier from each zone centroid. Finally, the mean Euclidean distance
with distributed average of gradients features. Table 12 sum- was calculated along with the mean angular values of each zone
marizes the performance details of several classifiers for the and these values corresponds to the feature set.
recognition of handwritten Telugu script. As shown in Table 12, Bhattacharya et al. [121] proposed a two stage recognition
the best recognition accuracy of 95.76% [89] is achieved by SVM system for handwritten Tamil script. In the first stage of this
16 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

Table 12
Recognition accuracies for handwritten Telugu script.
Methodology Dataset size Feature extraction Classification technique Recognition accuracy (%)
Sarkhel et al. [89] 45,217 MMCNN SVM 95.76
Sastry et al. [108] Not-specified 3D features Decision Tree 93.10
Angadi et al. [111] 45,133 Automatic CNN with SGD optimizer 92.40
Pal et al. [95] 10,872 Directional Quadratic classifier 90.90
Lakshmi [103] 18,000 Block wise pixel count KNN and SVM 90.80
Manisha et al. [109] 23,875 Hybrid and zone based features KNN 88.15
Sastry et al. [102] 19,250 Zoning method Nearest Neighbour classifier 78.00

recognition scheme, the character image was partitioned into concentrated in 32 curvature directions. Then concatenate these
7 × 7 blocks of equal size. Then each block was scanned along two feature sets and apply PCA on concatenated feature set in
both vertical and horizontal directions. Finally, the total number order to obtain final reduced feature vector.
of transitions from black to white and vice versa were counted Raju et al. [132] proposed a gradient and run length count
in each scan for obtaining the feature vector. In the second stage based feature extraction technique (GF-RLC) along with three
of recognition scheme, the chain code histogram features were other simple features, namely, position of centroid, character
computed for contour representation of input character image. code and aspect ratio for identification of Malayalam charac-
Shanthi et al. [122] used the pixel based density features. In ters. Manjusha et al. [133] have designed feature descriptors
this work, the character was segmented into non overlapping based on scattering convolution network for the classification of
blocks. Then the pixel densities were computed for each block of handwritten Malayalam characters.
character image and used as features. Subashini et al. [123] pro- Table 13, elucidates the features extraction techniques for
posed the SIFT algorithm for constructing local invariant feature handwritten Gujarati, Oriya, Tamil and Malayalam scripts.
vector for each character image. They generated the codebook 6.3.2. Classification and recognition
from the obtained feature vector set for each character image
A. Neural Network based techniques
using K means clustering. The bag-of-keypoints were then calcu-
Jose et al. [124] proposed a feed-forward BPNN using sigmoid
lated for total number of images. These features are then finally
activation function to identify handwritten Tamil characters. The
fed to SVM for classification.
FFBPNN consists of three layers and based on gradient descent
In [124] the localized features were obtained using two-
back-propagation approach. In this work, the FFBPNN is trained
dimensional wavelet transformation for handwritten Tamil char-
using wavelet decomposition features to classify handwritten
acter recognition. Abirami et al. [125] have extracted six statis-
samples.
tical features from the character boundaries to identify Tamil
Manuel et al. [136] used a back-propagation based MLP
characters. These six different features are freeman directional
(BPMLP) to classify handwritten Malayalam characters. The
code, slope angle, aspect ratio, curvature, linearity and curliness.
BPMLP was trained using curvelet-based features. This work
The research work reported in [126] was based on two step
reported a recognition accuracy of 95.99% on 2120 handwritten
procedure of feature selection and extraction. The features were
character samples.
firstly selected using zoning and 8 directional chain-code tech-
Extreme Learning Machine (ELM) [137], is a fast learning neu-
nique. Then the selected features were extracted using bound
ral network, in which the input weights and hidden biases are
boxing algorithm and sub line direction, and fed to the classifier.
chosen randomly. Along with the extremely fast learning prop-
Raj et al. [127] used zoning technique to partition the charac-
erty, this algorithm also provides a good generalization ability.
ter image into different zones. Then the chain code algorithm was
The work reported in [130] claimed an accuracy of 95.59% using
applied on each zone of the character image for selecting essential
wavelet feature based ELM on 9000 sample images.
character portion features. Finally, features were extracted from
Raju et al. [132] proposed a FFBPNN classifier for the recogni-
each selected portion of the character image using pixel points
tion of handwritten Malayalam characters. In this approach, the
count, pixel based on zoning division (horizontal and vertical),
FFBPNN with sigmoid activation function was trained by using
pixel location based on axis, and pixel location based on row, col-
combination of run length count and gradient features along with
umn and diagonal. The method in [128] was based on statistical
other simple features like centroid, character code and aspect ra-
analysis in which the pixel points present in the character image
tio. Using FFBPNN, a recognition accuracy of 99.78% was achieved
were analysed and represented using quad tree algorithm. The
after evaluating 19,800 handwritten characters.
research work reported in [129] was based on the extraction of
twelve directional features for 36 blocks and thereby generating B. SVM based techniques
a feature vector of dimension 432. Shanthi et al. [122] used a SVM classifier based on statisti-
The wavelet energy based features (WEF) were extracted cal learning theory to recognize handwritten Tamil characters.
from character image using wavelet transform in [130]. WEF The pixel density based features were computed for training
represents wavelet energy distribution of handwritten charac- the SVM classifier using max-win voting approach. Experimental
ters in various directions at different decomposition levels. The evaluations on 41,489 handwritten characters claim a recogni-
wavelet coefficient amplitude increases with the increase in scale tion accuracy of 82.04% with 5-fold cross-validation. Subashini
of wavelet decomposition. In order to discriminate handwritten et al. [123] presented a linear SVM classifier trained using local
character images, the wavelet energy of different levels possess invariant SIFT descriptor features to classify handwritten Tamil
different powers. Thus, the feature vector represents patterns of characters.
characters on the basis of wavelet energy for classification. Shyni et al. [126] proposed a weighted SVM, in which the
Jomy et al. [131] used a combination of gradient and curvature weight factor was computed using Lagrangian Theorem. The
based feature extraction technique along with principal compo- weighted SVM classifier was trained using chain-code, zoning,
nent analysis for dimension reduction. Firstly, gradient features bound box and sub-line detection features to identify 6000 char-
were computed using directional information obtained from arc acter samples with a recognition accuracy of 88%. Raj et al. [128]
tangent of gradient. Secondly, curvature features were obtained also used multiclass SVM based on Lagrangian theorem for the
using biquadratic interpolation technique on strength of gradient classification of handwritten Tamil characters They implemented
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 17

Table 13
Description of feature extraction techniques for Gujarati, Oriya, Tamil and Malayalam scripts.
Features Category Reference
Gradient, Curvature and PCA Statistical and structural features Pal et al. [117]
Gradient and F-ratio Statistical features Wakabayashi et al. [134]
Pixel density Statistical features Shanthi and Duraiswamy [122]
SIFT feature descriptors Statistical features Subashini and Kodikara [123]
Centroid and moment-based features Statistical features Patel and Desai [113]
CL, Connected-disconnected components and End points Structural features Thaker and Kumbharana [116]
Gradient based features and Run length count Statistical features Prasad and Kulkarni [115]
Zone, Chain-code Statistical and structural features Raj and Abirami [127]
Quad tree feature Statistical and structural features Raj et al. [128]
Automatic features using LSTM Non-explicit features Jino et al. [135]
Zone based mean angular values and mean ED Statistical features Sethy et al. [120]
Reduced scattering convolutional network Non-explicit features Manjusha et al. [133]

