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

net/publication/343273822

Handwritten Optical Character Recognition (OCR): A Comprehensive

Systematic Literature Review (SLR)

Article  in  IEEE Access · July 2020

DOI: 10.1109/ACCESS.2020.3012542

CITATIONS READS

215 6,548

4 authors, including:

Jamshed Memon Rizwan Ahmed Khan

Kudelski Security Salim Habib University Karachi and University of Lyon France
29 PUBLICATIONS   739 CITATIONS    45 PUBLICATIONS   908 CITATIONS   

SEE PROFILE SEE PROFILE

Mueen Uddin
University of Doha for Science & Technology Qatar
122 PUBLICATIONS   2,526 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Multipath Routing Protocol for Mobile Ad-hoc Networks View project

Recognizing expressions of children in real life scenarios View project

All content following this page was uploaded by Jamshed Memon on 14 August 2020.

The user has requested enhancement of the downloaded file.

Received June 24, 2020, accepted July 16, 2020, date of publication July 28, 2020, date of current version August 14, 2020.
Digital Object Identifier 10.1109/ACCESS.2020.3012542

Handwritten Optical Character Recognition

(OCR): A Comprehensive Systematic
Literature Review (SLR)
JAMSHED MEMON 1, MAIRA SAMI 2, RIZWAN AHMED KHAN 3, AND MUEEN UDDIN 4
1 School of Computing, Quest International University Perak, Ipoh 30250, Malaysia
2 Department of Computer Science, Shaheed Zulfiqar Ali Bhutto Institute of Science and Technology, Karachi 75600, Pakistan
3 Faculty of IT, Barrett Hodgson University, Karachi 74900, Pakistan
4 Department of Software Engineering, Faculty of Science and Technology, Ilma University, Karachi 75190, Pakistan

Corresponding author: Maira Sami ([email protected])

ABSTRACT Given the ubiquity of handwritten documents in human transactions, Optical Character
Recognition (OCR) of documents have invaluable practical worth. Optical character recognition is a science
that enables to translate various types of documents or images into analyzable, editable and searchable data.
During last decade, researchers have used artificial intelligence / machine learning tools to automatically
analyze handwritten and printed documents in order to convert them into electronic format. The objective of
this review paper is to summarize research that has been conducted on character recognition of handwritten
documents and to provide research directions. In this Systematic Literature Review (SLR) we collected,
synthesized and analyzed research articles on the topic of handwritten OCR (and closely related topics)
which were published between year 2000 to 2019. We followed widely used electronic databases by
following pre-defined review protocol. Articles were searched using keywords, forward reference searching
and backward reference searching in order to search all the articles related to the topic. After carefully
following study selection process 176 articles were selected for this SLR. This review article serves the
purpose of presenting state of the art results and techniques on OCR and also provide research directions by
highlighting research gaps.

INDEX TERMS Optical character recognition, classification, languages, feature extraction, deep learning.

I. INTRODUCTION input data. The offline system is a static system in which

PTICAL character recognition (OCR) is a system that con- input data is in the form of scanned images while in online
verts input text into machine-encoded format [1]. Today, OCR systems nature of input is more dynamic and is based on
is helping not only in digitizing the handwritten medieval the movement of pen tip having certain velocity, projection
manuscripts [2], but also helps in converting the typewritten angle, position and locus point. Therefore, an online system
documents into digital form [3]. This has made the retrieval is considered more complex and advance, as it resolves the
of the required information easier as one doesn’t have to overlapping problem of input data that is present in the offline
go through the piles of documents and files to search the system.
required information. Organizations are satisfying the needs One of the earliest OCR systems was developed in the
of digital preservation of historic data [4], law documents [5], 1940s, with the advancement in the technology over the time,
educational persistence [6] etc. the system became more robust to deal with both printed,
An OCR system depends mainly, on the extraction of and handwritten characters and this led to the commercial
features and discrimination/classification of these features availability of the OCR machines. In 1965, advance reading
(based on patterns). Handwritten OCR have received increas- machine ‘‘IBM 1287’’ was introduced at the ‘‘world fair’’ in
ing attention as a subfield of OCR. It is further categorized New York [10]. This was the first-ever optical reader, which
into offline system [7], [8] and online system [9] based on was capable of reading handwritten numbers. During the
The associate editor coordinating the review of this manuscript and
1970s, researchers focused on the improvement of response
approving it for publication was Jenny Mahoney. time and performance of the OCR system.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
142642 VOLUME 8, 2020
J. Memon et al.: Handwritten OCR: A Comprehensive SLR

The next two decades from 1980 till 2000, the software 1) To summarize existing research work (machine learn-
system of OCR was developed and deployed in educational ing techniques and databases) on different languages of
institutes, census OCR [11] and for recognition of stamped handwritten character recognition systems.
characters on metallic bar [12]. In the early 2000s, binariza- 2) To highlight research weakness in order to eliminate
tion techniques were introduced to preserve historical doc- them through additional research.
uments in digital form and provide researchers with access 3) To identify new research areas within the domain of
to these documents [13]–[16]. Some of the challenges of OCR.
binarization of historical documents were the use of non- We will follow the strategies proposed by
standard fonts, printing noise and spacing. In mid of 2000, Kitchenham et al. [22]. Following the proposed strategy,
multiple applications were introduced that were helpful for in subsequent sub- sections review protocol, inclusion and
differently-abled people. These applications helped these exclusion criteria, search strategy process, selection process
people in developing reading and writing skills. and data extraction and synthesis processes are discussed.
In the current decade, researchers have worked on different
machine learning approaches which include Support Vector A. REVIEW PROTOCOL
Machine (SVM), Random Forests (RF), k Nearest Neigh-
Following the philosophy, principles and measures of the
bor (kNN), Decision Tree (DT) [17]–[19], Neural Networks
Systematic Literature Review (SLR) [22], this systematic
etc. Researchers combined these machine learning tech-
study was initialized with the development of comprehensive
niques with image processing techniques to increase the accu-
review protocol. This protocol identifies review background,
racy of the optical character recognition system. Recently
search strategy, data extraction, research questions and qual-
researchers have focused on developing techniques for the
ity assessment criteria for the selection of study and data
digitization of handwritten documents, primarily based on
analysis.
deep learning [20] approach. This paradigm shift has been
The review protocol is what that creates a distinction
sparked due to adaption of cluster computing and GPUs and
between an SLR and traditional literature review or narrative
better performance by deep learning architectures [21], which
review [22]. It also enhances the consistency of the review
includes Recurrent Neural Networks (RNN), Convolutional
and reduces the researchers’ biases. This is due to the fact that
Neural Network (CNN), Long Short-Term Memory (LSTM)
researchers have to present a search strategy and the criteria
networks etc.
for the inclusion of exclusion of any study in the review.
This Systematic Literature Review (SLR) serves not only
the purpose of presenting literature in the domain of OCR for
different languages but also highlight research directions for B. INCLUSION AND EXCLUSION CRITERIA
a new researcher by highlighting weak areas of current OCR Setting up an inclusion and exclusion criteria makes sure
systems that need further investigation. that only articles that are relevant to study are included. Our
This article is organized as follows. Section II discusses criteria include research studies from journals, conferences,
review methodology employed in this article. Review symposiums and workshops on the optical character recogni-
methodology section includes review protocol, inclusion and tion of English, Urdu, Arabic, Persian, Indian and Chinese
exclusion criteria, search strategy, selection process, qual- languages. In this SLR, we considered studies that were
ity assessment criteria and metadata synthesis of selected published from January 2000 to December 2019.
studies. Statistical data from selected studies are presented Our initial search based on the keywords only resulted
in Section III. Section IV presents research question and in 1150 research articles related to handwritten OCRs of
their motivation. Section V discusses different classifications different languages (refer Figure 1 for complete overview
methods which are used for handwritten OCR. This section of the selection process). After a thorough review of the
will also elaborate on structural and statistical models for articles, we excluded articles that were not clearly related
optical character recognition. Section VI presents different to a handwritten OCR, but appeared in the search, because
databases (for specific language) which are available for of keyword match. Additionally, articles were also excluded
research purpose. Section VII presents research overview based on duplicity, non-availability of full text and if the
of language-specific research in OCR, while Section VIII studies were not related to any of our research questions.
highlights research trends. Section IX summarizes findings
and also highlights gaps in research that need the attention of C. SEARCH STRATEGY
the research community. Search strategy comprises of automatic and manual search,
as shown in Figure 1. An automatic search helped in identi-
fying primary studies and to achieve a broader perspective.
II. REVIEW METHODS Therefore, we extended the review by the inclusion of addi-
As mentioned above, this Systematic Literature Review tional studies. As recommended by Kitchenham et al. [22],
(SLR) aims to identify and present literature on OCR by for- the manual search strategy was applied to the references of the
mulating research questions and selecting relevant research studies that are identified after the application of automatic
studies. Thus, in summary, this review aims: search.

VOLUME 8, 2020 142643

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

FIGURE 1. Complete overview of studies selection process.

For automatic search, we used standard databases which D. STUDY SELECTION PROCESS
contain the most relevant research articles. These databases A tollgate approach was adopted for the selection of
include IEEE Explore, ISI Web of Knowledge, Scopus— study [23]. Therefore, after searching keywords in all rele-
Elsevier and Springer. While there is plenty of literature avail- vant databases, we extracted 1150 research studies through
able in the magazine, working papers, newspapers, books automatic search. Majority of these 1150 studies, 625 were
and blogs, we did not choose them for this review article duplicate studies and were eliminated. Inclusion and exclu-
as concepts discussed in these sources are not subjected to sion criteria based upon title, abstracts, keywords and the
review process; thus their quality cannot be reliably verified. type of publication was applied to the remaining 525 studies.
General keywords derived from our research questions and This resulted in the exclusion of 268 studies and leaving
the title of the study were used to search for research articles. 257 studies. In the next stage, the selection criteria were
Our aim was to identify as many relevant articles as possible applied, thus further 102 studies were excluded, and we were
from the main set of keywords. All possible permutations left with 155 studies.
of Optical character recognition concepts were tried in the Once we finished the automatic search stage, we started a
search, such as ‘‘optical character recognition’’, ‘‘pattern manual search procedure to guarantee the exhaustiveness of
recognition and OCR’’, ‘‘pattern matching and OCR’’ etc. the search results. We performed screening of the remaining
Once the primary data were obtained by using search 155 studies and went through the references to check relevant
strings, the data analysis phase of the obtained research research articles that could have been left search during the
papers began with the intention of considering their relevance automatic search. Manual search added 46 further studies.
to research questions and inclusion and exclusion criteria of After adding these studies, pre-final list of 201 primary stud-
the study. After that, a bibliography management tool, i.e. ies was obtained.
Mendeley, was used for storing all related research articles Next and final stage was to apply the quality assessment
to be used for referencing purpose. Mendeley also helped in criteria (QAC) on the pre-final list of 201 studies. Quality
identifying duplicate studies. assessment criteria were applied at the end as this is the
A manual search was performed with an automatic search final step through which a final list of studies for SLR was
to make sure that we had not missed anything. This deduced. QAC usually identifies studies whose quality is
was achieved through forward and backwards referencing. not helpful in answering the research question. After apply-
Furthermore, for data extraction, all the results were imported ing QAC, 25 studies were excluded, and we were left with
into a spreadsheet. Snowballing, which is an iterative pro- 176 primary studies. Refer Figure 1 for complete step-by-step
cess in which references of references are verified to iden- overview of selection process.
tify more relevant literature, was applied to primary studies Table 1 shows the distribution of the primary / selected
in order to extract more relevant primary studies. Set of studies among various publication sources, before and after
primary studies post snowball process was then added to applying above mentioned selection process. The same is also
Mendeley. shown in Figure 2.

142644 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

TABLE 1. Distribution of databases of selected studies before and after 176 studies were finally selected for this review article (refer
applying selection process.
to Figure 1 for complete overview of the selection process).

F. DATA EXTRACTION AND SYNTHESIS

During this phase, metadata of selected studies (176) was
extracted. As stated earlier, we used Mendeley and MS Excel
to manage the metadata of these studies. The main objective
of this phase was to record the information that was obtained
from the initial studies [22]. The data containing study ID
(to identify each study), study title, authors, publication year,
publishing platform (conference proceedings, journals, etc.),
citation count, and the study context (techniques used in
the study) were extracted and recorded in an excel sheet.
This data was extracted after a thorough analysis of each
study to identify the algorithms and techniques proposed by
the researchers. This also helped us to classify the studies
according to the languages on which the techniques were
applied. Table 2 shows the fields of the data extracted from
research studies.

TABLE 2. Extracted meta-data fields of selected studies.

FIGURE 2. Distribution of sources / databases of selected studies after

applying selection process.