the SVM classifier using one against all approach to predict 4200 like Euclidean distance and triangular distance. The proposed
handwritten samples and indicates an accuracy of 88.25%. The classifier along with Gabor phase XNOR pattern features (GPXNP)
multiclass SVM classifier proposed in [138] has been trained using provides a recognition accuracy of 86.33% for handwritten sam-
discrete value features extracted from Z-ordering, strip-tree, and ples. The work reported in [119] was classified using nearest
quad-tree approaches to identify handwritten Tamil characters. In neighbour classifier on the basis of BESAC features. The pro-
this multiclass SVM classifier, the one-vs-all class model is used posed approach claimed a recognition accuracy of 99.48% for
along with the divide and conquer technique. 7800 handwritten Oriya characters.
Jomy et al. [131] used a SVM classifier with ten-fold cross-
(3) Template Matching:
validation to identify handwritten Malayalam characters. The
A template matching technique based on structural features
multiclass SVM classifier operates using RBF kernel, with kernel
was presented in [112] for the recognition of handwritten Gu-
parameter (γ ) equals to 0.02. While evaluating 13,200 handwrit-
jarati characters.
ten Malayalam characters, a recognition accuracy of 97.96% was
achieved. Manjusha et al. [133] presented a SVM classifier with (4) Decision Tree Classifier:
linear kernel to recognize handwritten Malayalam characters. The research work presented in [116] proposed a decision
The linear SVM was trained using scattering convolutional net- tree classifier to identify handwritten Gujarati characters. In the
work features extracted from 29,302 character images, provides proposed model, several features like closed loops, connected-
a recognition accuracy of 91.05%. disconnected components and end points were fed to decision
tree classifier which claims an accuracy of 88.78%.
C. Miscellaneous techniques
(5) Adaptive Neuro Fuzzy Classifier
(1) Quadratic Discriminant Function Classifier: Prasad et al. [114] implemented an Adaptive Neuro Fuzzy
Pal et al. [117] proposed a quadratic classifier with five-fold (ANF) classification technique to recognize handwritten Gujarati
cross-validation to recognize handwritten Oriya characters. In characters. This ANF classification model was trained using
this work, gradient and curvature features were fed to quadratic GPXNP features.
classifier which achieves a recognition accuracy of 94.6%. Wak-
abayashi et al. [134] presented a Quadratic Discriminant Function (6) Hidden Markov Classifier
(QDF) classifier with five-fold cross-validation to identify hand- Abirami et al. [125] presented a symbol modelling Hidden
written Oriya characters. F-ratio based weighted features were Markov Model (HMM) for the classification of handwritten Tamil
fed to QDF classifier, which claims an accuracy of 95.14% for characters. Since the handwritten character recognition is consid-
18,190 handwritten character samples. ered to be very complex problem due to several dissimilarities
Pal et al. [95] used a quadratic classifier for the recognition in characters like different size, dimension, thickness, orientation
of handwritten Tamil characters. In this work, 400-dimensional and style. HMM are proved to be very efficient classifier in such
directional feature vector was fed to the quadratic classifier to scenarios.
recognize 10,216 Tamil characters and claims recognition accu- HMM mainly depends upon state probabilities and states are
racy of 96.73%. Moni et al. [129] proposed a modified QDF to assigned on the basis of features. In this work, Baum Welch
classify handwritten Malayalam characters. The MQDF classifier algorithm was used along with three probabilities to train HMM.
was trained using 12 directional code features and claims an ac- The HMM transition was represented using 5 states and prob-
curacy of 95.42% for 19,800 handwritten samples. In comparison ability value was assigned on the basis of reference value. The
to QDF, the MQDF enhances the recognition accuracy by more proposed HMM model was trained using freeman directional
than 10% and also reduces the computational cost. code, curvature, aspect ratio, slope angle and curliness features
Raju et al. [132] used a simplified QDF classifier for the recog- and claims a recognition accuracy of 85.00%.
nition of handwritten Malayalam characters. In this approach, the 6.3.3. Some observations on handwritten Gujarati, Oriya, Tamil and
SQDF was trained using combination of run length count and Malayalam scripts
gradient features along with other simple features like centroid, Table 14 summarizes the performance details of various clas-
character code and aspect ratio. Using SQDF, a recognition accu- sifiers for the recognition of handwritten Gujarati scripts. The
racy of 99.66% was reported after evaluating 19,800 handwritten best recognition accuracy reported so far for handwritten Gujarati
characters. script is 99.48% [139]. This recognition accuracy is obtained by
using Polynomial kernel SVM classifier with structural decom-
(2) K Nearest Neighbour Classifier: position. As seen from Table 15, a wide variety of classifiers are
Prasad et al. [115] used a weighted kNN classifier for the used for the identification of handwritten Oriya scripts. Among
recognition of handwritten Gujarati characters. In the proposed them, the quadratic discriminant classifier provides the optimal
approach, the conventional kNN algorithm was improved by ex- recognition accuracy of 97.40% [120]. Table 16 compiles the ex-
ploiting combination of new feature weights, distance measures perimental results obtained for handwritten Tamil script. The
18 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

Table 14
Recognition accuracies for handwritten Gujarati script.
Methodology Dataset size Feature extraction Classification technique Recognition accuracy (%)
Sharma et al. [139] 20,500 Structural decomposition Polynomial kernel SVM 99.48
Pareek et al. [140] 10,000 Automatic CNN with Adam optimizer 97.21
Thaker and Kumbharana [116] 750 Closed loops, CDC and End points Decision tree classifier 88.78
Prasad and Kulkarni [115] 16,560 GPXNP Weighted KNN 86.33
Prasad et al. [112] Unspecified Structural features Template matching 71.66
Prasad and Kulkarni [114] 16,560 GPXNP Adaptive NFC using feature selection 68.67
Patel and Desai [113] Unspecified Centroid and moment based features Tree classifier and KNN 63.10

Table 15
Recognition accuracies for handwritten Oriya script.
Methodology Dataset size Feature extraction Classification technique Recognition accuracy (%)
Sethy et al. [120] 9400 Zone based mean angular values and mean ED Quadratic discriminant classifier 97.40
Wakabayashi et al. [134] 18,190 Gradient and F-ratio Modified QDF 95.14
Dash et al. [119] 7800 Binary ESAC K-Nearest Neighbour 95.01
Pal et al. [117] 18,190 Gradient, Curvature and PCA Quadratic classifier 94.60
Padhi [118] Not specified Average angle, Average distance and SD BPNN and Genetic Algorithm 94.00
Dash et al. [141] 10,200 Tetrolet + SCC Nearest Neighbour 93.24

Table 16
Recognition accuracies for handwritten Tamil script.
Methodology Dataset size Feature extraction Classification technique Recognition
accuracy (%)
Sarkhel et al. [89] 82,000 MMCNN SVM 98.79
Kavitha et al. [142] 82,928 Automatic CNN 97.70
Pal et al. [95] 10,216 Directional Quadratic classifier 96.73
Raj et al. [138] 10,000 Z-ordering and Tree representation Hierarchical SVM 90.30
Bhattacharya et al. [121] 77,609 Directional, Chain code histogram K-means algorithm and 89.66
Multi-layer perceptron
Raj and Abirami [127] 12,000 Zone, Chain-code and Statistical features SVM 89.00
Jose and Wahi [124] 3100 Wavelet decomposition BPNN 89.00
Raj et al. [128] 4200 Quad tree feature SVM 88.25
Shyni et al. [126] 6000 Chain code, zoning, bound box and Weighted SVM 88.00
sub-line detection
Abirami et al. [125] 3360 Freeman directional code, curvature, etc. HMM 85.00
Shanthi and Duraiswamy [122] 41,489 Pixel density SVM 82.04
Subashini and Kodikara [123] 8000 SIFT feature descriptors SVM 81.62

Table 17
Recognition accuracies for handwritten Malayalam script.
Methodology Dataset size Feature extraction Classification technique Recognition
accuracy (%)
Raju et al. [132] 19,800 Gradient based features, run Simplified quadratic discriminant 99.78(MLP),
length count function and MLP 99.66(SQDF)
Jomy et al. [131] 13,200 Gradient, Curvature and PCA SVM with RBF kernel 97.96
Jino et al. [135] 18,000 Automatic features Stacked Long Short Term Memory 97.00
Manuel et al. [136] 2120 Curvelet transform Multi-layer perceptron 95.99
Chacko et al. [130] 9000 Wavelet Energy features Extreme Learning Machine 95.59
Moni and Raju [129] 19,800 Directional features Modified QDF 95.42
Kishna et al. [143] Unspecified Texture features Hybrid HMM with ANN 93.40
Manjusha et al. [133] 29,302 Reduced scattering Linear SVM 91.05
convolutional network

best accuracy of 98.79% is achieved by SVM classifier with the • One of the basic issues in Indic script OCRs is to find a
extraction of non-explicit MMCNN features from standardized significantly good benchmark dataset. Therefore, for Indic
HPL-iso-tamil-char dataset [89]. As shown in Table 17, the max- scripts, there is a necessity for the construction of exhaus-
imum recognition accuracy for Malayalam script is 99.78% and tive and standardized datasets. Most of the work done on In-
achieved by training MLP classifier GBF-RLC features [132].
dic scripts OCR is based on relatively small datasets gathered
in the laboratory environment.
7. Analysis of work done on handwritten Indic sripts OCR
• Parameters like accuracy, runtime and error rate are neces-
sary to perform comparative as well as experimental analy-
• Due to infinite variations in handwriting styles, the task of
sis of feature extraction and classification models.
handwritten script recognition becomes quite complex and
challenging. • From the survey, it is clearly evident that the accuracy of the
• It is evident from the studies that for handwritten script model relies on the size and quality of the dataset.
recognition, high accuracy rates are achieved by using hy- • Deep learning models provide very high recognition accu-
brid features i.e. the combination of statistical and structural racies, but only when the size of the training dataset is
features. sufficiently large.
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 19

• In hybrid classification models, the majority voting scheme the distinguishing element is lost in pre-processing phase
is mostly used by the researchers to aggregate the results or if the distinguishing element is too small to identify due
obtained from different classifiers. to variability in the writing styles. Thus, these characters
• Non-explicit feature extraction techniques such as ANN, need peculiar attention in order to get recognized correctly.
HMM, CNN, etc. helps in extracting more invariant and A two-stage recognition scheme can be used for handling
robust features from handwritten character images. these confusing characters, in which the first stage groups
• Accuracy of handwritten script recognition system depends the similar shaped characters into one class resulting in
largely upon the features with high discriminating power. reduced number of recognition classes. Then the second
Hence, there is an immense need to investigate feature stage classifies the similar shaped characters within the
selection techniques in order to achieve the most accurate class by extracting character specific features and thereby
results. immensely improving the recognition results.
• No standardized dataset is available for handwritten Guru- • Demand for multi-script OCR: India is a multi-script and
mukhi, Gujarati and Malayalam scripts. multi-lingual country with most of the people communi-
• Studies reported on Indic scripts indicate that various re- cating in their own regional languages, so there is a huge
searchers have created their in-house datasets to carry out need for the development of multi-script character recog-
the research. nition system for Indic scripts. To design a multi-script
• Most of the standardized and benchmark datasets available recognition system, it is necessary to recognize different
for Indic scripts belong to Devanagari and Bengali scripts. scripts documents before delivering them to the recognition
Other than these, some standardized datasets are also avail- system of individual scripts. The research work reported on
able for Kannada, Telugu, Oriya, Tamil scripts. multi-script Indic scripts is still at its infancy, so this area
• The work reported on handwritten Indic scripts signifies needs great attention to support the community of active
that principal component analysis is one of the highly used researchers in this field.
and efficient dimension reduction techniques. • Development of OCR for historical and degraded documents:
• The success rate of the classifier entirely depends upon the Majority of research work reported on Indic scripts recogni-
features extracted. So, in order to achieve good accuracy, tion are based on good quality documents. The work done
the feature extraction technique and classification algorithm on poor quality degraded documents is almost negligible.
should be chosen carefully. Therefore, experiments should be made in this direction for
• Studies reveal that among all Indic scripts, maximum work furthering the research on degraded, noisy and historical
has been reported on Devanagari and Gurumukhi scripts, documents. The recognition of these degraded historical and
and have yielded satisfactory results. ancient archives will help several memory institutions to
• From the above findings, it is analysed that SVM and ANN digitalize their manuscript collection. Thus, contributing to-
are very common classifiers that provide good recognition wards the preservation and advancement of ancient heritage
accuracies for handwritten Indic scripts. But again some lat- and will be very useful to philologist and historians.
est findings prove that recognition accuracies can be further • Combination of multiple classifiers: To significantly improve
improved by using CNN and other hybrid techniques. the classification performance, future research should intent
towards combining several classifiers so as to obtain an
8. Challenges and future perspectives ideal combination (ensemble of classifiers). The fusion of
classifiers can be performed in two ways i.e. sequential and
The research works discussed in this survey depict the efforts parallel fusion. The sequential fusion is mostly used for
of several researchers towards the recognition of handwritten In- handling large category set, whereas parallel fusion deals
dic scripts. The recognition accuracies claimed by applying a wide with improving the accuracy classifier.
variety of feature extraction and classification techniques are very • Post-recognition error detection and correction: The character
persuasive. But certain areas like constrained database, multi- recognition segment for handwritten Indic scripts are highly
script and degraded ancient documents, confusing characters and prone to give inaccurate results owing to the structural
many more still need more exploration. Therefore, this section complexities of characters in Indic scripts. So, in order to
highlights the number of challenges and future directions, which improve the recognition accuracy, it is recommended to use
may improve the performance and recognition rates of existing grapheme features, linguistic information or script specific
works when given proper scrutiny. knowledge. The research work reported on the use of post-
recognition error detection and correction is significantly
• Lack of benchmark and standard datasets: The creation of less, hence more endeavour is required to explore this field.
benchmark and standard datasets is one of the most indis-
pensable aspect for any character recognition system. Due 9. Proposed framework
to the inadequacy of comprehensive standardized datasets
for Indic scripts, researcher have created their own in-house Based on the aforementioned future directions, it is contem-
datasets. But these in-house datasets do not provide any plated that for the handwritten recognition of Indic scripts, there
established standards or parameters for evaluating the per- is a huge need to develop novel hybrid classification approaches.
formance of algorithms. Thus, successful research in this So, in this work we have directed our efforts to propose a hybrid
direction needs publicly available standard and accurate classification technique with an aim to outperform existing state-
datasets. Any effort infused in these activities will go a long of-the-art techniques. For better understanding, this proposed
way, furthering the research towards the advancement of algorithm has been demonstrated and explained in detail with
handwritten Indic scripts character recognition systems. the help of an algorithm and example.
• Resolution of confusing and similar characters: There are sev- Deep neural networks (DNN) have proven to outperform tra-
eral characters in Indic scripts that resemble to each other a ditional machine learning techniques for solving complex hand-
lot, thus the recognition of these similar shaped characters written Indic scripts recognition problems. However, building a
becomes quite difficult and challenging. The distinction of handcrafted successful DNN right from scratch requires a lot of
these similar shape characters becomes more complex if problem domain knowledge. For training deep CNN architectures
20 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

Fig. 9. Proposed framework.

millions of parameters are involved, thus a single training run activation function like rectified linear units (ReLu).
may take several days and hence this approach consumes a signif-
⎨Z n = X ,
⎧
n=1
icant amount of computational resources and time. Thus, a novel
framework is proposed based on the improved particle swarm Zn = fn (Yn ), n>1 (3)
Yn = gn (Zn−1 , Kn )
⎩
optimization (PSO) algorithm, to automatically search and create
optimal deep convolutional neural network (CNN) architecture ⎨Yn = Kn Yn−1 ,
⎧
if nth layer is convolution
for the classification and recognition of handwritten Indic scripts
Yn = (⊞x,y , Yn−1 ), if nth layer is pooling (4)
as shown in Fig. 9.
Zn = fn (Yn ),
⎩
if nth layer is fully-connected
The PSO is a population based meta-heuristic global opti-
mization technique commonly used in continuous and discrete The proposed framework consists of improved PSO algorithm
optimization problems. PSO mimics the flying behaviour of flock including the following four stages that help it to search and
of birds searching for food. In PSO algorithm, a single solution construct the optimal CNN architecture. In the first step, particles
is known as particle and the group of all solutions is known as are initialized with random CNN architectures. Then in the next
swarm. The PSO is based on the idea that every particle in the step fitness is evaluated in order to determine the loss. The third
swarm has knowledge about its position, velocity, its personal step deals with velocity computation and finally, the particles are
best configuration (p-best) and global best configuration of the updated leading the search towards best CNN architecture. The
swarm (g-best) [144]. In every iteration, the p-best of particles general layout of the proposed model is shown in Algorithm 1.
and g-best of the swarm are adjusted and updated in such a way The detailed description of improved PSO–CNN is given in the
that all particles lead towards the same optimal position. The following subsections.
velocity and position kth particle of at nth iteration is updated
using Eqs. (1) and (2) respectively. 9.1. Representing CNN architectures with swarm initialization

Vkn+1 = ω ∗ Vkn + c1 r1 ∗ (p − bestkn − Xkn ) + c2 r2 ∗ (g − best n − Xkn ) (1) The initialization of population or swarm is the first stage of
the proposed algorithm. This step constructs N particles corre-
Xkn+1 = Xkn + Vkn+1 (2) sponding to random CNN architectures, with a different number
In CNN classifier [145], layers are aligned together in such a of layers. For every particle, the first layer is always a convolution
manner that output of one layer will be fed as input to the suc- layer (Conv) and the last layer is always a fully-connected layer
ceeding layer. Mathematically, the CNN architecture is depicted in (FC) in order to build feasible CNN architectures. In addition
Eq. (3), where X denotes the input data, Zn indicates the output to this, the fully-connected layer is always placed at the end
of nth layer, the activation function corresponding to nth layer i.e. it cannot be placed in between convolution and pooling lay-
ers. Once, a fully-connected layer is placed in the architecture
is represented as fn (), the kernel operation corresponding to nth
then every other layer that follows it is also a fully-connected
layer is gn (), the kernel for nth layer is Kn and the kernel operation
layer. Typically, the function of fully connected layer classifies the
output before activation function for nth layer is represented by
reduced feature vector obtained from convolution and pooling
Yn . The CNN classifier is mainly composed of three layers, namely
layers.
convolution, pooling and fully-connected layers, represented in
Eq. (4). The convolution layer is responsible for constructing
Example. Three randomly initialized particles are illustrated in
feature map with the help of sliding kernel or filter over the
Fig. 10 with each particle representing a unique CNN architecture.
input data on the basis of stride size. Then, the pooling layer
is responsible for down-sampling on the convolved features in
a x × y window ⊞x,y and thereby further reducing the output 9.2. Evaluate the fitness
parameters. Finally, the fully-connected layer operates in a way
similar to classical ANN. In this step, the flatten output is fed to The fitness is evaluated by compiling the particles into full-
the fully-connected layer which gives the final output using an fledged CNN architectures and then training these architectures
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 21