E. QUALITY ASSESSMENT CRITERIA

Quality Assessment Criteria (QAC) is based on the principle
to make a decision related to the overall quality of the selected
set of studies [22]. Following criteria were used to assess the
quality of selected studies. This criterion helped us to identify III. STATISTICAL RESULTS FROM SELECTED STUDIES
the strength of inferences and helped us in selecting the most In this section, statistical results of the selected studies are
relevant research studies for our research. presented with respect to their publication sources, citation
Quality Assessment criteria questions: count status, temporal view, type of languages and type of
1) Are topics presented in a research paper relevant to the research methodologies.
objectives of this review article?
2) Does research study describes the context of the A. PUBLICATION SOURCES OVERVIEW
research? In this review, most of the included studies are published
3) Does research article explains the approach and in reputed journals and leading conferences. Therefore, con-
methodology of research with clarity? sidering the quality of research studies, we believe that this
4) Is data collection procedure explained If data collection systematic review can be used as a reference to find latest
is done in the study? trends and to highlight research directions for further studies
5) Is the process of data analysis explained with proper in the domain of handwritten OCR. Figure 3 shows the distri-
examples? bution of studies derived from different publication sources.
We evaluated 201 selected studies by using the abovemen- Majority of included studies (107) were published in research
tioned quality assessment questions in order to determine the journals (61%), followed by (61) publications in conference
credibility of a particular acknowledged study. These five articles (34%). Whereas, few (5) articles were published in
QA schema is inspired by [23]. The quality of the study was workshop proceedings and only (3) relevant articles were
measured depending upon the score of each QA question. found to be presented in symposiums.
Each question was assigned 2 marks, and the study’s quality
was considered to be selected if it scored greater than or equal B. RESEARCH CITATIONS
to 5 at the scale of 10. Thus, studies below the score of 5 Citation count was obtained from Google Scholar. Overall,
were not included in the research. Following this criterion, selected studies have good citation count, which shows that

VOLUME 8, 2020 142645

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

the reference figure, it can be noticed that there is a variation

in the publication count through these years. Statistics show
sudden increase in a number of publications in the domain of
handwritten character recognition in the years 2002, 2007 and
2009. The number of publications remained steady in the
remaining years of the 2000s. After 2010 there is again a
steady increase in the number publications, i.e. 59 publica-
tions in 8 years from 2010-17. During the last two years,
we have seen a steep rise in the number of the publication.
We found 55 new studies in the last two years as compared
59 studies in the previous 8 years. This is conceivably not
surprising since the concept of handwritten character recog-
nition is catching the interest of more researcher because of
the advancement of the research work in the fields of deep
learning and computer vision. We believe that application
FIGURE 3. Study distribution per publication source.
areas of handwritten OCRs will further increase in the coming
years. This is to be noted that these number of studies only
include research articles which are related to our research
the quality of selected studies is worthy of being added in questions.
the review and also implies that researchers are actively
working in this area of research. As presented in Figure 4,
approximately 95% of the selected studies have at least one D. LANGUAGE SPECIFIC RESEARCH
citation, except the research articles, which are published The distributions/number of selected studies with respect
recently in 2019. Among selected studies, 33 studies have to investigated scripting languages are shown in Figure 6.
more than 100 citations, 15 studies have been cited between A total number of selected studies are 176, and out of these
61-100 times, 25 studies were cited between 33-60 times, 172 studies, the English language has the highest contribution
16 studies were cited between 16-30 times and 68 studies of 53 studies in the domain of handwritten character recog-
were cited between 1 and 15 times. Overall, we predict nition, 44 studies related to the Arabic language, 37 studies
that selected studies citations will increase further because are on the Indian scripts, 23 on the Chinese language, 118 on
research articles are constantly being published in this the Urdu language, while 14 studies were conducted on the
domain. Persian language. Some of the selected articles discussed
Table 3 provides details of research publications with more multiple languages.
than 100 citations each. These articles can be considered to Figure 7 represents publications count each year with
have a strong impact on the researchers working to build respect to language. Reference figure shows compiled
robust OCR system. temporal view of handwritten OCR researches done in
different languages throughout the mentioned era of
C. TEMPORAL VIEW 2000-2019, in this time period there are certain research
The distribution of number of studies over the period under articles that cover more than one language of handwritten
study (2000 - 2019) can be seen in Figure 5. According to OCR.

FIGURE 4. Citation count of selected studies. Numeric value within bar shows number of studies that have been cited
x times (corresponding values on the x-axis).

142646 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

TABLE 3. Research publications with more than 100 citations.

VOLUME 8, 2020 142647

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

V. CLASSIFICATION METHODS OF HANDWRITTEN OCR

In handwritten OCR an algorithm is trained on a known
dataset, and it discovers how to accurately categorize/classify
the alphabets and digits. Classification is a process to learn a
model on a given input data and map or label it to predefined
category or classes [17]. In this section, we have discussed
the most prevalent classification techniques in OCR research
studies beginning from 2000 till 2019.

A. ARTIFICIAL NEURAL NETWORKS (ANN)

Biological neuron inspired architecture, Artificial Neural
FIGURE 5. Publications over the years. On the y -axis is the number of
publications.
Networks (ANN) consists of numerous processing units
called neurons [56]. These processing elements (neurons)
work together to model given input data and map it to pre-
defined class or label [57]. The main unit in neural networks
is nodes (neuron). Weights associated with each node are
adjusted to reduce the squared error on training samples
in a supervised learning environment (training on labelled
samples/data). Figure 8 presents a pictorial representation of
Multi-Layer Perceptron (MLP) that consists of three layers,
i.e. (input, hidden and output).
Feedforward networks / Multi-Layer Perceptron (MLP)
achieved renewed interest of research community in the mid
1980s as by that time ‘‘Hopfield network’’ provided the way
to understand human memory and calculate the state of a
neuron [59]. Initially, the computational complexity of find-
FIGURE 6. Number of selected studies with respect to investigated ing weights associated with neurons hindered the application
language. Numeric value within bar shows number of selected studies for
the given language. of neural networks. With the advent of deep (many layers)
neural architectures, i.e. Recurrent Neural Network (RNN)
and Convolutional Neural Networks (CNN), neural networks
IV. RESEARCH QUESTIONS have established itself as one of the best classification tech-
Research questions play an important role in a systematic lit- nique for recognition tasks including OCR [60]–[63]. Refer
erature review because these questions determine the search Sections VIII and IX-B for current and future research trends.
queries and keywords that will be used to explore research The early implementation of MLP for handwritten OCR
publications. As discussed above, we chose research ques- was done by Shamsher et al. [64] on the Urdu language. The
tions which not only help seasoned researchers but also to researchers proposed feed-forward neural network algorithm
researchers entering in the domain of optical character recog- of MLP (Multi-Layer Perceptrons) [65]. Liu and Suen [66]
nition to understand where the research in this field stands as used MLP on Farsi and Bangla numerals. One hidden layer
of today. This review article answers research questions pre- was used with the connecting weights estimated by the error
sented in Table 4. Reference table also presents the motivation backpropagation (BP) algorithm that minimized the squared
for each research question. error criterion. On the other hand, Cirecsan et al. [30] trained
five MLPs with two to nine hidden layers and varying num-
bers of hidden units for the recognition of English numerals.
TABLE 4. Research questions and motivation. Recently, Convolutional Neural Network (CNN) has
reported great success in character recognition task [67].
A convolutional neural network has been widely used
for classification and recognition of almost all the lan-
guages that have been reviewed for this systematic literature
review [68]–[74].

B. KERNEL METHODS
A number of powerful kernel-based learning models,
e.g. Support Vector Machines (SVMs), Kernel Fisher
Discriminant Analysis (KFDA) and Kernel Principal Com-
ponent Analysis (KPCA) have shown practical relevance for

142648 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

FIGURE 7. Selected studies count each year with respect to specific language. y -axis shows the number of selected studies.
Specific color within each bar represents specific language as shown in the legend.

digit recognition, image classification, face detection, object

detection, and text classification [75]. Kernel Fisher Dis-
criminant Analysis (KFDA) and Kernel Principal Component
Analysis (KPCA) are also some of the most significant kernel
methods being used in offline handwritten character recogni-
tion system [76]. A number of researchers still believe that
SVM performs better than most of the other techniques in
classifying the handwritten characters. This is the reason why
SVM is still being used for the purpose of classification of
characters in HCR [77]–[80].
Previously, Boukharouba and Bennia [75] and
Yang et al. [81] used SVM for recognition of Urdu and
Arabic handwritten digits. SVMs have also been success-
FIGURE 8. An architecture of Multilayer Perceptron (MLP) [58]. fully implement in image classification and affect recogni-
tion [82], [83], text classification [84] and face and object
detection [85], [86].
classification problems. For instance, in the context of optical
pat- tern, text categorization, time-series prediction, these
models have significant relevance. C. STATISTICAL METHODS
In support vector machine, kernel performs mapping of Statistical classifiers can be parametric and non-parametric.
feature vectors into a higher dimensional feature space in Parametric classifiers have fixed (finite) number of param-
order to find a hyperplane, which is linearly separates classes eters, and their complexity is not a function of the size
by as much margin as possible. Given a training set of labeled of input data. Parametric classifiers are generally fast in
examples {(xi , yi ), i = 1 . . . l } where xi ∈ <n and yi ∈ learning concept and can even work with the small training
{−1, 1}, a new test example x is classified by the following set. Example of parametric classifiers is Logistic Regression
function: (LR), Linear Discriminant Analysis (LDA), Hidden Markov
1
Model (HMM) etc.
X On the other hand, non-parametric classifiers are more
f (x) = sgn( αi yi K (xi , x) + b) (1)
flexible in learning concepts but usually grow in complexity
i=1
with the size of input data. K Nearest Neighbor (K NN),
where: Decision Trees (DT) are examples of non-parametric tech-
1) K (.,.) is a kernel function niques as their number of parameters grows with the size of
2) b is the threshold parameter of the hyperplane the training set.
3) αi are Lagrange multipliers of a dual optimization prob-
lem that describe the separating hyperplane 1) NON-PARAMETRIC STATISTICAL METHODS
Before the popularization of deep learning methodology, One of the most used and easy to train statistical model for
SVM was one of the most robust technique for handwritten classification is k nearest neighbor (kNN) [42], [87], [88].

VOLUME 8, 2020 142649

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

It is a non-parametric statistical method, which is widely used metrics are used, most common ones are Euclidean distance,
in optical character recognition. Non-parametric recognition city block distance, cross-correlation, normalized correlation
does not involve a-priori information about the data. etc.
kNN finds a number of training samples closest to a new In template matching, either template matching technique
example-based on target function. Based on the value of the employs a rigid shape matching algorithm or deformable
targeted function, it infers the value of the output class. The shape matching algorithm. Thus, creating a different family
probability of an unknown sample q belonging to class y can of template matching. Taxonomy of template matching tech-
be calculated as follows: niques is presented in Figure 9.
k∈K Wk .1(ky =y)
P
p(y|q) = P (2)
k∈K Wk
1
Wk = (3)
d(k, q)
where;
1) K is the set of nearest neighbors
2) ky the class of k
3) d(k, q) the Euclidean distance of k from q, respectively.
Researchers have been found to use kNN for over a decade
FIGURE 9. An overview of template matching techniques.
now, and they believe that this algorithm achieves relatively
good performance for character recognition in their experi-
ments performed on different datasets [2], [18], [62], [88]. One of the most applicable approaches for character recog-
kNN classifies object / ROI based on the majority vote of nition is deformable template matching (refer Figure 10) as
its neighbours (class) as it assigns class most prevalent among different writers can write character by deforming them in a
its k nearest neighbours. If k = 1, then the object is simply particular way specific to writer. In this approach, a deformed
assigned to a class of that single nearest neighbour [57]. image is used to compare it with a database of known images.
Thus, matching/classification is performed with deformed
2) PARAMETRIC STATISTICAL METHODS shapes as a specific writer could have deformed charac-
ter in a particular way [36]. Deformable template matching
As mentioned above, parametric techniques models concepts
is further divided into parametric and free form matching.
using fixed (finite) number of parameters as they assume
Prototype matching, which is sub-class of parametric
sample population/training data can be modelled by a proba-
deformable matching, matching of done based on a stored
bility distribution that has a fixed set of parameters. In OCR
prototype (deformed) [94].
research studies, generally, characters are classified accord-
ing to some decision rules such as a maximum likelihood or
Bayes method once parameters of the model are learned [36].
Hidden Markov Model (HMM) was one of the most fre-
quently used parametric statistical method earlier in 2000.
HMM, models system/data that is assumed to be a Markov
process with hidden states, wherein Markov process prob-
ability of one states only depends on previous state [36].
It was first used in speech recognition during the 1990s
before researchers started using it in recognition of optical
characters [89]–[91]. It is believed that HMM provides better
results even when the availability of lexicons is limited [41].

D. TEMPLATE MATCHING TECHNIQUES

FIGURE 10. (a) Digit deformations (b) Deformed template superimposed
As the names suggest, template matching is an approach in on target image [36].
which images (a small part of an image) is matched with a
certain predefined template. Usually, template matching tech- Apart from the deformable template matching approach,
niques employ a sliding window approach in which template second sub-class of template matching is rigid template
image or feature are sliders on the image to determine the sim- matching. As the name suggests, rigid template matching
ilarity between the two. Based on used similarity (or distance) does not take into account shape deformations. This approach
metric classification of different objects are obtained [92]. usually works with features extraction/matching of the image
In OCR, template matching technique is used to clas- with a template. One of the most common approaches used
sify character after matching it with the predefined tem- in OCR to extract shape features is Hough transform, like
plate(s) [93]. In literature, different distance (similarity) Arabic [95] and Chinese [96].

142650 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

FIGURE 11. (a) Primitive and relations (b) Directed graph for capital letter R and E [100].