Algorithm 1 Procedure PSO-CNN

1: Input: The swarm or population size (N), training data (Dtrain ), the number (etrain ) of training epochs, maximum number of iterations
(max)
2: Output: The optimal CNN architecture
3: X = X1 , X2 , . . . , XN //Initialize population with each particle corresponding to random CNN architecture
4: for k = 1 to N do
5: p-bestk = Xk //Initialize the personal best
6: Xk _loss, p-bestk _loss = L (Xk , Dtrain , etrain ) //Evaluate the fitness
7: end for
8: g-best = Xk for all particles m, k ̸ = m and L(Xk , Dtrain , etrain ) < L(Xm , Dtrain , etrain )
9: g-best_loss = Xk _loss
10: for iter = 1 to max do
11: for k = 1 to N do
12: Xk _velocity = Update_Velocity(Xk , ω)
13: Xk = Update_Particle(Xk )
14: Xk _loss = L(Xk , Dtrain , etrain ) // Evaluate the fitness
15: if Xk _loss < p-bestk _loss then
16: p-bestk = Xk
17: p-bestk _loss = Xk _loss
18: if p-bestk _loss < g-best_loss then
19: g-best = Xk
20: g-best_loss = Xk _loss
21: end if
22: end if
23: end for
24: end for
25: return g-best //G-best will return the optimal CNN architecture

Fig. 10. Representation of three particles with different CNN architectures. Fig. 11. Example of computing the difference (g-best - X).

for the total number of training epochs (etrain ). The loss func-
tions computed for different particles are compared in order to
evaluate the fitness. Thus, the aim of the algorithm is to iden-
tify a particle architecture with a minimum loss, irrespective of
the number of hyper-parameters. On the basis of this evaluated
fitness, the g-best and p-bestk are updated in order to direct
the search towards an optimal solution. When the evaluation
terminates, then the obtained g-best will represent the optimal
CNN architecture.
Fig. 12. Example of computing the difference (p-best - X).
9.3. Velocity computation

The velocity of every kth particle (Xk ) depends upon the differ- indicates that this layer position should remain the same while
ences: (g-best – Xk ) and (p-bestk – Xk ), thus we have to compute updating the given particle architecture.
these differences in order to determine the velocity for each The difference is always computed with respect to g-best and
particle. While computing the differences we need to separate p-best, as shown in Figs. 11 and 12 respectively. For example, if g-
FC layers from convolution and pooling layers, in order to avoid best and Xk have different layer types then the result will keep the
the chance of ending-up with FC layers between convolution and layer from g-best with its hyper-parameters. If g-best has fewer
pooling layers. For each particle, the difference depends only on layers as compared to Xk then the corresponding extra layer is
the type of layer. truncated. On the contrary, if g-best has more layers than Xk , then
For example, as shown in Fig. 11 the third layer of both g-best extra layer is padded to the final difference. The same procedure
and Xk is a convolution layer, therefore difference will be 0, which is followed for computing the difference with respect to p-best.
22 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

Table 18
Comparative analysis of the proposed framework with some popular
contemporaries on CMATERdb 3.1.2 database.
Database Work reference Recognition accuracy (%)
Bengali basic Sarkhel et al. [73] 87.28
Character Basu et al. [52] 80.58
CMATERdb 3.1.2 Rabby et al. [65] 98.00
Gupta et al. [94] 86.10
Sarkhel et al. [92] 86.53
Keserwani et al. [85] 98.56
Roy et al. [93] 86.40
Dash et al. [141] 94.78
Present work 98.93

Fig. 13. Example of velocity computation.

Table 19
Comparative analysis of the proposed framework with some popular
contemporaries on CMATERdb 3.1.3 database.
Database Work reference Recognition accuracy (%)
Bengali compound Pramanik and Bag [75] 88.74
Character Keserwani et al. [85] 98.12
CMATERdb 3.1.3 Roy et al. [91] 90.33
Sarkhel et al. [89] 98.12
Sarkhel et al. [92] 78.38
Das et al. [37] 87.50
Present work 98.50

Fig. 14. Example of updated particle architecture.

After computing the differences, the final velocity is evaluated
by drawing uniformly distributed random numbers between [0,
1) and then comparing them with inertia weight (ωi ) for every ith 10.1. Benchmark datasets
iteration. If ωi < r, then the algorithm will choose a layer from
(p-best – X), otherwise it will choose a layer from (g-best – X). The Computational experiments are conducted on two standard
ωi is evaluated for every ith iteration using the following Eq. (5). image classification datasets to analyse the performance of the
ωi = ωi−1 × β (5) proposed model. The benchmark datasets used are CMATERdb
3.1.2 [35] and CMATERdb 3.1.3 [36], briefly discussed in Sec-
where β is the decrement counter, randomly generated between tion 5.2. Furthermore, the script category, type, and size of the
(0, 1). The inertia weight plays a significant role in the fast conver- chosen benchmark datasets are summarized in Table 5. The orig-
gence of proposed algorithm. The improved PSO algorithm works inal images in the benchmark datasets have non-uniform di-
globally for large values, however the algorithm shift towards mensions, so all handwritten images are firstly normalized into
local or optimal search space for small ω values. As shown in 28 × 28 pixel resolution while preserving their aspect ratios. Sub-
Eq. (5), the inertia weight for ith iteration depends upon, the sequently, each 3-channel RGB image of the handwritten charac-
previous value of inertia weight (ωi−1 ) and decrement counter ter dataset is binarized that reduces the number of parameters in
(β ). The inertia weight provides an appropriate direction to the
the proposed architecture.
particles in the swarm towards optimal solution. Thus, this rapid
tuning of inertia weight maintains a good balance between global
exploration and local exploitation. 10.2. Experimental setup and overall performance

Example. An example of final velocity computation is shown in In the designed PSO–CNN model, the parameters setting re-
Fig. 13. lated to the PSO algorithm are based on the conventions used in
PSO communities. Initially, the inertia weight value is predefined
9.4. Update the particle as 0.9, which decreases with each iteration according to Eq. (5).
The population size is initialized to 30 particles with each par-
The particles are updated on the basis of computed velocity ticle representing random CNN architecture, and the maximum
i.e. layers are added to the particle architecture or removed from number of PSO iterations is 20. Since the fitness evaluation is an
the particle architecture with respect to its velocity. expensive process in CNN architecture due to the requirement of
a large amount of training and testing, therefore, the maximum
Example. An example of updating the particle architecture is number of epochs for particle training is kept to 5. Finally, at
shown in Fig. 14. the end of optimization, the obtained optimal CNN architecture
is trained for 100 epochs.
10. Experimental analysis Due to the stochastic nature of the proposed algorithm, the
results are averaged over five independent experimental runs
In order to investigate the performance of the proposed model, for maintaining the consistency in the results to be compared.
the following sub-sections firstly outlines the employed bench- The proposed model has reported significantly improved results
mark datasets. Then the experimental setup and overall perfor- on both the publicly available standard datasets. Furthermore,
mance are illustrated and comprehensively analysed. Finally, the the computational results achieved on the CMATERdb 3.1.2 and
chosen peer competitors are reviewed, and comparative analysis CMATERdb 3.1.3 datasets are provided in Tables 18 and 19, re-
is performed with these state-of-the-art techniques. spectively.
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 23

Fig. A.15. Quality assessment form.

10.3. Comparison with the state-of-the-art analysis has been performed with an intent to determine the
efficiency of recognition schemes by comparing several exist-
In this section, the superiority of the proposed model is proved ing feature extraction and classification methods based on their
by comparing its performance with some other popular con- recognition rates. Finally, we have discussed various challenges
temporary works. Peer competitors for the proposed PSO–CNN in Indic scripts, which provide a way forward for future research.
model, which have been extensively discussed and reviewed in Based on the extensive study conducted in this article, it has
Section 6, are employed to perform the comparisons on the been contemplated that there is a need to develop hybrid feature
chosen standard image classification datasets. Specifically, the extraction and classification approaches for achieving the best
state-of-the-art methods, that have reported promising recogni- accuracy results. So, a novel framework based on the improved
tion accuracy on the chosen benchmarks are MMCNN+SVM [89], PSO algorithm to automatically construct optimal CNN architec-
SL-DCNN [91], Two-pass GA+SVM [37], NSHA+NSGA-II+AFS+ ture has been proposed, which yield better performance than the
SVM [73], Shape decomposition+BP-MLP [75], SVM+RBF [92], existing traditional state-of-the-art techniques.
BornoNet [65], ABCO+SVM [93], CDM+BP-MLP [52], unified- From the perspective of future work, the demand for man and
CNN [85], Multiobjective harmony search+SVM [94] and machine interaction has elevated this field. The researchers have
Tetrolet+SCC+Nearest Neighbour [141]. Finally, Tables 18 and immensely contributed towards the advancement of character
19 illustrates the comparative analysis of the proposed PSO–CNN recognition systems that can be implemented practically, to aid
model with the above benchmark state-of-the-art algorithms. society by enhancing the interface between humans and com-
puters. Intensive research available so far in this field is mainly
11. Conclusion focused on scripts like Chinese, Latin, Arabic and Japanese, and
the work done on Indic scripts is still at its infancy. So, this survey
In this paper, we have presented an exhaustive survey of vari- has been performed to compile significant contributions in the
ous feature extraction and classification techniques for the recog- field that will encourage and provide in-depth information to
nition of handwritten Indic scripts. It also provides a detailed the novice researchers in this field of handwritten Indic scripts
description of related benchmark and standardized datasets. It is recognition.
clearly evident from the survey that the recognition accuracy of
the model relies on the size and quality of the dataset. The lack of Declaration of competing interest
significantly good benchmark datasets is one of the basic issues in
several Indic scripts. So, there is a necessity for the construction The authors declare that they have no known competing finan-
of exhaustive and standardized datasets for Indic scripts which cial interests or personal relationships that could have appeared
provides a novel future aspect in this direction. A systematic to influence the work reported in this paper.
24 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

Appendix A. Quality assessment form HOG Histogram of Oriented Gradient

ISI Indian Statistical Institute
ISI-HBND ISI Handwritten Bangla Numeral Database
The quality assessment form is shown in Fig. A.15 ISI-HDND ISI Handwritten Devanagari Numeral Database
ISI-HOND ISI Handwritten Oriya Numeral Database
KHTD Kannada Handwritten Text Dataset
Appendix B. Data items extracted from all papers
KNN K Nearest Neighbour
LBP Local Binary Patterns
Data item Description LR Longest Run
LSTM Long Short Term Memory
Bibliographic data Author, title, source and publication MDP Modified Division Point
year MIL Mirror Image Learning
Articles category Journal or conference proceeding MLP Multi Layer Perceptron
Aim of study It outlines the objectives and research MMCNN Multicolumn Multiscale Convolutional Neural
area on which the article is focused Network
Approaches for Feature extraction and classification MQDF Modified Quadratic Discriminant Function
handwritten Indic methods available for handwritten NFC Neuro Fuzzy Classifier
scripts recognition Indic scripts recognition NN Neural Network
Datasets for It indicates the dataset name, script NSGA Non-Dominated Sorting Genetic Algorithm
handwritten Indic type, dataset size, image dimensions NSHA Non-Dominated Sorting Harmony-Search
scripts recognition etc. Algorithm
Data Analysis Important parameters such as NVD Normalized Vector Distance
recognition accuracy, error rate, run OCR Optical Character Recognition
time etc. PCA Principal Component Analysis
Tools It determines the tools used in the PSO Particle Swarm Optimization
studies to perform experimental QDF Quadratic Discriminant Function
evaluations RBF Radial Basis Function
Results It refers to the results and conclusion RLC Run Length Count
obtained from primary study SCC Sparse Concept Coding
SCI Science Citation Index
SD Standard deviation
Appendix C. List of abbreviations
SGD Stochastic Gradient Decent
SIFT Scale Invariant Feature Transform
ABCO Artificial Bee Colony Optimization SLDCNN Supervised Layerwise Deep Convolutional
AFS Axiomatic Fuzzy Set Neural Network
ANFC Adaptive Neuro Fuzzy Classifier SVM Support Vector Machine
ANN Artificial Neural Network TM Template Matching
BDD Background Directional Distribution WEF Wavelet Energy Features
BESAC Binary External Symmetry Axis Constellation ZD Zoning Density
BPNN Back Propagation Neural Network
CC Chain Code
CDC Connected Disconnected Components
References
CDM Character Digit Modified Shape
CDPDF Contour Direction Probability Distribution
[1] J. Mantas, An overview of character recognition methodologies, Pattern
Function Recognit. 19 (6) (1986) 425–430.
CL Closed Loops [2] S. Mori, C.Y. Suen, K. Yamamoto, Historical review of OCR research and
CMATER Center for Microprocessor Applications for development, Proc. IEEE 80 (7) (1992) 1029–1058.
Training Education and Research [3] A.K. Bhunia, S. Mukherjee, A. Sain, A.K. Bhunia, P.P. Roy, U. Pal, Indic
handwritten script identification using offline-online multi-modal deep
CMNN Class Modular Neural Network
network, Inf. Fusion 57 (2020) 1–14.
DCNN Deep Convolutional Neural Network [4] J. Qiao, G. Wang, W. Li, M. Chen, An adaptive deep Q-learning strategy
DCT Discrete Cosine Transformation for handwritten digit recognition, Neural Netw. 107 (2018) 61–71.
DHCD Devanagari Handwritten Character Dataset [5] Y. Mizukami, A handwritten Chinese character recognition system us-
DT Decision Tree ing hierarchical displacement extraction based on directional features,
DWT Discrete Wavelet Transformation Pattern Recognit. Lett. 19 (7) (1998) 595–604.
[6] A. Qaroush, B. Jaber, K. Mohammad, M. Washaha, E. Maali, N. Nayef, An
ED Euclidean distance
efficient, font independent word and character segmentation algorithm
ELM Extreme Learning Machine for printed arabic text, J. King Saud Univ. - Comput. Inf. Sci. (2019).
FBT Fan Beam Transformation [7] Z. Xie, Y. Huang, L. Jin, Y. Liu, Y. Zhu, L. Gao, X. Zhang, Weakly supervised
FFANN Feed Forward Artificial Neural Network precise segmentation for historical document images, Neurocomputing
FFT Fast Fourier Transformation 350 (2019) 271–281.
[8] S. Naz, K. Hayat, M.I. Razzak, M.W. Anwar, S.A. Madani, S.U. Khan,
FR Fisher Ratio
The optical character recognition of Urdu-like cursive scripts, Pattern
FZ-NVD Fuzzy Zoning based Normalized Vector Distance Recognit. 47 (3) (2014) 1229–1248.
GA Genetic Algorithm [9] G.A. Montazer, H.Q. Saremi, V. Khatibi, A neuro-fuzzy inference engine
GPXNP Gabor Phase XNOR Pattern for Farsi numeral characters recognition, Expert Syst. Appl. 37 (9) (2010)
GRNN Generalized Regression Neural Network 6327–6337.
[10] H. Akram, S. Khalid, et al., Using features of local densities, statistics
HMM Hidden Markov Model
and HMM toolkit (HTK) for offline Arabic handwriting text recognition,
J. Electr. Syst. Inf. Technol. 4 (3) (2017) 387–396.
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 25