Second sub-class of rigid template matching is correlation- As mentioned above, writing structural components are
based matching. In this technique, initially, image similarity extracted using pattern primitives, i.e. edge, contours, con-
is calculated and based on similarity features from specific nected component geometry etc. The relation between these
regions are extracted and compared [36], [97]. structures can be defined mathematically using graphs (refer
Figure 11 for an example showing how letter ‘‘R’’ and ‘‘E’’
E. STRUCTURAL PATTERN RECOGNITION can be modelled using graph theory). Then considering spe-
Another classification technique that was used by OCR cific graph architecture different structures can be classi-
research community before the popularization of kernel fied using graph similarity measure i.e. similarity flooding
methods and neural networks / deep learning approach was algorithm [103], SimRank algorithm [104], Graph similarity
structural pattern recognition. Structural pattern recognition scoring [105] and vertex similarity method [106]. In one
aims to classify objects based on a relationship between its study [107], graph distance is used to segment overlapping
pattern structures and usually structures are extracted using and joined characters as well.
pattern primitives (refer Figure 11 for an example of pattern
primitives), i.e. edge, contours, connected component geom- 2) GRAMMAR BASED METHODS
etry etc. One of such image primitive that has been used In graph theory, syntactic analysis is also used to find similar-
in OCR is Chain Code Histogram (CCH) [98], [99]. CCH ities in structural graph primitives using the concept of gram-
effectively describes image / character boundary / curve, thus mar [108]. The benefit of using grammar concepts in finding
helping in classify character [57], [75]. Prerequisite condition the similarity in graphs comes from the fact that this area
to apply CCH for OCR is that image should be in binary is well researched and techniques are well developed. There
format, and boundaries should be well defined. Generally, for are different types of grammar based on restriction rules,
handwritten character recognition, this condition makes CCH for example, unrestricted grammar, context-free grammar,
difficult to use. Thus, different research studies and publicly context-sensitive grammar and regular grammar. Explanation
available datasets use/provide binarized images [87]. of these grammar and corresponding applied restrictions are
In research studies of OCR, structural models can be out the scope of this survey article.
further subdivided on the basis of the context of structure, In OCR literature, usually, strings and trees are used
i.e. graphical methods and grammar-based methods. Both of to represent models based on grammar. With well-defined
these models are presented in the next two sub-sections. grammar, a string is produced that then can be robustly
clas- sified to recognize the character. The tree structure can
1) GRAPHICAL METHODS also model hierarchical relations between structural primi-
A graph (G) is a way to mathematically describe a relation tives [92]. Trees can also be classified by analyzing grammar
between connected objects and is represented by an ordered that defines the tree, thus classifying specific character [109].
pair of nodes (N ) and edges (E). Generally, for OCR, E rep-
resents the arc of writing stroke connecting N . The particular VI. DATASETS
arrangement of N and E define characters / digits / alphabets. Generally, for evaluating and benchmarking different OCR
Trees (undirected graph, where the direction of the connec- algorithms, standardized databases are needed/used to enable
tion is not defined), directed graphs (where the direction of a meaningful comparison [55]. Availability of a dataset
edge to a node is well defined) are used in different research containing enough amount of data for training and testing
studies to represent characters mathematically [101], [102]. purpose is always a fundamental requirement for a quality

VOLUME 8, 2020 142651

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

research [110], [111]. Research in the domain of optical

character recognition mainly revolves around six different
languages, namely, English, Arabic, Indian, Chinese, Urdu
and Persian / Farsi script. Thus, there are publicly available
datasets for these languages such as MNIST, CEDAR, CEN-
PARMI, PE92, UCOM, HCL2000 etc.
Following subsections presents an overview of most used
datasets for the above mentioned languages.

A. CEDAR
This legacy dataset, CEDAR, was developed by the
researchers at the University of Buffalo in 2002 and is con-
sidered among the first few large databases of handwritten
characters [40]. In CEDAR, the images were scanned at 300
dpi. Example character images from the CEDAR database are
shown in Figure 12.

FIGURE 13. Sample image from CHARS74K dataset [112].

image detection, while [113] used the dataset for recognizing

characters in early Indian printed documents. Joe et al. [114]
used CNN to recognize offline handwritten characters written
in Kannada script.
It is to be noted that Kannada is one of many Indian
scripts we have included in this research. There are various
datasets for Indian language, depending on the script that has
been used. For example, CMATERDB is a dataset for Indian
script called Bangla [115], [116] and Kaggle’s Tamil hand-
written character dataset is another such dataset for Tamil
script [117].

C. MNIST
The MNIST dataset is considered as one of the most used/
cited dataset for handwritten digits [30], [42], [118]–[121].
FIGURE 12. Sample image from CEDAR dataset [42]. It is the subset of the NIST dataset, and that is why it
is called modified NIST or MNIST. The dataset consists
of 60,000 training and 10,000 test images. Samples are nor-
B. CHARS74K malized into 20 × 20 grayscale images with reserved aspect
Chars74k [112] dataset was introduced by researchers at the ratio, and the normalized images are of size 28 × 28. The
university of surrey in 2009. The dataset contains 74,000 dataset greatly reduces the time required for pre-processing
images of English and Kannada (Indian) scripts. The database and formatting, because it is already in a normalized form.
contains street scenes taken in Bangalore, India. One thou-
sand nine hundred twenty-two images of signboards, hoard- D. UCOM
ings, advertisements and products in supermarkets were The UCOM is an Urdu language dataset available for
photographed. Segmentation of individual characters was research [122]. The authors claim that this dataset could
done manually, and results were presented in bounding box be used for both character recognition as well as writer
segmentation. Bag of visual words technique was used for identification. The dataset consists of 53,248 characters and
object categorization, and eventually, 62 different classes 62,000 words written in nasta’liq (calligraphy) style, scanned
were created for English and 657 classes for Kannada. at 300 dpi. The dataset was created based on the writing
A number of researchers have used CHARS74k dataset for of 100 different writers where each writer wrote 6 pages
recognition of Kannada script. Naiemi [78] applied histogram of A4 size. The dataset evaluation is based on 50 text line
of oriented gradients features on CHARS74k dataset for spam images as train dataset and 20 text line images as test dataset

142652 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

FIGURE 16. Sample writings from IFN/ENIT dataset [37].

11,000 training, 2,000 verification and 5,000 samples for

testing purpose.
Another similar, but larger dataset of Farsi numerals was
produced by Khosravi and Kabir [51] in 2007. This dataset
contains 102,352 digits extracted from registration forms of
high school and undergraduate students. Later in 2009 [126],
CENPARMI released another larger, extended version of
Farsi dataset. This larger dataset contains 432,357 images of
dates, words, isolated letters, isolated digits, numeral strings,
special symbols, and documents. Refer Figure 17 for exam-
FIGURE 14. Sample handwritten digits from MNIST dataset [42]. ples images from CENPARMI Farsi language dataset.

with reported error rate between 0.004 -0.006%. Example

characters from the dataset are presented in Figure 15.

FIGURE 15. Example hand written characters from UCOM dataset [122].

FIGURE 17. CENPARMI dataset example images [51].

E. IFN/ENIT
The IFN/ENIT [37] is the most popular Arabic database
of handwritten text. It was developed in 2002 by the G. HCL2000
researchers at Technical University Braunschweig, Germany The HCL2000 is a handwritten Chinese character database,
for the advancement of research and development of refer Figure 18 to see sample images. The dataset is publicly
Arabic handwriting recognition systems. The dataset contains available for researchers. The dataset contains 3,755 fre-
26459 handwritten images of the names of towns and vil- quently used Chinese characters written by 1,000 different
lages in Tunisia. These images consist of 212,211 characters subjects. The database is unique in a way that it contains two
written by 411 different writers, refer Figure 16. Since its sub-datasets, one is handwritten Chinese characters dataset,
inception, the dataset has been widely used by the researchers while the other is corresponding writer’s information dataset.
for the efficient recognition of Arabic characters [41], [48], This information is provided so that research can be con-
[123], [124]. ducted not only based on character recognition, but also on
the writer’s background such as age, gender, occupation and
F. CENPARMI education [127].
The CENter for Pattern Recognition and Machine Intelli-
gence (CENPARMI) introduced the first version of Farsi H. IAM
dataset in 2006 [51], [125]. This dataset contains 18,000 The IAM [128] is a handwritten database of English
samples of Farsi numerals. These numerals are divided into language based on Lancaster-Oslo/Bergen (LOB) corpus.

VOLUME 8, 2020 142653

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

FIGURE 18. HCL2000 dataset sample images [127].

Data were collected from 400 different writers who pro-

duced 1,066 forms of English text containing a vocabulary FIGURE 20. Data from UNESCO’s report on ‘‘world’s languages in
danger’’ [129].
of 82,227 words. Data consists of full English language
sentences. The dataset was also used for writer identifica-
tion [48]. Researchers were able to successfully identify This section presents state-of-the-art results for six lan-
writer 98% of the time during experiments on IAM dataset. guages which are usually studied by researchers.
Writing sample from the IAM dataset are presented in
Figure 19. A. ENGLISH LANGUAGE
The English Language is the most widely used language in
the world. It is the official language of 53 countries and
articulated as a first language by around 400 million people.
Bilinguals use English as an international language. Charac-
ter recognition for the English language has been extensively
studied throughout many years. In this systematic literature
review, the English language has the highest number of
publications, i.e. 45 publications after concluding the study
selection process (refer Section II-D and Section III-D). The
OCR systems for the English language occupy a significant
place as a large number of studies have been done in the era
of 2000-2018 on the English language.
The English language OCR systems have been used suc-
cessfully in a wide array of commercial applications. The
most cited study for English language handwritten OCR is by
Plamondon and Srihari [35] in 2000, which have more than
2900 citations, refer Table 3. The objective of the research by
FIGURE 19. Sample Image IAM dataset [128].
Plamondon et al. was to present a broad review of state of the
art in the field of automatic processing of handwriting. This
VII. LANGUAGES paper explained the phenomenon of pen-based computers and
As mentioned above, researchers working in the domain of achieved the goal of automatic processing of electronic ink by
optical character recognition have mainly investigated six dif- mimicking and extending the pen-paper metaphor. To identify
ferent languages, which are English, Arabic, Indian, Chinese, the shape of the character, structural and rule-based models
Urdu and Persian. This is one of the future work to built OCR like (SOFM) self-organized feature map, (TDNN) time-delay
systems for other languages as well. neural network and (HMM) hidden Markov model was used.
According to the United Nations Educational, Scientific Another comprehensive overview of character recognition
and Cultural Organization (UNESCO) report on ‘‘world’s presented in [36] by Arica et al. has more than 500 citations.
languages in danger,’’ at least 43% of languages spoken in the Arica et al. concluded that characters are natural entities,
world are endangered [129]. These large number of languages and it is practically impossible for character recognition to
need the attention of OCR research community as well to pre- impose a strict mathematical rule on the patterns of char-
serve this heritage from extinction or at least to build such a acters. Neither the structural nor the statistical models can
system that translates documents from endangered languages signify a complex pattern alone. The statistical and structural
to electronic form for reference. Data from UNESCO’s report information for many characters pattern can be combined by
on ‘‘world’s languages in danger’’ is presented in Figure 20. neural networks (NNs) or harmonic markov models (HMM).

142654 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

Connell and Jain [9] demonstrated a template-based sys- 32 Farsi alphabets into 8 different classes. A multilayer per-
tem for online character recognition, which is capable of ceptron (MLP) (refer Figure 8 for an overview of MLP) was
representing different handwriting styles of a particular char- used as a classifier for this purpose. The classification rate for
acter. They used decision trees for efficient classification of characters and digits were 87.26% and 91.37% respectively.
charac- ters and achieved 86% accuracy. However, in another research [141], researchers achieved
Every language has specific way of writing and have some a recognition rate of 99.5% by using RBF kernel-based sup-
diverse features that distinguished it with other language. port vector machine. Broumandnia and Shanbehzadeh [142]
We believe that to efficiently recognize handwritten and conducted research on Farsi character recognition and claims
machine printed text of the English language, researchers to propose the fastest approach of recognizing Farsi character
have used almost all of the available feature extraction and using Fast Zernike wavelet moments and artificial neural net-
classification techniques. These feature extraction and clas- works (ANN). This model improves on average recognition
sification techniques include but not limited to HOG [130], speed by 8 times.
bidirectional LSTM [131], directional features [132], multi- Liu and Suen [66] presented results of handwritten Bangla
layer perceptron (MLP) [119], [133], [134], hidden markov and Farsi numeral recognition on binary and grayscale
model(HMM) [26], [52], [54], [62], Artificial neural net- images. The researchers applied various character recogni-
work (ANN) [135]–[137] and support vector machine tion methods and classifiers on the three public datasets such
(SVM) [29], [67]. as ISI Bangla numerals, CENPARMI Farsi numerals, and
Recently trend is shifting away from using hand- IFHCDB Farsi numerals and claimed to have achieved the
crafted features and moving towards deep neural networks. highest accuracies on the three datasets, i.e. 99.40%, 99.16%,
Convolutional Neural Network (CNN) architecture, a class and 99.73%, respectively.
of deep neural networks, has achieved classification results In another research, Boukharouba and Bennia [75] pro-
that exceed state-of-the-art results specifically for visual posed SVM based system for efficient recognition of hand-
stimuli/input [138]. LeCun [20] proposed CNN architecture written digits. Two feature extraction techniques, namely,
based on multiple stages where each stage is further based on chain code histogram (CCH) [143] and white-black transition
multiple layers. Each stage uses feature maps, which are basi- information, were discussed. The feature extraction algorithm
cally arrays containing pixels. These pixels are fed as input to used in the research did not require digits to be normalized.
multiple hidden layers for feature extraction and a connected SVM classifier, along with RBF kernel method, was used for
layer, which detects and classifies object [55]. A recent study classification of handwritten Farsi digits named ‘hoda’. This
by [69] used fully convolutional neural network(FCNN) on system maintains high performance with less computational
IAM and RIMES datasets. Results were promising, and complexity as compared to previous systems as the features
researchers achieved the character error rate(CER) and word used were computationally simple.
error rate(WER) of 4.7%, 8.22%, 2.46%, 5.68% respectively. Researchers have also used Convolutional Neural Network
Jayasundara [139] proposed a novel technique called capsule (CNN) in conjunction with other techniques for the recog-
networks(CapsNet) for the handwritten character recogni- nition of characters. These techniques have been applied on
tion with very small datasets. Research claims that these different datasets to check the accuracy of techniques [74],
techniques require a very small number of training samples [87], [144]–[146].
for each class. These samples can be as low as 200. It is
also claimed that the proposed technique can produce results C. URDU LANGUAGE
similar to state-of-the-art systems, with only 10% of the data. Urdu is curvasive language like Arabic, Farsi and many
When the proposed technique was applied to small datasets, other [147]. In the Urdu language, a notable early attempt to
it achieved the accuracy of 90.46%. improve the methods for OCR is by Javed et al. in 2009 [148].
Their study focuses on the Nasta’liq (calligraphy) style-
B. FARSI/PERSIAN SCRIPT specific pre-processing stage in order to overcome the chal-
Farsi, also known as the Persian Language, is mainly spoken lenges posed by the Nasta’liq style of Urdu handwriting.
in Iran and partly in Afghanistan, Iraq, Tajikistan and The steps proposed include page segmentation into lines and
Uzbekistan by approximately 120 million people. The further line segmentation into sub-ligatures, followed by base
Persian script is considered to be similar to Arabic, Urdu, identification and base-mark association. 94% of the ligatures
Pashto and Dari languages. Its nature is also cursive, so the were accurately separated with proper mark association.
appearance of the letter changes with respect to positions. Later in 2009, the first known dataset for Urdu hand-
The script comprises of 32 characters, and unlike the Arabic writing recognition was developed at Centre for Pattern
language, the writing direction of the Farsi language is mostly Recognition and Machine Intelligence (CENPARMI) [149].
but not exclusively from right to left. Sagheer et al. [149] focused on the methods involving data
Mozaffari et al. [140] proposed a novel handwritten char- collection, data extraction and pre-processing. The dataset
acter recognition method for isolated alphabets and digits of stores dates, isolated digits, numerical strings, isolated letters,
Farsi and Arabic language by using fractal codes. On the special symbols and 57 words. As an experiment, Support
basis of the similarities of the characters, they categorized the Vector Machine (SVM) using a Radial Base Function/kernel