[11] U. Pal, R. Jayadevan, N. Sharma, Handwriting recognition in indian [39] U. Bhattacharya, B. Chaudhuri, Umapada, Databases for research on recog-
regional scripts: a survey of offline techniques, ACM Trans. Asian Lang. nition of handwritten characters of Indian scripts, in: Eighth International
Inf. Process. (TALIP) 11 (1) (2012) 1–35. Conference on Document Analysis and Recognition (ICDAR’05), IEEE,
[12] U. Pal, B. Chaudhuri, Indian script character recognition: a survey, Pattern 2005, pp. 789–793.
Recognit. 37 (9) (2004) 1887–1899. [40] HPLabs, Isolated handwritten telugu character dataset, 2006, URL http:
[13] A. Datta, A generalized formal approach for description and analysis of //lipitk.sourceforge.net/datasets/teluguchardata.htm.
major Indian scripts, IETE J. Res. 30 (6) (1984) 155–161. [41] M. Agrawal, A.S. Bhaskarabhatla, S. Madhvanath, Data collection for hand-
[14] R. Sharma, B.N. Kaushik, N.K. Gondhi, Devanagari and gurmukhi script writing corpus creation in Indic scripts, in: International Conference on
recognition in the context of machine learning classifiers, J. Artif. Intell. Speech and Language Technology and Oriental COCOSDA (ICSLT-COCOSDA
11 (2) (2018) 65–70. 2004), New Delhi, India (November 2004), Citeseer, 2004.
[15] P.K. Singh, R. Sarkar, M. Nasipuri, Offline script identification from [42] A. Alaei, P. Nagabhushan, U. Pal, A benchmark kannada handwritten
multilingual indic-script documents: a state-of-the-art, Comp. Sci. Rev. document dataset and its segmentation, in: 2011 International Conference
15 (2015) 1–28. on Document Analysis and Recognition, IEEE, 2011, pp. 141–145.
[16] M. Yadav, R.K. Purwar, M. Mittal, Handwritten Hindi character [43] K.S. Dash, N.B. Puhan, G. Panda, Odia character recognition: a directional
recognition: a review, IET Image Process. 12 (11) (2018) 1919–1933. review, Artif. Intell. Rev. 48 (4) (2017) 473–497.
[17] K. Ubul, G. Tursun, A. Aysa, D. Impedovo, G. Pirlo, T. Yibulayin, Script [44] A. Chahi, Y. Ruichek, R. Touahni, et al., An effective and conceptu-
identification of multi-script documents: a survey, IEEE Access 5 (2017) ally simple feature representation for off-line text-independent writer
6546–6559. identification, Expert Syst. Appl. 123 (2019) 357–376.
[18] S. Bag, G. Harit, A survey on optical character recognition for Bangla and [45] P. Inkeaw, J. Bootkrajang, S. Marukatat, T. Gonçalves, J. Chaijaruwanich,
Devanagari scripts, Sadhana 38 (1) (2013) 133–168. Recognition of similar characters using gradient features of discriminative
[19] R. Jayadevan, S.R. Kolhe, P.M. Patil, U. Pal, Offline recognition of de- regions, Expert Syst. Appl. (2019).
vanagari script: A survey, IEEE Trans. Syst. Man Cybern. C 41 (6) (2011) [46] M.A. Dhali, C.N. Jansen, J.W. de Wit, L. Schomaker, Feature-extraction
782–796. methods for historical manuscript dating based on writing style
[20] B. Chaudhuri, U. Pal, A complete printed Bangla OCR system, Pattern development, Pattern Recognit. Lett. 131 (2020) 413–420.
Recognit. 31 (5) (1998) 531–549. [47] Ø.D. Trier, A.K. Jain, T. Taxt, Feature extraction methods for character
[21] S.M. Obaidullah, K. Santosh, C. Halder, N. Das, K. Roy, Automatic Indic recognition-a survey, Pattern Recognit. 29 (4) (1996) 641–662.
script identification from handwritten documents: page, block, line and [48] B. Bataineh, S.N.H.S. Abdullah, K. Omar, A novel statistical feature ex-
word-level approach, Int. J. Mach. Learn. Cybern. 10 (1) (2019) 87–106. traction method for textual images: Optical font recognition, Expert Syst.
[22] A.K. Bhunia, P.P. Roy, A. Mohta, U. Pal, Cross-language framework for Appl. 39 (5) (2012) 5470–5477.
word recognition and spotting of Indic scripts, Pattern Recognit. 79 (2018) [49] C.-Z. Shi, S. Gao, M.-T. Liu, C.-Z. Qi, C.-H. Wang, B.-H. Xiao, Stroke detector
12–31. and structure based models for character recognition: a comparative
[23] N.H. Khan, A. Adnan, Urdu optical character recognition systems: Present study, IEEE Trans. Image Process. 24 (12) (2015) 4952–4964.
contributions and future directions, IEEE Access 6 (2018) 46019–46046. [50] N. Sharma, U. Pal, F. Kimura, S. Pal, Recognition of off-line handwritten
[24] N.R. Soora, P.S. Deshpande, A novel local skew correction and segmenta- devnagari characters using quadratic classifier, in: Computer Vision,
tion approach for printed multilingual Indian documents, Alexandria Eng. Graphics and Image Processing, Springer, 2006, pp. 805–816.
J. 57 (3) (2018) 1609–1618. [51] S. Arora, D. Bhattacharjee, M. Nasipuri, D.K. Basu, M. Kundu, Combining
[25] V. Bansal, R. Sinha, Segmentation of touching and fused devanagari multiple feature extraction techniques for handwritten devnagari char-
characters, Pattern Recognit. 35 (4) (2002) 875–893. acter recognition, in: 2008 IEEE Region 10 and the Third International
[26] R. Ghosh, C. Vamshi, P. Kumar, RNN based online handwritten word Conference on Industrial and Information Systems, IEEE, 2008, pp. 1–6.
recognition in Devanagari and Bengali scripts using horizontal zoning, [52] S. Basu, N. Das, R. Sarkar, M. Kundu, M. Nasipuri, D.K. Basu, A hierar-
Pattern Recognit. 92 (2019) 203–218. chical approach to recognition of handwritten Bangla characters, Pattern
[27] P.P. Roy, A.K. Bhunia, A. Das, P. Dey, U. Pal, HMM-based Indic handwritten Recognit. 42 (7) (2009) 1467–1484.
word recognition using zone segmentation, Pattern Recognit. 60 (2016) [53] N. Das, B. Das, R. Sarkar, S. Basu, M. Kundu, M. Nasipuri, Handwritten
1057–1075. bangla basic and compound character recognition using MLP and SVM
[28] N. Das, J.M. Reddy, R. Sarkar, S. Basu, M. Kundu, M. Nasipuri, D.K. Basu, A classifier, 2010, arXiv preprint arXiv:1002.4040.
statistical–topological feature combination for recognition of handwritten [54] B. Singh, A. Mittal, D. Ghosh, An evaluation of different feature extractors
numerals, Appl. Soft Comput. 12 (8) (2012) 2486–2495. and classifiers for offline handwritten devnagari character recognition, J.
[29] P.K. Singh, R. Sarkar, N. Das, S. Basu, M. Kundu, M. Nasipuri, Bench- Pattern Recognit. Res. 2 (2011) 269–277.
mark databases of handwritten Bangla-Roman and Devanagari-Roman [55] K. Kale, S. Chavan, M. Kazi, Y. Rode, Handwritten and printed devana-
mixed-script document images, Multimedia Tools Appl. 77 (7) (2018) gari compound using multiclass svm classifier with orthogonal moment
8441–8473. feature, Int. J. Comput. Appl. 71 (24) (2013).
[30] S. Basu, C. Chaudhuri, M. Kundu, M. Nasipuri, D.K. Basu, Text line [56] M. Kumar, R. Sharma, M. Jindal, Efficient feature extraction techniques
extraction from multi-skewed handwritten documents, Pattern Recognit. for offline handwritten Gurmukhi character recognition, Nat. Acad. Sci.
40 (6) (2007) 1825–1839. Lett. 37 (4) (2014) 381–391.
[31] U. Bhattacharya, B.B.U. Chaudhuri, Handwritten numeral databases of [57] S. Bag, G. Harit, P. Bhowmick, Recognition of Bangla compound char-
Indian scripts and multistage recognition of mixed numerals, IEEE Trans. acters using structural decomposition, Pattern Recognit. 47 (3) (2014)
Pattern Anal. Mach. Intell. 31 (3) (2008) 444–457. 1187–1201.
[32] V.P. Agnihotri, Offline handwritten devanagari script recognition, IJ Inf. [58] A. Singh, K.A. Maring, Handwritten devanagari character recognition using
Technol. Comput. Sci. 8 (1) (2012) 37–42. SVM and ANN, Int. J. Adv. Res. Comput. Commun. Eng. 4 (8) (2015)
[33] S. Acharya, A.K. Pant, P.K. Gyawali, Deep learning based large scale 123–128.
handwritten devanagari character recognition, in: 2015 9th Interna- [59] S. Shelke, S. Apte, A fuzzy based classification scheme for unconstrained
tional Conference on Software, Knowledge, Information Management and handwritten devanagari character recognition, in: 2015 International
Applications (SKIMA), IEEE, 2015, pp. 1–6. Conference on Communication, Information & Computing Technology
[34] N. Das, R. Sarkar, S. Basu, M. Kundu, M. Nasipuri, D.K. Basu, A genetic (ICCICT), IEEE, 2015, pp. 1–6.
algorithm based region sampling for selection of local features in hand- [60] O. Surinta, M.F. Karaaba, L.R. Schomaker, M.A. Wiering, Recognition of
written digit recognition application, Appl. Soft Comput. 12 (5) (2012) handwritten characters using local gradient feature descriptors, Eng. Appl.
1592–1606. Artif. Intell. 45 (2015) 405–414.
[35] N. Das, S. Basu, R. Sarkar, M. Kundu, M. Nasipuri, et al., An improved [61] A. Aggarwal, K. Singh, Handwritten gurmukhi character recognition, in:
feature descriptor for recognition of handwritten bangla alphabet, 2015, 2015 International Conference on Computer, Communication and Control
arXiv preprint arXiv:1501.05497. (IC4), IEEE, 2015, pp. 1–5.
[36] N. Das, K. Acharya, R. Sarkar, S. Basu, M. Kundu, M. Nasipuri, A benchmark [62] M. Kumar, M. Jindal, R. Sharma, Offline handwritten gurmukhi character
image database of isolated bangla handwritten compound characters, Int. recognition: analytical study of different transformations, Proc. Nat. Acad.
J. Doc. Anal. Recognit. (IJDAR) 17 (4) (2014) 413–431. Sci. India Sect. A 87 (1) (2017) 137–143.
[37] N. Das, R. Sarkar, S. Basu, P.K. Saha, M. Kundu, M. Nasipuri, Handwritten [63] M. Yadav, R. Purwar, Hindi handwritten character recognition using
Bangla character recognition using a soft computing paradigm embedded multiple classifiers, in: 2017 7th International Conference on Cloud
in two pass approach, Pattern Recognit. 48 (6) (2015) 2054–2071. Computing, Data Science & Engineering-Confluence, IEEE, 2017, pp.
[38] R. Sarkar, N. Das, S. Basu, M. Kundu, M. Nasipuri, D.K. Basu, CMATERdb1: a 149–154.
database of unconstrained handwritten Bangla and Bangla–English mixed [64] N. Kumar, S. Gupta, A novel handwritten Gurmukhi character recognition
script document image, Int. J. Doc. Anal. Recognit. (IJDAR) 15 (1) (2012) system based on deep neural networks, Int. J. Pure Appl. Math. 117 (21)
71–83. (2017) 663–678.
26 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302

[65] A.S.A. Rabby, S. Haque, S. Islam, S. Abujar, S.A. Hossain, BornoNet: Bangla [92] R. Sarkhel, A.K. Saha, N. Das, An enhanced harmony search method
handwritten characters recognition using convolutional neural network, for bangla handwritten character recognition using region sampling, in:
Procedia Comput. Sci. 143 (2018) 528–535. 2015 IEEE 2nd International Conference on Recent Trends in Information
[66] M. Jangid, S. Srivastava, Handwritten devanagari character recognition Systems (ReTIS), IEEE, 2015, pp. 325–330.
using layer-wise training of deep convolutional neural networks and [93] A. Roy, N. Das, R. Sarkar, S. Basu, M. Kundu, M. Nasipuri, Region selection
adaptive gradient methods, J. Imaging 4 (2) (2018) 41. in handwritten character recognition using artificial bee colony optimiza-
[67] U. Pal, T. Wakabayashi, F. Kimura, Comparative study of devnagari tion, in: 2012 Third International Conference on Emerging Applications
handwritten character recognition using different feature and classi- of Information Technology, IEEE, 2012, pp. 183–186.
fiers, in: 2009 10th International Conference on Document Analysis and [94] A. Gupta, R. Sarkhel, N. Das, M. Kundu, Multiobjective optimization for
Recognition, IEEE, 2009, pp. 1111–1115. recognition of isolated handwritten Indic scripts, Pattern Recognit. Lett.
[68] G. Singh, S. Lehri, Recognition of handwritten hindi characters using 128 (2019) 318–325.
backpropagation neural network, Int. J. Comput. Sci. Inf. Technol. 3 (4) [95] U. Pal, N. Sharma, T. Wakabayashi, F. Kimura, Handwritten character
(2012) 4892–4895. recognition of popular south Indian scripts, in: Summit on Arabic and
[69] U. Jain, D. Sharma, Recognition of isolated handwritten characters of Chinese Handwriting Recognition, Springer, 2006, pp. 251–264.
gurumukhi script using neocognitron, Int. J. Comput. Appl. 10 (8) (2010). [96] S. Sangame, R. Ramteke, R. Benne, Recognition of isolated handwritten
[70] K.S. Siddharth, R. Dhir, R. Rani, Handwritten Gurumukhi character recog- Kannada vowels, Adv. Comput. Res. 1 (2) (2009) 52–55.
nition using zoning density and background directional distribution [97] B. Dhandra, M. Hangarge, G. Mukarambi, Spatial features for handwritten
features, Int. J. Comput. Sci. Inf. Technol. 2 (3) (2011) 1036–1041. kannada and english character recognition, Int. J. Comput. Appl. (2010)
[71] G. Sinha, R. Rani, R. Dhir, Handwritten gurmukhi character recognition 146–150, Special Issue on RTIPPR-10.
using K-NN and SVM classifier, Int. J. Adv. Res. Comput. Sci. Softw. Eng. [98] G. Rajput, R. Horakeri, Shape descriptors based handwritten character
2 (6) (2012) 288–293. recognition engine with application to Kannada characters, in: 2011 2nd
[72] S. Singh, A. Aggarwal, R. Dhir, Use of gabor filters for recognition of International Conference on Computer and Communication Technology
handwritten gurmukhi character, Int. J. Adv. Res. Comput. Sci. Softw. Eng. (ICCCT-2011), IEEE, 2011, pp. 135–141.
2 (5) (2012). [99] G. Mukarambi, B. Dhandra, M. Hangarge, A zone based character recog-
[73] R. Sarkhel, N. Das, A.K. Saha, M. Nasipuri, A multi-objective approach nition engine for kannada and english scripts, Procedia Eng. 38 (2012)
towards cost effective isolated handwritten Bangla character and digit 3292–3299.
recognition, Pattern Recognit. 58 (2016) 172–189. [100] S.A. Vaidya, B.R. Bombade, A novel approach of handwritten character
[74] U. Pal, T. Wakabayashi, F. Kimura, Handwritten Bangla compound charac- recognition using positional feature extraction, Int. J. Comput. Sci. Mobile
ter recognition using gradient feature, in: 10th International Conference Comput. 2 (6) (2013) 179–186.
on Information Technology (ICIT 2007), IEEE, 2007, pp. 208–213. [101] B. Dhandra, G. Mukarambi, Kannada handwritten vowels recognition
[75] R. Pramanik, S. Bag, Shape decomposition-based handwritten compound based on normalized chain code and wavelet filters, Int. J. Comput. Appl.
character recognition for Bangla OCR, J. Vis. Commun. Image Represent. 975 (2014) 8887.
50 (2018) 123–134. [102] P.N. Sastry, T.V. Lakshmi, N.K. Rao, T. Rajinikanth, A. Wahab, Telugu
[76] A. Garg, M.K. Jindal, A. Singh, Offline handwritten gurmukhi character handwritten character recognition using zoning features, in: 2014 Inter-
recognition: k-NN vs. SVM classifier, Int. J. Inf. Technol. (2019) 1–8. national Conference on IT Convergence and Security (ICITCS), IEEE, 2014,
[77] J.A. Sánchez, V. Romero, A.H. Toselli, M. Villegas, E. Vidal, A set of pp. 1–4.
benchmarks for handwritten text recognition on historical documents, [103] T.V. Lakshmi, Multi-stage strategy to classify handwritten characters of
Pattern Recognit. 94 (2019) 122–134. telugu, Int. J. Curr. Res. Rev. 9 (20) (2017) 39.
[78] S. Mandal, S.M. Prasanna, S. Sundaram, GMM posterior features for [104] S. Karthik, K.S. Murthy, Deep belief network based approach to recognize
improving online handwriting recognition, Expert Syst. Appl. 97 (2018) handwritten Kannada characters using distributed average of gradients,
421–433. Cluster Comput. 22 (2) (2019) 4673–4681.
[79] M. Labani, P. Moradi, F. Ahmadizar, M. Jalili, A novel multivariate filter [105] N.S. Rani, N. Chandan, A.S. Jain, H. Kiran, Deformed character recognition
method for feature selection in text classification problems, Eng. Appl. using convolutional neural networks, Int. J. Eng. Technol. 7 (3) (2018)
Artif. Intell. (ISSN: 0952-1976) 70 (2018) 25–37. 1599–1604.
[80] J.M. Alonso-Weber, M. Sesmero, A. Sanchis, Combining additive input [106] S. Pasha, M. Padma, Handwritten kannada character recognition using
noise annealing and pattern transformations for improved handwritten wavelet transform and structural features, in: 2015 International Con-
character recognition, Expert Syst. Appl. 41 (18) (2014) 8180–8188. ference on Emerging Research in Electronics, Computer Science and
[81] T. Akter, S. Desai, Developing a predictive model for nanoimprint Technology (ICERECT), IEEE, 2015, pp. 346–351.
lithography using artificial neural networks, Mater. Des. 160 (2018) [107] S. Angadi, S.H. Angadi, Structural features for recognition of hand written
836–848. kannada character based on SVM, Int. J. Comput. Sci. Eng. Inf. Technol. 5
[82] P. Ajmire, Handwritten devanagari vowel recognition using artificial (2) (2015) 25–32.
neural network, Int. J. Adv. Res. Comput. Sci. 8 (7) (2017). [108] P.N. Sastry, R. Krishnan, B.V.S. Ram, Classification and identification of
[83] D. Khanduja, N. Nain, S. Panwar, A hybrid feature extraction algorithm Telugu handwritten characters extracted from palm leaves using decision
for devanagari script, ACM Trans. Asian Low-Resour. Lang. Inf. Process. tree approach, J. Appl. Eng. Sci. 5 (3) (2010) 22–32.
15 (1) (2016) 2. [109] C.N. Manisha, E.S. Reddy, Y.S. Krishna, Glyph-based recognition of offline
[84] M. Jangid, S. Srivastava, Deep convnet with different stochastic optimiza- handwritten Telugu characters: GBRoOHTC, in: 2016 IEEE International
tions for handwritten devanagari character, in: Advances in Computer Conference on Computational Intelligence and Computing Research
Communication and Computational Sciences, Springer, 2019, pp. 51–60. (ICCIC), IEEE, 2016, pp. 1–6.
[85] P. Keserwani, T. Ali, P.P. Roy, Handwritten bangla character and numeral [110] L.R. Ragha, M. Sasikumar, Feature analysis for handwritten kannada
recognition using convolutional neural network for low-memory GPU, Int. kagunita recognition, Int. J. Comput. Theory Eng. 3 (1) (2011) 94.
J. Mach. Learn. Cybern. 10 (12) (2019) 3485–3497. [111] A. Angadi, V. Kumari Vatsavayi, S. Keerthi Gorripati, A deep learning
[86] V.N. Vapnik, The nature of statistical learning, Theory (1995). approach to recognize handwritten telugu character using convolution
[87] S.R. Narang, M. Jindal, S. Ahuja, M. Kumar, On the recognition of neural networks, Int. J. Inf. Syst. Manage. Sci. 1 (2) (2018).
devanagari ancient handwritten characters using SIFT and gabor features, [112] J.R. Prasad, U.V. Kulkarni, R.S. Prasad, Template matching algorithm for
Soft Comput. (2020). gujrati character recognition, in: 2009 Second International Conference on
[88] A. Kataria, M. Singh, A review of data classification using k-nearest Emerging Trends in Engineering & Technology, IEEE, 2009, pp. 263–268.
neighbour algorithm, Int. J. Emerg. Technol. Adv. Eng. 3 (6) (2013) [113] C. Patel, A. Desai, Gujarati handwritten character recognition using hybrid
354–360. method based on binary tree-classifier and k-nearest neighbour, Int. J.
[89] R. Sarkhel, N. Das, A. Das, M. Kundu, M. Nasipuri, A multi-scale deep Eng. Res. Technol. (IJERT) 2 (6) (2013) 2337–2345.
quad tree based feature extraction method for the recognition of isolated [114] J.R. Prasad, U. Kulkarni, Gujarati character recognition using adaptive
handwritten characters of popular indic scripts, Pattern Recognit. 71 neuro fuzzy classifier with fuzzy hedges, Int. J. Mach. Learn. Cybern. 6
(2017) 78–93. (5) (2015) 763–775.
[90] H. Kaur, S. Rani, Handwritten gurumukhi character recognition using [115] J.R. Prasad, U. Kulkarni, Gujrati character recognition using weighted k-
convolution neural network, Int. J. Comput. Intell. Res. 13 (5) (2017) NN and mean χ 2 distance measure, Int. J. Mach. Learn. Cybern. 6 (1)
933–943. (2015) 69–82.
[91] S. Roy, N. Das, M. Kundu, M. Nasipuri, Handwritten isolated Bangla [116] H.R. Thaker, C. Kumbharana, Structural feature extraction to recognize
compound character recognition: A new benchmark using a novel deep some of the offline isolated handwritten gujarati characters using decision
learning approach, Pattern Recognit. Lett. 90 (2017) 15–21. tree classifier, Int. J. Comput. Appl. (2014).
 R. Sharma and B. Kaushik / Computer Science Review 38 (2020) 100302 27