VOLUME 8, 2020 142655

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

(RBF) was used for classification of isolated Urdu digits. The Researchers claim to have achieved very good accuracy in
experiment resulted in a high recognition rate of 98.61%. recognizing the characters. The error rate was 6.04–7.93%
To facilitate multilingual OCR, Hangarge and Dhandra during various experiments. During another study,
[118] proposed a texture-based method for handwritten script Rafeeq et al. [156] used a deep neural network with dropout
identification of three major scripts: English, Devnagari and regularization. Ligatures were categorized, and the K-Means
Urdu. Data from the documents were segmented into text algorithm is used to cluster the ligatures. Researchers claim
blocks and / or lines. In order to discriminate the scripts, that their proposed technique achieved 94.71% accuracy as
the proposed algorithm extracts fine textural primitives from compared to neural networks which achieved only 74.31%
the input image based on stroke density and pixel density. accuracy.
For experiments, k-nearest neighbour classifier was used for
classification of the handwritten scripts. The overall accuracy D. CHINESE LANGUAGE
for tri-script and bi-script classification peaked up to 88.6% Our research includes 23 research publications on the OCR
and 97.5% respectively. system of Chinese language after concluding the study selec-
A study by Pathan et al. [7] in 2012 proposed an approach tion process (refer Section II-D and Section III-D). One
based on the invariant moment technique to recognize the of the Earliest research on the Chinese language was done
handwritten isolated Urdu characters. A dataset comprising in 2000 by Fu et al. [157]. The researchers used self-growing
of 36800 isolated single and multi-component characters probabilistic decision-based neural networks (SPDNNs) to
was created. For multi-component letters, primary and sec- develop a user adaptation module for character recognition
ondary components were separated, and invariant moments and personal adaption. The resulting recognition accuracy
were calculated for each. The researchers used SVM for peaked up to 90.2% in ten adapting cycles.
classification, which resulted in an overall performance rate Later in 2005, a comparative study of applying feature
of 93.59%. Similarly, Raza et al. [150] created an offline sen- vector-based classification methods to character recognition
tence database with automatic line segmentation. It comprises by Liu and Fujisawa [67] found that discriminative classifiers
of 400 digitised forms by 200 different writers. such as an artificial neural network (ANN) and support vector
Obaidullah et al. [151] proposed a handwritten numeral machines (SVM) gave higher classification accuracies than
script identification (HNSI) framework to identify numeral statistical classifiers when the sample size was large. How-
text written in Bangla, Devanagari, Roman and Urdu. The ever, in the study SVM demonstrated better accuracies than
framework is based on a combination of daubechies wavelet neural networks in many experiments.
decomposition [152] and spatial domain features. A dataset In another study Bai and Huo [45] evaluated the use
of 4000 handwritten numeral word image for these scripts of 8-directional features to recognize online handwritten
was created for this purpose. In terms of average accuracy Chinese characters. Following a series of processing steps,
rate, multi-layer perceptron (MLP) (refer Figure 8 for a pic- blurred directional features were extracted at uniformly
torial depiction of MLP) proves to be better than NBTree, sampled locations using a derived filter, which forms a
PART, Random Forest, SMO and Simple Logistic classifiers. 512-dimensional vector of raw features. This, in comparison
In 2018, Asma and Kashif [153] presented a compara- to an earlier approach of using 4-directional features, resulted
tive analysis of raw images and meta-features from UCOM in a much better performance.
dataset. CNN (Convolutional Neural Network) and an LSTM In 2009, Zhang et al. [127] presented HCL2000, a
(Long short-term memory), which is a recurrent neural large-scale handwritten Chinese Character database. It stores
network-based architecture were used on Urdu language 3,755 frequently used characters along with the information
dataset. Researchers claim that CNN provided accuracy of its 1000 different writers. HCL2000 was evaluated using
of 97.63% and 94.82% on thickness graph and raw images, three different algorithms; Linear Discriminant Analysis
respectively. While the accuracy of LSTM was 98.53% and (LDA), Locality Preserving Projection (LPP) and Marginal
99.33%. Naseer and Zafar [153] and Tayyab et al. [154] Fisher Analysis (MFA). Prior to the analysis, the Nearest
proposed an OCR model based on CNN and BDLSTM Neighbor classifier assigns input image to a character group.
(Bi-Directional LSTM). This model was applied to a dataset The experimental results show MFA and LPP to be better than
containing Urdu news tickers, and results were compared LDA.
with google’s vision cloud OCR. Researchers found that their Yin et al. [53] proposed ICDAR 2013 competition which
proposed model worked better than google’s cloud vision received 27 systems for 5 tasks – classification on extracted
OCR in 2 of the 4 experiments. feature data, online/offline isolated character recognition and
In 2019 Ahmed et al. [155] proposed a technique based online/offline handwritten text recognition. Techniques used
on one-dimensional BLSTM classifier that used recurrent in the systems were inclusive of LDA, Modified quadratic
neural network(RNN), long-short term memory(LSTM) discriminant function (MFQD), Compound Mahalanobis
and bidirectional recurrent neural networks(BRNN) for the Function (CMF), convolutional neural network (CNN) and
recognition of handwritten Urdu written in Nasta’liq style. multilayer perceptron (MLP). It was explored that the meth-
Researchers also presented a new dataset of 500 writ- ods based on neural networks proved to be better for recog-
ers named Urdu-Nasta’liq handwritten dataset (UNHD). nizing both isolated character and handwritten text.

142656 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

During the study in 2016 on accurate recognition of the review included restrictive lexicons and restrictions on the
multilingual scene characters, Tian et al. [130] proposed text appearance.
an extension of Histogram of Oriented Gradient (HOG), In 2009, Graves and Schmidhuber [24] introduced a
Cooccurrence HOG (Co-HOG) and Convolutional Co-HOG globally trained offline handwriting recognizer based on
(ConvCo-HOG) features. The experimental results show the multi-directional recurrent neural networks and connectionist
efficiency of the approaches used and higher recognition temporal classification. It takes raw pixel data as input. The
accuracy of multilingual scene texts. system had an overall accuracy of 91.4%, which also won the
In 2018, researchers on Chinese script used neural international Arabic recognition competition.
networks to recognize CAPTCHA (Completely Automated Another notable attempt for Arabic OCR was made by
Public Turing test to tell Computers and Humans Apart) Lutf et al. [170] in 2014, which primarily focused on the
recognition [158], Medical document recognition [159], speciality of the Arabic writing system. The researcher pro-
License plate recognition [160] and text recognition in his- posed a novel method with minimum computation cost for
torical documents [161]. Researchers used Convolutional Arabic font recognition based on diacritics. Flood-fill based
Neural Network(CNN) [158], [161], Convolutional Recur- and clustering-based algorithms were developed for diacritics
rent Neural Network(CRNN) [159] and Single Deep Neural segmentation. Further, diacritic validation is done to avoid
Network(SDNN) [160] during these studies. misclassification with isolated letters. Compared to other
During 2019 [162], [163] used techniques based on recur- approaches, this method is the fastest with an average recog-
rent neural network(RNN) for the recognition of online and nition rate of 98.73% for 10 most popular Arabic fonts.
offline handwritten text, respectively. On the other hand, An Arabic handwriting synthesis system devised by
Gan et al. [73] used 1-dimensional CNN for the recogni- Elarian et al. [171] in 2015 synthesizes words from
tion of online handwritten Chinese characters. 1-dimensional segmented characters. It uses two concatenation models:
CNN seems to have performed better as recognition accuracy ExtendedGlyphs connection and the Synthetic-Extensions
of [73] is 98.1% as compared to [163] where the accuracy connection. The impact of the results from this system shows
of 83% was achieved. Zhu et al. [164] proposed a new neural significant improvement in the recognition performance of an
network structure for Chinese handwritten character recog- HMM-based Arabic text recognizer.
nition. Researchers adaptively assigned different weights Akram et l. [172] discussed an analytical approach to
to category-classifiers depending on the quality of data. develop a recognition system based on HMM Toolkit (HTK).
Maximum accuracy of 93.74% was achieved during the This approach requires no priori segmentation. Features of
experiments on three different datasets. local densities and statistics are extracted using a vertical
sliding windows technique, where each line image is trans-
formed into a series of extracted feature vectors. HTK is used
E. ARABIC SCRIPT in the training phase, and Viterbi algorithm is used in the
Research on handwritten Arabic OCR systems has passed recognition phase. The system gave an accuracy of 80.26%
through various stages over the past two decades. Studies for words with ‘‘Arabic-numbers’’ database and 78.95% with
in the early 2000s focused mainly on the neural net- IFN / ENIT database.
work methods for recognition and developed variants of In a study conducted in 2016 by Elleuch et al. [173],
databases [165]. In 2002, Pechwitz et al. [37] developed convolutional neural network (CNN) based on support vector
the first IFN/ENIT-database to allow for the training and machine (SVM) is explored for recognizing offline handwrit-
testing of Arabic OCR systems. This is one of the highly ten Arabic. The model automatically extracts features from
cited databases and has been cited more than 470 times. raw input and performs classification.
Another database was developed by Mozaffari et al. [166] an In 2018, researchers applied the technique of DCNN (deep
Mozaffari and Soltanizadeh [167] in 2006. It stores grey-scale CNN) for recognizing the offline and handwritten Arabic
images of isolated offline handwritten 17,740 Arabic / Farsi characters [174]. An accuracy of 98.86% was achieved when
numerals and 52,380 characters. Another notable dataset the strategy of DCNN using transfer learning was applied
containing Arabic handwritten text images was introduced to two datasets. In another similar study [175] an OCR
by Mezghani et al. [168]. The dataset has an open vocab- technique based on HOG (Histograms of Oriented Gradient)
ulary written by multiple writers (AHTID/ MW). It can [176] for feature extraction and SVM for character classifi-
be used for word and sentence recognition, and writer cation was used on the handwritten dataset. The dataset con-
identification [169]. tained names of Jordanian cities, towns and villages yielded
A survey by Lorigo and Govindaraju [18] provides a an accuracy of 99%. However, when the researchers used
comprehensive review of the Arabic handwriting recognition multichannel neural network for segmentation and CNN for
methodologies and databases used until 2006. This includes recognition of machine-printed characters, the experiments
research studies carried out on IFN/ENIT database. These on 18pt font showed an overall accuracy of 94.38%.
studies mostly involved artificial neural networks (ANNs), In 2019, Sahlol et al. [177] applied hybrid machine
Hidden Markov Models (HMM), holistic and segmentation- learning approach on CENPARMI dataset. The technique
based recognition approaches. The limitations pointed out by used the rough neighbourhood sets along with binary whale