[117] U. Pal, T. Wakabayashi, F. Kimura, A system for off-line Oriya handwritten [132] G. Raju, B.S. Moni, M.S. Nair, A novel handwritten character recognition
character recognition using curvature feature, in: 10th International system using gradient based features and run length count, Sadhana 39
Conference on Information Technology (ICIT 2007), IEEE, 2007, pp. (6) (2014) 1333–1355.
227–229. [133] K. Manjusha, M.A. Kumar, K. Soman, On developing handwritten character
[118] D. Padhi, Novel hybrid approach for odia handwritten character image database for malayalam language script, Eng. Sci. Technol. Int. J.
recognition system, IJARCSSE 2 (5) (2012). 22 (2) (2019) 637–645.
[119] K.S. Dash, N.B. Puhan, G. Panda, BESAC: Binary external symmetry [134] T. Wakabayashi, U. Pal, F. Kimura, Y. Miyake, F-ratio based weighted
axis constellation for unconstrained handwritten character recognition, feature extraction for similar shape character recognition, in: 2009 10th
Pattern Recognit. Lett. 83 (2016) 413–422. International Conference on Document Analysis and Recognition, IEEE,
[120] A. Sethy, P.K. Patra, D.R. Nayak, Off-line odia handwritten character 2009, pp. 196–200.
recognition: A hybrid approach, in: Computational Signal Processing and [135] P. Jino, J. John, K. Balakrishnan, Offline handwritten malayalam character
Analysis, Springer, 2018, pp. 247–257. recognition using stacked LSTM, in: 2017 International Conference on In-
[121] U. Bhattacharya, S. Ghosh, S. Parui, A two stage recognition scheme telligent Computing, Instrumentation and Control Technologies (ICICICT),
for handwritten tamil characters, in: Ninth International Conference on IEEE, 2017, pp. 1587–1590.
Document Analysis and Recognition (ICDAR 2007), Vol. 1, IEEE, 2007, pp. [136] M. Manuel, S. Saidas, Handwritten malayalam character recognition using
511–515. curvelet transform and ANN, Int. J. Comput. Appl. 121 (6) (2015).
[122] N. Shanthi, K. Duraiswamy, A novel SVM-based handwritten tamil [137] S.M. Salaken, A. Khosravi, T. Nguyen, S. Nahavandi, Extreme learning
character recognition system, Pattern Anal. Appl. 13 (2) (2010) 173–180. machine based transfer learning algorithms: A survey, Neurocomputing
[123] A. Subashini, N. Kodikara, Bag-of-keypoints approach for tamil handwrit- 267 (2017) 516–524.
ten character recognition using SVMs, in: 2011 International Conference [138] M.A.R. Raj, S. Abirami, Structural representation-based off-line tamil
on Advances in ICT for Emerging Regions (ICTer), IEEE, 2011, pp. 102–107. handwritten character recognition, Soft Comput. 24 (2) (2020) 1447–
[124] T.M. Jose, A. Wahi, Recognition of tamil handwritten characters us- 1472.
ing daubechies wavelet transforms and feed-forward backpropagation [139] A.K. Sharma, P. Thakkar, D.M. Adhyaru, T.H. Zaveri, Handwritten gujarati
network, Int. J. Comput. Appl. 64 (8) (2013). character recognition using structural decomposition technique, Pattern
[125] S. Abirami, V. Essakiammal, R. Baskaran, Statistical features based char- Recognit. Image Anal. 29 (2) (2019) 325–338.
acter recognition for offline handwritten tamil document images using [140] J. Pareek, D. Singhania, R.R. Kumari, S. Purohit, Gujarati handwritten
HMM, Int. J. Comput. Vis. Robot. 5 (4) (2015) 422–440. character recognition from text images, Procedia Comput. Sci. 171 (2020)
[126] S. Shyni, M.A.R. Raj, S. Abirami, Offline Tamil handwritten character 514–523.
recognition using sub line direction and bounding box techniques, Indian [141] K.S. Dash, N. Puhan, G. Panda, Sparse concept coded tetrolet transform
J. Sci. Technol. 8 (S7) (2015) 110–116. for unconstrained odia character recognition, 2020, arXiv preprint arXiv:
[127] M.A.R. Raj, S. Abirami, Offline Tamil handwritten character recognition 2004.01551.
using statistical features, Adv. Nat. Appl. Sci. 9 (6 SE) (2015) 367–375. [142] B. Kavitha, C. Srimathi, Benchmarking on offline handwritten tamil
[128] R. Raj, M. Antony, S. Abirami, Offline tamil handwritten character character recognition using convolutional neural networks, J. King Saud
recognition using statistical based quad tree, 2016. Univ.-Comput. Inf. Sci. (2019).
[129] B.S. Moni, G. Raju, Modified quadratic classifier and directional features [143] N.T. Kishna, S. Francis, Intelligent tool for malayalam cursive handwritten
for handwritten malayalam character recognition, Int. J. Comput. Appl. character recognition using artificial neural network and hidden Markov
(2011) 30–34. model, in: 2017 International Conference on Inventive Computing and
[130] B.P. Chacko, V.V. Krishnan, G. Raju, P.B. Anto, Handwritten character Informatics (ICICI), IEEE, 2017, pp. 595–598.
recognition using wavelet energy and extreme learning machine, Int. J. [144] G. Xu, K. Luo, G. Jing, X. Yu, X. Ruan, J. Song, On convergence analysis of
Mach. Learn. Cybern. 3 (2) (2012) 149–161. multi-objective particle swarm optimization algorithm, European J. Oper.
[131] J. Jomy, K. Balakrishnan, K. Pramod, A system for offline recognition of Res. (2020).
handwritten characters in Malayalam script, Int. J. Image Graph. Signal [145] J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu, X. Wang,
Process. 5 (4) (2013) 53. G. Wang, J. Cai, et al., Recent advances in convolutional neural networks,
Pattern Recognit. 77 (2018) 354–377.