VOLUME 8, 2020 142657

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

optimization algorithm. Researcher claims that the proposed perceptrons in a final attempt to recognize the input numeral.
technique consumes less amount of time in recognizing The proposed scheme provided 99.27% recognition accu-
the characters as compared to VGGnet, Resnet, Nasnet, racy vs the fuzzy modelling technique, which provided the
Mobilenet, Inception, and Xception. Alrehali et al. [71] used accuracy of 95%.
CNN on various datasets of historical Arabic manuscripts Desai [28] used neural networks for the numeral recog-
and achieved an accuracy of 74% to 88%. In an interest- nition of Gujrati script. The researcher used a multi-layer
ing study Ali and Suresha [79] used classifier fusion tech- feed-forward neural network for the classification of digits.
nique based on a fusion of features moments invariants(MI), However, the recognition rate was low at 82%.
runlength matrix(RLM), statistical properties of intensity his- Kumar et al. [184], [185] proposed a method for line
togram(SFIH) and wavelet decomposition(WD) and clas- segmentation of handwritten Devanagari text. An accuracy
sifiers modified quadratic discriminate functions(MQDF), of 91.5% for line segmentation and 98.1% for word seg-
support vector machine(SVM) and random forest(RF). mentation was achieved. Perwej and Chaturvedi [186] used
Researcher claim that the fusion technique provided accuracy backpropagation based neural network for the recogni-
of 97% to 99.8%, which is among the highest in Arabic tion of handwritten characters. The results showed that
handwritten character recognition. the highest recognition rate of 98.5% was achieved.
Obaidullah et al. [151] proposed Handwritten Numeral
F. INDIAN SCRIPT Script Identification or HNSI framework based on four
Indian script is collection of scripts used in the sub-continent indic scripts, namely, Bangla, Devanagari, Roman and Urdu.
namely Devanagari [138], Bangla [116], Hindi [178], The researchers used different classifiers, namely NBTree,
Gurmukhi [63], Kannada [179] etc. One of the earliest PART, Random Forest, SMO, Simple Logistic and MLP
research on Devanagari (Hindi) script was proposed and evaluated the performance against the true positive rate.
in 2000 by Lehal and Bhatt [180]. The research was con- Performance of MLP was found to be better than the rest.
ducted on Devanagari script and English numerals. The MLP was then used for bi and tri-script identification.
researchers used data that was already in an isolated form in Bi-script combination of Bangla and Urdu gave the highest
order to avoid the segmentation phase. The research is based accuracy rate of 90.9% on MLP, while the highest accuracy
on statistical and structural algorithms [181]. The results rate of 74% was achieved in a tri-script combination of
of Devanagari scripts were better than English numerals. Bangla, roman and Urdu.
Devanagari had a recognition rate of 89% with 4.5 confusion In a multi dataset experiment [116], researchers applied
rate, while English numerals had a recognition rate of 78% a lightweight model based on 13 layers of CNN with
with confusion rate of 18%. 2-sub layers on four datasets of Bangla language. An accu-
Patil and Subbareddy [182] was the first researcher to use racy of 98%, 96.81%, 95.71%, and 96.40% was achieved
neural network approach for the identification of Indian doc- when the model was applied on CMATERdb, ISI,
uments. The researchers propose a system capable of read- BanglaLekhaIsolated dataset and mixed datasets respectively.
ing English, Hindi and Kannada scripts. A modular neural CNN based model was also applied on ancient documents
network was used for script identification while a two-stage written in Devanagari or Sanskrit script in another study.
feature extraction system was developed, first to dilate the Results, when compared with Google’s vision, OCR gave an
document image and second to find average pixel distribution accuracy of 93.32% vs 92.90%.
in the resulting images. In 2019 sahare and Dhok [77] conducted research on old
Sharma et al. [46] proposed a scheme based on quadratic Indian documents in Devanagari script. These documents
classifier for the recognition of the Devanagari script. The had text diffusion due to degrading. Researchers used graph
researchers used 64 directional features based on chain distance theory to carve up the characters that were over-
code histogram [143] for feature recognition. The proposed lapped due to diffusion before apply support vector machine
scheme resulted in 98.86% and 80.36% accuracy in recogniz- to confirm the segmentation results. Paper also proposes a
ing Devanagari characters and numeral, respectively. Fivefold set of features based on the geometrical properties of the
cross-validation was used for the computation of results. characters. These features claim to provide character recog-
Two research studies [50], [183] presented in 2007 were nition accuracy of 98.8% and 99.6%. During a study on
based on the use of fuzzy modelling for character recogni- Kannada, researchers used the HOG feature extraction
tion of Indian script. The researchers claim that the use of method along with SVM classifier for spam image detection
reinforcement learning on a small database of 3500 Hindi with an accuracy of 94.2% [78]. Research on Indian scripts
numerals helped achieve a recognition rate of 95%. is very diverse, and a number of researchers are involved in
Another research carried out on Hindi numerals [25] used research on multiple scripts. This is the reason why a number
a relatively large dataset of 22,556 isolated numeral samples of research articles on character recognition of Indian scripts
of Devanagari and 23,392 samples of Bangla scripts. The are growing each year. researchers have used techniques like
researchers used three Multi-layer perceptron classifiers to Tesseract OCR and google multilingual OCR [113], Convo-
classify the characters. In case of a rejection, a 4th percep- lutional Neural Network (CNN) [70], [114], Deep Belief Net-
tron was used based on the output of the previous three work with the distributed average of gradients feature [187],

142658 VOLUME 8, 2020

 TABLE 5. Summary of frequently used feature extraction and classification techniques: Data corresponding to last three years (2017-2019). Studies corresponding to ‘‘Indian’’ script do include research on
scripts belonging to Devanagari, Bangla, Hindi, Kannada etc.

VOLUME 8, 2020
J. Memon et al.: Handwritten OCR: A Comprehensive SLR

142659
 TABLE 5. (Continued.) Summary of frequently used feature extraction and classification techniques: Data corresponding to last three years (2017-2019). Studies corresponding to ‘‘Indian’’ script do include
research on scripts belonging to Devanagari, Bangla, Hindi, Kannada etc.

142660
VOLUME 8, 2020
J. Memon et al.: Handwritten OCR: A Comprehensive SLR
VOLUME 8, 2020
J. Memon et al.: Handwritten OCR: A Comprehensive SLR

TABLE 5. (Continued.) Summary of frequently used feature extraction and classification techniques: Data corresponding to last three years (2017-2019). Studies corresponding to ‘‘Indian’’ script do include
research on scripts belonging to Devanagari, Bangla, Hindi, Kannada etc.

142661
 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

Modified Neural Network with the aid of elephant herding specific technique models a different style of characters
optimization [188], VGG (Visual Geometry Group) [117] and and quality of the dataset.
SVM classifier with the polynomial and linear kernel [80]. 3) Most of the published research studies propose a solu-
tion for one language or even a subset of a language.
Publicly available datasets also include stimuli that are
VIII. RESEARCH TRENDS
aligned well with each other and fail to incorporate
Characters written by different individuals create large intr-
examples that correspond well with real-life scenarios,
aclass variability, which makes it difficult for classifiers to
i.e. writing styles, distorted strokes, variable character
perform robustly. Lately, the research in the domain of optical
thickness and illumination [213].
character recognition has moved towards a deep learning
4) It was also observed that researchers are increasingly
approach [189], [190] with little emphasis on handcrafted
using Convolutional Neural Networks(CNN) for the
features. Deep learning approach has produced improved
recognition of handwritten and machine-printed char-
classification accuracy at the cost of increased computational
acters. This is due to the fact that CNN based archi-
complexity, especially during the training phase.
tectures are well suited for recognition tasks where
In this section, we have analyzed hand character recog-
input is an image. CNN was initially used for object
nition research trend in the last three years (2017-2019).
recognition tasks in images, e.g. the ImageNet Large
Our analysis is summarized in Table 5. Table 5 includes
Scale Visual Recognition Challenge (ILSVRC) 216].
script under investigation, techniques or classification tech-
AlexNet [215], GoogLeNet [216] and ResNet [217] are
nique employed for OCR, year of publication and respec-
some of the CNN based architectures widely used for
tive reference number. This table gives a holistic view of
visual recognition tasks.
how researchers working on some of the widely used lan-
guages are trying to solve the problem of optical character
recognition.
B. FUTURE WORK
Table 5 highlights the fact that the bulk of recent
1) As mentioned in Section VII, research in OCR domain
publications have employed a deep learning approach in
is usually done on some of the most widely spoken
some form. Especially CNN is being used extensively for
languages. This is partially due to non-availability of
the recognition of optical characters. This is partially due
datasets on other languages. One of the future research
to the availability of large datasets. Researchers usually
direction is to conduct research on languages other
employ a deep learning approach for a language that has large
than widely spoken languages, i.e. regional languages
enough dataset for deep learning to learn meaningful model.
and endangered languages. This can help preserve the
As stated above, although frameworks based on deep learning
cultural heritage of vulnerable communities and will
methods have obtained improved classification accuracy but
also create a positive impact on strengthening global
at the cost of increased computational complexity. There
synergy.
are few recent studies that have utilized classical feature
2) Another research problem that needs the attention of
extraction approach in combination of feature selection algo-
research community is to build systems that can rec-
rithms and have obtained state-of-the-art result, for example,
ognize on-screen characters and text in different con-
[177], [191], [192].
ditions in daily life scenarios, e.g. text in captions or
news tickers, text on signboards, text on billboards etc.
IX. CONCLUSION AND FUTURE WORK This is the domain of ‘‘recognition / classification /
A. CONCLUSION text in the wild’’. This is a complex problem to solve
1) Optical character recognition has been around for the as a system for such a scenario needs to deal with
last eight (8) decades. However, initially, products background clutters, variable illumination condition,
that recognize optical characters were mostly devel- variable camera angles, distorted characters and vari-
oped by large technology companies. Development able writing styles [213].
of machine learning and deep learning has enabled 3) To build a robust system for ‘‘text in the wild’’,
individual researchers to develop algorithms and tech- researchers need to come up with challenging datasets
niques, which can recognize handwritten manuscripts that are comprehensive enough to incorporate all
with greater accuracy. possible variations in characters. One such effort
2) In this literature review, we systematically extracted is [218]. In another attempt, the research community
and analyzed research publications on six widely spo- has launched ‘‘ICDAR 2019: Robustreading chal-
ken languages. We explored that some techniques lenge on multilingual scene text detection and recog-
perform better on one script than on another, e.g. nition’’ [219]. Aim of this challenge invites research
multilayer perceptron classifier gave better accuracy studies that propose a robust system for multi-lingual
on Devanagri, and Bangla numerals [25], 140] but gave text recognition in daily life or ‘‘in the wild’’ scenario.
average results for other languages [119], [133], [134]. Recently report for this challenge has been published
The difference may have been due to the fact of how and winner methods for different tasks in the challenge

142662 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

are all based on different deep learning architectures, [14] B. Gatos, I. Pratikakis, and S. J. Perantonis, ‘‘An adaptive binarization
e.g. CNN, RNN or LSTM. technique for low quality historical documents,’’ in Proc. Int. Workshop
Document Anal. Syst. Berlin, Germany: Springer, 2004, pp. 102–113.
4) Characters written by different individuals create large [15] J. He, Q. D. M. Do, A. C. Downton, and J. H. Kim, ‘‘A comparison of
intra-class variability, which makes it difficult for clas- binarization methods for historical archive documents,’’ in Proc. 8th Int.
sifiers to perform robustly. Although with the increas- Conf. Document Anal. Recognit. (ICDAR), 2005, pp. 538–542.
[16] T. Sari, L. Souici, and M. Sellami, ‘‘Off-line handwritten Arabic character
ing utilization of complex deep learning architectures, segmentation algorithm: ACSA,’’ in Proc. 8th Int. Workshop Frontiers
obtained classification accuracy has improved at the Handwriting Recognit., 2002, pp. 452–457.
same time computational complexity (especially dur- [17] T. M. Mitchell, Machine Learning, 1st ed. New York, NY, USA:
McGraw-Hill, 1997.
ing the training phase of classifier) has grown. This
[18] L. M. Lorigo and V. Govindaraju, ‘‘Offline Arabic handwriting recogni-
creates a hurdle in the development of a real-time, tion: A survey,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 28, no. 5,
robust system for hand character recognition. pp. 712–724, May 2006.
5) Published research studies have proposed various [19] R. A. Khan, A. Meyer, H. Konik, and S. Bouakaz, ‘‘Saliency-based
framework for facial expression recognition,’’ Frontiers Comput. Sci.,
systems for OCR but one aspect that needs to improve vol. 13, no. 1, pp. 183–198, Feb. 2019.
is the commercialization of research. Commercializa- [20] Y. LeCun, Y. Bengio, and G. Hinton, ‘‘Deep learning,’’ Nature, vol. 521,
tion of research will help to build low-cost real-life no. 7553, pp. 436–444, 2015.
[21] T. M. Breuel, A. Ul-Hasan, M. A. Al-Azawi, and F. Shafait, ‘‘High-
systems for OCR that can turn lots of invaluable infor- performance OCR for printed English and Fraktur using LSTM net-
mation into searchable/digital data [220]. works,’’ in Proc. 12th Int. Conf. Document Anal. Recognit., Aug. 2013,
pp. 683–687.
[22] B. Kitchenham, R. Pretorius, D. Budgen, O. P. Brereton, M. Turner,
ACKNOWLEDGMENT M. Niazi, and S. Linkman, ‘‘Systematic literature reviews in software
(Jamshed Memon, Maira Sami, Rizwan Ahmed Khan, and engineering—A tertiary study,’’ Inf. Softw. Technol., vol. 52, no. 8,
Mueen Uddin contributed equally to this work.) pp. 792–805, 2010.
[23] S. Nidhra, M. Yanamadala, W. Afzal, and R. Torkar, ‘‘Knowledge transfer
challenges and mitigation strategies in global software development—
REFERENCES A systematic literature review and industrial validation,’’ Int. J. Inf.
Manage., vol. 33, no. 2, pp. 333–355, Apr. 2013.
[1] C. C. Tappert, C. Y. Suen, and T. Wakahara, ‘‘The state of the art in [24] A. Graves and J. Schmidhuber, ‘‘Offline handwriting recognition with
online handwriting recognition,’’ IEEE Trans. Pattern Anal. Mach. Intell., multidimensional recurrent neural networks,’’ in Proc. Adv. Neural Inf.
vol. 12, no. 8, pp. 787–808, Aug. 1990, doi: 10.1109/34.57669. Process. Syst., 2009, pp. 545–552.
[2] M. Kumar, S. R. Jindal, M. K. Jindal, and G. S. Lehal, ‘‘Improved [25] U. Bhattacharya and B. B. Chaudhuri, ‘‘Handwritten numeral databases
recognition results of medieval handwritten Gurmukhi manuscripts using of indian scripts and multistage recognition of mixed numerals,’’ IEEE
boosting and bagging methodologies,’’ Neural Process. Lett., vol. 50, Trans. Pattern Anal. Mach. Intell., vol. 31, no. 3, pp. 444–457, Mar. 2009.
pp. 43–56, Sep. 2018.
[26] A. Graves, M. Liwicki, S. Fernández, R. Bertolami, H. Bunke, and
[3] M. A. Radwan, M. I. Khalil, and H. M. Abbas, ‘‘Neural networks pipeline J. Schmidhuber, ‘‘A novel connectionist system for unconstrained hand-
for offline machine printed Arabic OCR,’’ Neural Process. Lett., vol. 48, writing recognition,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 31,
no. 2, pp. 769–787, Oct. 2018. no. 5, pp. 855–868, May 2009.
[4] P. Thompson, R. T. Batista-Navarro, G. Kontonatsios, J. Carter, E. Toon, [27] T. Plötz and G. A. Fink, ‘‘Markov models for offline handwriting recog-
J. McNaught, C. Timmermann, M. Worboys, and S. Ananiadou, ‘‘Text nition: A survey,’’ Int. J. Document Anal. Recognit., vol. 12, no. 4, p. 269,
mining the history of medicine,’’ PLoS ONE, vol. 11, no. 1, pp. 1–33, 2009.
Jan. 2016. [28] A. A. Desai, ‘‘Gujarati handwritten numeral optical character reorga-
[5] K. D. Ashley and W. Bridewell, ‘‘Emerging AI & Law approaches to nization through neural network,’’ Pattern Recognit., vol. 43, no. 7,
automating analysis and retrieval of electronically stored information in pp. 2582–2589, Jul. 2010.
discovery proceedings,’’ Artif. Intell. Law, vol. 18, no. 4, pp. 311–320, [29] G. Vamvakas, B. Gatos, and S. J. Perantonis, ‘‘Handwritten character
Dec. 2010, doi: 10.1007/s10506-010-9098-4. recognition through two-stage foreground sub-sampling,’’ Pattern Recog-
[6] R. Zanibbi and D. Blostein, ‘‘Recognition and retrieval of mathemat- nit., vol. 43, no. 8, pp. 2807–2816, Aug. 2010.
ical expressions,’’ Int. J. Document Anal. Recognit., vol. 15, no. 4, [30] D. C. Cireşan, U. Meier, L. M. Gambardella, and J. Schmidhuber, ‘‘Deep,
pp. 331–357, Dec. 2012, doi: 10.1007/s10032-011-0174-4. big, simple neural nets for handwritten digit recognition,’’ Neural Com-
[7] I. K. Pathan, A. A. Ali, and R. J. Ramteke, ‘‘Recognition of offline put., vol. 22, no. 12, pp. 3207–3220, Dec. 2010.
handwritten isolated Urdu character,’’ Adv. Comput. Res., vol. 4, no. 1, [31] J. Pradeep, E. Srinivasan, and S. Himavathi, ‘‘Diagonal based feature
pp. 117–121, 2012. extraction for handwritten character recognition system using neural
[8] M. T. Parvez and S. A. Mahmoud, ‘‘Offline Arabic handwritten text network,’’ in Proc. 3rd Int. Conf. Electron. Comput. Technol. (ICECT),
recognition: A survey,’’ ACM Comput. Surv., vol. 45, no. 2, p. 23, 2013. vol. 4, Apr. 2011, pp. 364–368.
[9] S. D. Connell and A. K. Jain, ‘‘Template-based online character recogni- [32] D. C. Ciresan, U. Meier, L. M. Gambardella, and J. Schmidhuber,
tion,’’ Pattern Recognit., vol. 34, no. 1, pp. 1–14, Jan. 2001. ‘‘Convolutional neural network committees for handwritten character
[10] S. Mori, C. Y. Suen, and K. Yamamoto, ‘‘Historical review of OCR classification,’’ in Proc. Int. Conf. Document Anal. Recognit., Sep. 2011,
research and development,’’ Proc. IEEE, vol. 80, no. 7, pp. 1029–1058, pp. 1135–1139.
Jul. 1992. [33] V. Patil and S. Shimpi, ‘‘Handwritten English character recognition using
[11] R. A. Wilkinson, J. Geist, S. Janet, P. J. Grother, C. J. Burges, R. Creecy, neural network,’’ Elixir Comput. Sci. Eng., vol. 41, pp. 5587–5591, Nov.
B. Hammond, J. J. Hull, N. Larsen, T. P. Vogl, ‘‘The first census optical 2011.
character recognition system conference,’’ US Dept. Commerce, Nat. [34] K. Gregor, I. Danihelka, A. Graves, D. J. Rezende, and D. Wierstra,
Inst. Standards Technol., Gaithersburg, MD, USA, Tech. Rep. 4912, 1992, ‘‘DRAW: A recurrent neural network for image generation,’’ 2015,
vol. 184. arXiv:1502.04623. [Online]. Available: http://arxiv.org/abs/1502.04623
[12] Z. M. Kovács-V, ‘‘A novel architecture for high quality hand-printed [35] R. Plamondon and S. N. Srihari, ‘‘Online and off-line handwriting
character recognition,’’ Pattern Recognit., vol. 28, no. 11, pp. 1685–1692, recognition: A comprehensive survey,’’ IEEE Trans. Pattern Anal. Mach.
Nov. 1995. Intell., vol. 22, no. 1, pp. 63–84, Jan. 2000.
[13] C. Wolf, J.-M. Jolion, and F. Chassaing, ‘‘Text localization, enhancement [36] N. Arica and F. T. Yarman-Vural, ‘‘An overview of character recognition
and binarization in multimedia documents,’’ in Proc. Object Recognit. focused on off-line handwriting,’’ IEEE Trans. Syst., Man, Cybern. C,
Supported User Interact. Service Robots, vol. 2, 2002, pp. 1037–1040. Appl. Rev., vol. 31, no. 2, pp. 216–233, May 2001.

VOLUME 8, 2020 142663

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

[37] M. Pechwitz, S. S. Maddouri, V. Märgner, N. Ellouze, and H. Amiri, [61] S. N. Srihari, X. Yang, and G. R. Ball, ‘‘Offline Chinese handwriting
‘‘IFN/ENIT-database of handwritten Arabic words,’’ in Proc. CIFED, recognition: An assessment of current technology,’’ Frontiers Comput.
vol. 2, 2002, pp. 127–136. Sci. China, vol. 1, no. 2, pp. 137–155, May 2007.
[38] M. S. Khorsheed, ‘‘Off-line Arabic character recognition—A review,’’ [62] J. Pradeep, E. Srinivasan, and S. Himavathi, ‘‘Neural network based
Pattern Anal. Appl., vol. 5, no. 1, pp. 31–45, 2002. recognition system integrating feature extraction and classification for
[39] I.-S. Oh and C. Y. Suen, ‘‘A class-modular feedforward neural network for english handwritten,’’ Int. J. Eng.-Trans. B, Appl., vol. 25, no. 2, p. 99,
handwriting recognition,’’ Pattern Recognit., vol. 35, no. 1, pp. 229–244, 2012.
Jan. 2002. [63] P. Singh and S. Budhiraja, ‘‘Feature extraction and classification tech-
[40] S. N. Srihari, S.-H. Cha, H. Arora, and S. Lee, ‘‘Individuality of hand- niques in OCR systems for handwritten Gurmukhi script—A survey,’’ Int.
writing,’’ J. Forensic Sci., vol. 47, no. 4, pp. 1–17, 2002. J. Eng. Res. Appl., vol. 1, no. 4, pp. 1736–1739, 2011.
[41] M. Pechwitz and V. Maergner, ‘‘HMM based approach for handwritten [64] I. Shamsher, Z. Ahmad, J. K. Orakzai, and A. Adnan, ‘‘OCR for printed
Arabic word recognition using the IFN/ENIT–database,’’ in Proc. 7th Int. Urdu script using feed forward neural network,’’ Proc. World Acad. Sci.,
Conf. Document Anal. Recognit., 2003, p. 890. Eng. Technol., vol. 23, pp. 172–175, Aug. 2007.
[42] C.-L. Liu, K. Nakashima, H. Sako, and H. Fujisawa, ‘‘Handwritten [65] R. Al-Jawfi, ‘‘Handwriting Arabic character recognition LeNet using
digit recognition: Benchmarking of state-of-the-art techniques,’’ Pattern neural network,’’ Int. Arab J. Inf. Technol., vol. 6, no. 3, pp. 304–309,
Recognit., vol. 36, no. 10, pp. 2271–2285, Oct. 2003. 2009.
[43] U. Pal and B. B. Chaudhuri, ‘‘Indian script character recognition: A sur- [66] C.-L. Liu and C. Y. Suen, ‘‘A new benchmark on the recognition of
vey,’’ Pattern Recognit., vol. 37, no. 9, pp. 1887–1899, Sep. 2004. handwritten bangla and farsi numeral characters,’’ Pattern Recognit.,
[44] C.-L. Liu, S. Jaeger, and M. Nakagawa, ‘‘Online recognition of Chinese vol. 42, no. 12, pp. 3287–3295, Dec. 2009.
characters: The state-of-the-art,’’ IEEE Trans. Pattern Anal. Mach. Intell., [67] C.-L. Liu and H. Fujisawa, ‘‘Classification and learning for character
vol. 26, no. 2, pp. 198–213, Feb. 2004. recognition: Comparison of methods and remaining problems,’’ in Proc.
Int. Workshop Neural Netw. Learn. Document Anal. Recognit., 2005,
[45] Z.-L. Bai and Q. Huo, ‘‘A study on the use of 8-directional features for
pp. 1–7.
online handwritten Chinese character recognition,’’ in Proc. 8th Int. Conf.
Document Anal. Recognit. (ICDAR), 2005, pp. 262–266. [68] H. Zhang and Z. Cheng, ‘‘An advanced pyramid network technology for
optical character recognition,’’ J. Phys., Conf. Ser., vol. 1302, Aug. 2019,
[46] N. Sharma, U. Pal, F. Kimura, and S. Pal, ‘‘Recognition of off-line
Art. no. 022042.
handwritten devnagari characters using quadratic classifier,’’ in Com-
puter Vision, Graphics and Image Processing. Berlin, Germany: Springer, [69] R. Ptucha, F. P. Such, S. Pillai, F. Brockler, V. Singh, and P. Hutkowski,
2006, pp. 805–816. ‘‘Intelligent character recognition using fully convolutional neural net-
works,’’ Pattern Recognit., vol. 88, pp. 604–613, Apr. 2019.
[47] A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, ‘‘Connectionist
[70] B. Dessai and A. Patil, ‘‘A deep learning approach for optical character
temporal classification: Labelling unsegmented sequence data with recur-
recognition of handwritten Devanagari script,’’ in Proc. 2nd Int. Conf.
rent neural networks,’’ in Proc. 23rd Int. Conf. Mach. Learn. (ICML),
Intell. Comput., Instrum. Control Technol. (ICICICT), vol. 1, Jul. 2019,
2006, pp. 369–376.
pp. 1160–1165.
[48] M. Bulacu, L. Schomaker, and A. Brink, ‘‘Text-independent writer identi-
[71] B. Alrehali, N. Alsaedi, H. Alahmadi, and N. Abid, ‘‘Historical Arabic
fication and verification on offline Arabic handwriting,’’ in Proc. 9th Int.
manuscripts text recognition using convolutional neural network,’’ in
Conf. Document Anal. Recognit. (ICDAR), Sep. 2007, pp. 769–773.
Proc. 6th Conf. Data Sci. Mach. Learn. Appl. (CDMA), Mar. 2020,
[49] M. Liwicki, A. Graves, S. Fernàndez, H. Bunke, and J. Schmidhuber, pp. 37–42.
‘‘A novel approach to on-line handwriting recognition based on bidi-
[72] E. Shaikh, I. Mohiuddin, A. Manzoor, G. Latif, and N. Mohammad,
rectional long short-term memory networks,’’ in Proc. 9th Int. Conf.
‘‘Automated grading for handwritten answer sheets using convolutional
Document Anal. Recognit. (ICDAR), 2007, pp. 1–5.
neural networks,’’ in Proc. 2nd Int. Conf. New Trends Comput. Sci.
[50] M. Hanmandlu and O. V. R. Murthy, ‘‘Fuzzy model based recognition of (ICTCS), Oct. 2019, pp. 1–6.
handwritten numerals,’’ Pattern Recognit., vol. 40, no. 6, pp. 1840–1854,
[73] J. Gan, W. Wang, and K. Lu, ‘‘A new perspective: Recognizing online
Jun. 2007.
handwritten Chinese characters via 1-dimensional CNN,’’ Inf. Sci.,
[51] H. Khosravi and E. Kabir, ‘‘Introducing a very large dataset of handwrit- vol. 478, pp. 375–390, Apr. 2019.
ten farsi digits and a study on their varieties,’’ Pattern Recognit. Lett., [74] S. Ghasemi and A. H. Jadidinejad, ‘‘Persian text classification via
vol. 28, no. 10, pp. 1133–1141, Jul. 2007. character-level convolutional neural networks,’’ in Proc. 8th Conf. AI
[52] A. Graves, M. Liwicki, H. Bunke, J. Schmidhuber, and S. Fernández, Robot., 10th RoboCup Iranopen Int. Symp. (IRANOPEN), Apr. 2018,
‘‘Unconstrained on-line handwriting recognition with recurrent neural pp. 1–6.
networks,’’ in Proc. Adv. Neural Inf. Process. Syst., 2008, pp. 577–584. [75] A. Boukharouba and A. Bennia, ‘‘Novel feature extraction technique for
[53] F. Yin, Q.-F. Wang, X.-Y. Zhang, and C.-L. Liu, ‘‘ICDAR 2013 Chinese the recognition of handwritten digits,’’ Appl. Comput. Informat., vol. 13,
handwriting recognition competition,’’ in Proc. 12th Int. Conf. Document no. 1, pp. 19–26, Jan. 2017.
Anal. Recognit. (ICDAR), Aug. 2013, pp. 1464–1470. [76] R. Verma and J. Ali, ‘‘A-survey of feature extraction and classification
[54] M. Zimmermann and H. Bunke, ‘‘Automatic segmentation of the IAM techniques in OCR systems,’’ Int. J. Comput. Appl. Inf. Technol., vol. 1,
off-line database for handwritten English text,’’ in Proc. 16th Int. Conf. no. 3, pp. 1–3, 2012.
Pattern Recognit., vol. 4, 2002, pp. 35–39. [77] P. Sahare and S. B. Dhok, ‘‘Robust character segmentation and recogni-
[55] R. A. Khan, A. Crenn, A. Meyer, and S. Bouakaz, ‘‘A novel database tion schemes for multilingual Indian document images,’’ IETE Tech. Rev.,
of children’s spontaneous facial expressions (LIRIS-CSE),’’ Image Vis. vol. 36, no. 2, pp. 209–222, Mar. 2019.
Comput., vols. 83–84, pp. 61–69, Mar. 2019. [78] F. Naiemi, V. Ghods, and H. Khalesi, ‘‘An efficient character recognition
[56] S. Rajasekaran and G. V. Pai, Neural Networks, Fuzzy Systems and method using enhanced HOG for spam image detection,’’ Soft Comput.,
Evolutionary Algorithms: Synthesis and Applications. New Delh, India: vol. 23, no. 22, pp. 11759–11774, Nov. 2019.
PHI Learning, 2017. [79] A. A. A. Ali and M. Suresha, ‘‘A novel features and classifiers fusion
[57] P. Vithlani and C. Kumbharana, ‘‘A study of optical character patterns technique for recognition of Arabic handwritten character script,’’ Social
identified by the different OCR algorithms,’’ Int. J. Sci. Res. Publications, Netw. Appl. Sci., vol. 1, no. 10, p. 1286, Oct. 2019.
vol. 5, no. 3, pp. 2250–3153, 2015. [80] V. A. Naik and A. A. Desai, ‘‘Multi-layer classification approach
[58] H. Sharif and R. A. Khan, ‘‘A novel framework for automatic detection of for online handwritten Gujarati character recognition,’’ in Computa-
autism: A study on corpus callosum and intracranial brain volume,’’ 2019, tional Intelligence: Theories, Applications and Future Directions, vol. 2.
arXiv:1903.11323. [Online]. Available: https://arxiv.org/abs/1903.11323 Singapore: Springer, 2019, pp. 595–606.
[59] A. K. Jain, J. Mao, and K. M. Mohiuddin, ‘‘Artificial neural networks: [81] L. Yang, C. Y. Suen, T. D. Bui, and P. Zhang, ‘‘Discrimination of similar
A tutorial,’’ Computer, vol. 29, no. 3, pp. 31–44, Mar. 1996. handwritten numerals based on invariant curvature features,’’ Pattern
[60] S. N. Nawaz, M. Sarfraz, A. Zidouri, and W. G. Al-Khatib, ‘‘An approach Recognit., vol. 38, no. 7, pp. 947–963, Jul. 2005.
to offline Arabic character recognition using neural networks,’’ in Proc. [82] J. Yang, K. Yu, Y. Gong, and T. Huang, ‘‘Linear spatial pyramid match-
10th IEEE Int. Conf. Electron., Circuits Syst. (ICECS), vol. 3, Dec. 2003, ing using sparse coding for image classification,’’ in Proc. IEEE Conf.
pp. 1328–1331. Comput. Vis. Pattern Recognit., Jun. 2009, vol. 1, no. 2, p. 6.

142664 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

[83] R. A. Khan, A. Meyer, H. Konik, and S. Bouakaz, ‘‘Framework [104] G. Jeh and J. Widom, ‘‘SimRank: A measure of structural-context
for reliable, real-time facial expression recognition for low resolu- similarity,’’ in Proc. 8th ACM SIGKDD Int. Conf. Knowl. Discov-
tion images,’’ Pattern Recognit. Lett., vol. 34, no. 10, pp. 1159–1168, ery Data Mining (KDD), 2002, pp. 538–543. [Online]. Available:
Jul. 2013. http://doi.acm.org/10.1145/775047.775126
[84] M. Haddoud, A. Mokhtari, T. Lecroq, and S. Abdeddaïm, ‘‘Com- [105] L. A. Zager and G. C. Verghese, ‘‘Graph similarity scoring and
bining supervised term-weighting metrics for SVM text classification matching,’’ Appl. Math. Lett., vol. 21, no. 1, pp. 86–94, 2008.
with extended term representation,’’ Knowl. Inf. Syst., vol. 49, no. 3, [Online]. Available: http://www.sciencedirect.com/science/article/
pp. 909–931, Dec. 2016. pii/S0893965907001012
[85] J. Ning, J. Yang, S. Jiang, L. Zhang, and M.-H. Yang, ‘‘Object track- [106] E. Leicht, P. Holme, and M. Newman, ‘‘Vertex similarity in networks,’’
ing via dual linear structured SVM and explicit feature map,’’ in Phys. Rev. E, Stat. Phys. Plasmas Fluids Relat. Interdiscip. Top., vol. 73,
Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2016, Feb. 2006, Art. no. 026120.
pp. 4266–4274. [107] P. Sahare and S. B. Dhok, ‘‘Multilingual character segmentation and
[86] Q.-Q. Tao, S. Zhan, X.-H. Li, and T. Kurihara, ‘‘Robust face detection recognition schemes for Indian document images,’’ IEEE Access, vol. 6,
using local CNN and SVM based on kernel combination,’’ Neurocomput- pp. 10603–10617, 2018.
ing, vol. 211, pp. 98–105, Oct. 2016. [108] M. Flasiński, ‘‘Graph grammar models in syntactic pattern recogni-
[87] Y. Akbari, M. J. Jalili, J. Sadri, K. Nouri, I. Siddiqi, and C. Djeddi, tion,’’ in Proc. Int. Conf. Comput. Recognit. Syst., R. Burduk, M.
‘‘A novel database for automatic processing of persian handwritten bank Kurzynski, and M. Wozniak, Eds. Springer, 2019. [Online]. Available:
checks,’’ Pattern Recognit., vol. 74, pp. 253–265, Feb. 2018. https://link.springer.com/chapter/10.1007/978-3-030-19738-4_1#citeas
[88] A. A. Chandio, M. Pickering, and K. Shafi, ‘‘Character classification and [109] A. Chaudhuri, K. Mandaviya, P. Badelia, and S. K. Ghosh, ‘‘Optical char-
recognition for Urdu texts in natural scene images,’’ in Proc. Int. Conf. acter recognition systems,’’ in Optical Character Recognition Systems for
Comput., Math. Eng. Technol. (iCoMET), Mar. 2018, pp. 1–6. Different Languages With Soft Computing. Cham, Switzerland: Springer,
[89] S. Alma’adeed, C. Higgens, and D. Elliman, ‘‘Recognition of off-line 2017, pp. 9–41.
handwritten Arabic words using hidden Markov model approach,’’ in [110] R. Hussain, A. Raza, I. Siddiqi, K. Khurshid, and C. Djeddi, ‘‘A compre-
Proc. 16th Int. Conf. Pattern Recognit., vol. 3, 2002, pp. 481–484. hensive survey of handwritten document benchmarks: Structure, usage
[90] S. Alma’adeed, C. Higgins, and D. Elliman, ‘‘Off-line recognition and evaluation,’’ EURASIP J. Image Video Process., vol. 2015, no. 1,
of handwritten Arabic words using multiple hidden Markov mod- p. 46, Dec. 2015.
els,’’ in Research and Development in Intelligent Systems XX. London, [111] R. A. Khan, É. Dinet, and H. Konik, ‘‘Visual attention: Effects of blur,’’
U.K.: Springer, 2004, pp. 33–40. in Proc. IEEE Int. Conf. Image Process., Sep. 2011, pp. 3289–3292.
[91] M. Cheriet, ‘‘Visual recognition of Arabic handwriting: Challenges and [112] T. E. de Campos, B. R. Babu, and M. Varma, ‘‘Character recognition in
new directions,’’ in Arabic and Chinese Handwriting Recognition. Berlin, natural images,’’ in Proc. VISAPP, vol. 7, 2009, pp. 1–8.
Germany: Springer, 2008, pp. 1–21. [113] C. Clausner, A. Antonacopoulos, T. Derrick, and S. Pletschacher,
[92] U. Pal, R. Jayadevan, and N. Sharma, ‘‘Handwriting recognition in indian ‘‘ICDAR2019 competition on recognition of early Indian printed
regional scripts: A survey of offline techniques,’’ ACM Trans. Asian Lang. documents—REID2019,’’ in Proc. Int. Conf. Document Anal. Recognit.
Inf. Process., vol. 11, no. 1, pp. 1–35, Mar. 2012. (ICDAR), Sep. 2019, pp. 1527–1532.
[93] V. L. Sahu and B. Kubde, ‘‘Offline handwritten character recognition [114] K. G. Joe, M. Savit, and K. Chandrasekaran, ‘‘Offline character
techniques using neural network: A review,’’ Int. J. Sci. Res., vol. 2, no. 1, recognition on segmented handwritten kannada characters,’’ in
pp. 87–94, 2013. Proc. Global Conf. Advancement Technol. (GCAT), Oct. 2019,
[94] M. E. Tiwari and M. Shreevastava, ‘‘A novel technique to read small and pp. 1–5.
capital handwritten character,’’ Int. J. Adv. Comput. Res., vol. 2, no. 2, [115] A. Choudhury, H. S. Rana, and T. Bhowmik, ‘‘Handwritten Bengali
p. 127, 2012. numeral recognition using HOG based feature extraction algorithm,’’ in
[95] S. Touj, S. Touj, N. B. Amara, N. B. Amara, H. Amiri, and H. Amiri, Proc. 5th Int. Conf. Signal Process. Integr. Netw. (SPIN), Feb. 2018,
‘‘Two approaches for Arabic script recognition-based segmentation using pp. 687–690.
the Hough transform,’’ in Proc. 9th Int. Conf. Document Anal. Recognit. [116] A. S. A. Rabby, S. Haque, S. Islam, S. Abujar, and S. A. Hossain,
(ICDAR ), vol. 2, Sep. 2007, pp. 654–658. ‘‘BornoNet: Bangla handwritten characters recognition using convolu-
[96] M.-J. Li and R.-W. Dai, ‘‘A personal handwritten Chinese character tional neural network,’’ Procedia Comput. Sci., vol. 143, pp. 528–535,
recognition algorithm based on the generalized Hough transform,’’ in Jan. 2018.
Proc. 3rd Int. Conf. Document Anal. Recognit. (ICDAR), Washington, [117] M. A. Pragathi, K. Priyadarshini, S. Saveetha, A. S. Banu, and
DC, USA, 1995, p. 828. [Online]. Available: http://dl.acm.org/ K. O. M. Aarif, ‘‘Handwritten tamil character recognition UsingDeep
citation.cfm?id= 839278.840308 learning,’’ in Proc. Int. Conf. Vis. Towards Emerg. Trends Commun. Netw.
[97] A. Chaudhuri, ‘‘Some experiments on optical character recognition sys- (ViTECoN), Mar. 2019, pp. 1–5.
tems for different languages using soft computing techniques,’’ Birla Inst. [118] M. Hangarge and B. V. Dhandra, ‘‘Offline handwritten script identifica-
Technol. Mesra, Patna, India, Tech. Rep., 2010. tion in document images,’’ Int. J. Comput. Appl., vol. 4, no. 6, pp. 6–10,
[98] J. Iivarinen and A. Visa, ‘‘Shape recognition of irregular objects,’’ in Jul. 2010.
Intelligent Robots and Computer Vision XV: Algorithms, Techniques, [119] C.-L. Liu, K. Nakashima, H. Sako, and H. Fujisawa, ‘‘Handwritten digit
Active Vision, and Materials Handling. Bellingham, WA, USA: SPIE, recognition using state-of-the-art techniques,’’ in Proc. 8th Int. Workshop
1996, pp. 25–32. Frontiers Handwriting Recognit., 2002, pp. 320–325.
[99] C.-L. Liu, K. Nakashima, H. Sako, and H. Fujisawa, ‘‘Handwritten digit [120] G. Vamvakas, B. Gatos, and S. Perantonis, ‘‘Hierarchical classification of
recognition: Investigation of normalization and feature extraction tech- handwritten characters based on novel structural features,’’ in Proc. 11th
niques,’’ Pattern Recognit., vol. 37, no. 2, pp. 265–279, Feb. 2004. Int. Conf. Frontiers Handwriting Recognition (ICFHR), Montreal, QC,
[100] J. P. M. de Sá, ‘‘Structural pattern recognition,’’ in Pattern Recognition. Canada, 2008, pp. 535–539.
Berlin, Germany: Springer, 2001, pp. 243–289, doi: 10.1007/978-3-642- [121] U. R. Babu, A. K. Chintha, and Y. Venkateswarlu, ‘‘Handwritten digit
56651-6_6. recognition using structural, statistical features and k-nearest neigh-
[101] S. Lavirott and L. Pottier, ‘‘Mathematical formula recognition using bor classifier,’’ Int. J. Inf. Eng. Electron. Bus., vol. 6, no. 1, p. 62,
graph grammar,’’ Proc. SPIE, vol. 3305, pp. 44–52, Apr. 1998, doi: 2014.
10.1117/12.304644. [122] S. B. Ahmed, S. Naz, S. Swati, M. I. Razzak, A. I. Umar, and A. A. Khan,
[102] F. Álvaro, J.-A. Sánchez, and J.-M. Benedí, ‘‘Recognition of on-line ‘‘Ucom offline dataset—An Urdu handwritten dataset generation,’’ Int.
handwritten mathematical expressions using 2D stochastic context-free Arab J. Inf. Technol., vol. 14, no. 2, pp. 239–245, 2017.
grammars and hidden Markov models,’’ Pattern Recognit. Lett., vol. 35, [123] H. El Abed and V. Margner, ‘‘The IFN/ENIT-database—A tool to develop
pp. 58–67, Jan. 2014. [Online]. Available: http://www.sciencedirect. Arabic handwriting recognition systems,’’ in Proc. 9th Int. Symp. Signal
com/science/article/pii/S016786551200308X Process. Appl. (ISSPA), Feb. 2007, pp. 1–4.
[103] S. Melnik, H. Garcia-Molina, and E. Rahm, ‘‘Similarity flooding: A ver- [124] V. Margner and H. El Abed, ‘‘Arabic handwriting recognition competi-
satile graph matching algorithm and its application to schema matching,’’ tion,’’ in Proc. 9th Int. Conf. Document Anal. Recognit. (ICDAR), vol. 2,
in Proc. 18th Int. Conf. Data Eng., 2002, pp. 117–128. 2007, pp. 1274–1278.

VOLUME 8, 2020 142665

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

[125] F. Solimanpour, J. Sadri, and C. Y. Suen, ‘‘Standard databases for recog- [148] S. T. Javed and S. Hussain, ‘‘Improving nastalique specific pre-
nition of handwritten digits, numerical strings, legal amounts, letters and recognition process for Urdu OCR,’’ in Proc. IEEE 13th Int. Multitopic
dates in farsi language,’’ in Proc. 10th Int. Workshop Frontiers Handwrit- Conf., Dec. 2009, pp. 1–6.
ing Recognit., 2006, pp. 1–6. [149] M. W. Sagheer, C. L. He, N. Nobile, and C. Y. Suen, ‘‘A new large Urdu
[126] P. J. Haghighi, N. Nobile, C. L. He, and C. Y. Suen, ‘‘A new large-scale database for off-line handwriting recognition,’’ in Proc. Int. Conf. Image
multi-purpose handwritten farsi database,’’ in Proc. Int. Conf. Image Anal. Anal. Process. Berlin, Germany: Springer, 2009, pp. 538–546.
Recognit. Berlin, Germany: Springer, 2009, pp. 278–286. [150] A. Raza, I. Siddiqi, A. Abidi, and F. Arif, ‘‘An unconstrained bench-
[127] H. Zhang, J. Guo, G. Chen, and C. Li, ‘‘HCL2000—A large-scale hand- mark Urdu handwritten sentence database with automatic line segmen-
written Chinese character database for handwritten character recogni- tation,’’ in Proc. Int. Conf. Frontiers Handwriting Recognit., Sep. 2012,
tion,’’ in Proc. 10th Int. Conf. Document Anal. Recognit. (ICDAR), 2009, pp. 491–496.
pp. 286–290. [151] S. M. Obaidullah, C. Halder, N. Das, and K. Roy, ‘‘Numeral script
[128] U.-V. Marti and H. Bunke, ‘‘The IAM-database: An english sentence identification from handwritten document images,’’ Procedia Comput.
database for offline handwriting recognition,’’ Int. J. Document Anal. Sci., vol. 54, pp. 585–594, Jan. 2015.
Recognit., vol. 5, no. 1, pp. 39–46, Nov. 2002. [152] I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA, USA: SIAM,
[129] C. Moseley, Ed., Atlas of the World’s Languages in Danger. Paris, France: 1992.
UNESCO Publishing, 2010. [153] A. Naseer and K. Zafar, ‘‘Comparative analysis of raw images and meta
[130] S. Tian, U. Bhattacharya, S. Lu, B. Su, Q. Wang, X. Wei, Y. Lu, feature based Urdu OCR using CNN and LSTM,’’ Int. J. Adv. Comput.
and C. L. Tan, ‘‘Multilingual scene character recognition with co- Sci. Appl., vol. 9, no. 1, pp. 419–424, 2018.
occurrence of histogram of oriented gradients,’’ Pattern Recognit., vol. 51, [154] Sami-Ur-Rehman, B. U. Tayyab, M. F. Naeem, A. Ul-Hasan, and
pp. 125–134, Mar. 2016. F. Shafait, ‘‘A multi-faceted OCR framework for artificial Urdu news
[131] A. H. Toselli, E. Vidal, V. Romero, and V. Frinken, ‘‘HMM word graph ticker text recognition,’’ in Proc. 13th IAPR Int. Workshop Document
based keyword spotting in handwritten document images,’’ Inf. Sci., Anal. Syst. (DAS), Apr. 2018, pp. 211–216.
vols. 370–371, pp. 497–518, Nov. 2016. [155] S. B. Ahmed, S. Naz, S. Swati, and M. I. Razzak, ‘‘Handwritten Urdu
[132] S. Deshmukh and L. Ragha, ‘‘Analysis of directional features–stroke character recognition using one-dimensional BLSTM classifier,’’ Neural
and contour for handwritten character Recognition,’’ in Proc. IEEE Int. Comput. Appl., vol. 31, no. 4, pp. 1143–1151, Apr. 2019.
Advance Comput. Conf. (IACC), Mar. 2009, pp. 1114–1118. [156] M. J. Rafeeq, Z. U. Rehman, A. Khan, I. A. Khan, and W. Jadoon, ‘‘Lig-
[133] S. Ahlawat and R. Rishi, ‘‘Off-line handwritten numeral recognition using ature categorization based nastaliq Urdu recognition using deep neural
hybrid feature set—A comparative analysis,’’ Procedia Comput. Sci., networks,’’ Comput. Math. Org. Theory, vol. 25, no. 2, pp. 184–195,
vol. 122, pp. 1092–1099, Jan. 2017. Jun. 2019.
[134] P. Sharma and R. Singh, ‘‘Performance of English character recognition [157] H.-T. Pao, Y. Y. Xu, H.-Y. Chang, and H.-C. Fu, ‘‘User adaptive hand-
with and without noise,’’ Int. J. Comput. Trends Technol., vol. 4, no. 3, writing recognition by self-growing probabilistic decision-based neural
pp. 400–403, 2013. networks,’’ IEEE Trans. Neural Netw., vol. 11, no. 6, pp. 1373–1384,
[135] C. I. Patel, R. Patel, and P. Patel, ‘‘Handwritten character recognition Nov. 2000.
using neural network,’’ Int. J. Sci. Eng. Res., vol. 2, no. 5, pp. 1–6, 2011. [158] D. Lin, F. Lin, Y. Lv, F. Cai, and D. Cao, ‘‘Chinese character CAPTCHA
recognition and performance estimation via deep neural network,’’ Neu-
[136] P. Zhang, T. D. Bui, and C. Y. Suen, ‘‘A novel cascade ensemble classifier
rocomputing, vol. 288, pp. 11–19, May 2018.
system with a high recognition performance on handwritten digits,’’
Pattern Recognit., vol. 40, no. 12, pp. 3415–3429, Dec. 2007. [159] Y. Zhao, W. Xue, and Q. Li, ‘‘A multi-scale CRNN model for Chinese
papery medical document recognition,’’ in Proc. IEEE 4th Int. Conf.
[137] S. Saha, N. Paul, S. K. Das, and S. Kundu, ‘‘Optical character recognition
Multimedia Big Data (BigMM), Sep. 2018, pp. 1–5.
using 40-point feature extraction and artificial neural network,’’ Int. J.
Adv. Res. Comput. Sci. Softw. Eng., vol. 3, no. 4, pp. 1–8, 2013. [160] Y. Luo, Y. Li, S. Huang, and F. Han, ‘‘Multiple Chinese vehicle license
plate localization in complex scenes,’’ in Proc. IEEE 3rd Int. Conf. Image,
[138] M. Avadesh and N. Goyal, ‘‘Optical character recognition for Sanskrit
Vis. Comput. (ICIVC), Jun. 2018, pp. 745–749.
using convolution neural networks,’’ in Proc. 13th IAPR Int. Workshop
[161] H. Yang, L. Jin, and J. Sun, ‘‘Recognition of Chinese text in historical
Document Anal. Syst. (DAS), Apr. 2018, pp. 447–452.
documents with page-level annotations,’’ in Proc. 16th Int. Conf. Fron-
[139] V. Jayasundara, S. Jayasekara, H. Jayasekara, J. Rajasegaran,
tiers Handwriting Recognit. (ICFHR), Aug. 2018, pp. 199–204.
S. Seneviratne, and R. Rodrigo, ‘‘TextCaps: Handwritten character
[162] H. Ren, W. Wang, and C. Liu, ‘‘Recognizing online handwritten Chinese
recognition with very small datasets,’’ in Proc. IEEE Winter Conf. Appl.
characters using RNNs with new computing architectures,’’ Pattern
Comput. Vis. (WACV), Jan. 2019, pp. 254–262.
Recognit., vol. 93, pp. 179–192, Sep. 2019.
[140] S. Mozaffari, K. Faez, and H. R. Kanan, ‘‘Recognition of isolated hand-
[163] X. Zhang and K. Yan, ‘‘An algorithm of bidirectional RNN for offline
written Farsi/Arabic alphanumeric using fractal codes,’’ in Proc. 6th IEEE
handwritten Chinese text recognition,’’ in Proc. Int. Conf. Intell. Comput.
Southwest Symp. Image Anal. Interpretation, Mar. 2004, pp. 104–108.
Cham, Switzerland: Springer, 2019, pp. 423–431.
[141] H. Soltanzadeh and M. Rahmati, ‘‘Recognition of persian handwritten
[164] Y. Zhu, F. Zhuang, J. Yang, X. Yang, and Q. He, ‘‘Adaptively trans-
digits using image profiles of multiple orientations,’’ Pattern Recognit.
fer category-classifier for handwritten Chinese character recognition,’’
Lett., vol. 25, no. 14, pp. 1569–1576, Oct. 2004.
in Proc. Pacific-Asia Conf. Knowl. Discovery Data Mining. Cham,
[142] A. Broumandnia and J. Shanbehzadeh, ‘‘Fast Zernike wavelet moments Switzerland: Springer, 2019, pp. 110–122.
for farsi character recognition,’’ Image Vis. Comput., vol. 25, no. 5, [165] N. Mezghani, A. Mitiche, and M. Cheriet, ‘‘On-line recognition
pp. 717–726, May 2007. of handwritten Arabic characters using a kohonen neural network,’’
[143] H. Freeman and L. Davis, ‘‘A corner-finding algorithm for chain-coded in Proc. 8th Int. Workshop Frontiers Handwriting Recognit., 2002,
curves,’’ IEEE Trans. Comput., vol. C-26, no. 3, pp. 297–303, Mar. 1977. pp. 490–495.
[144] G. Sokar, E. E. Hemayed, and M. Rehan, ‘‘A generic OCR using [166] S. Mozaffari, K. Faez, F. Faradji, M. Ziaratban, and S. M. Golzan,
deep siamese convolution neural networks,’’ in Proc. IEEE 9th Annu. ‘‘A comprehensive isolated Farsi/Arabic character database for handwrit-
Inf. Technol., Electron. Mobile Commun. Conf. (IEMCON), Nov. 2018, ten OCR research,’’ in Proc. 10th Int. Workshop Frontiers Handwriting
pp. 1238–1244. Recognit., 2006, pp. 1–6.
[145] B. Alizadehashraf and S. Roohi, ‘‘Persian handwritten character recog- [167] S. Mozaffari and H. Soltanizadeh, ‘‘ICDAR 2009 handwritten
nition using convolutional neural network,’’ in Proc. 10th Iranian Conf. Farsi/Arabic character recognition competition,’’ in Proc. 10th Int.
Mach. Vis. Image Process. (MVIP), Nov. 2017, pp. 247–251. Conf. Document Anal. Recognit. (ICDAR), 2009, pp. 1413–1417.
[146] S. A. A. A. Arani, E. Kabir, and R. Ebrahimpour, ‘‘Handwritten farsi word [168] A. Mezghani, S. Kanoun, M. Khemakhem, and H. E. Abed, ‘‘A database
recognition using NN-based fusion of HMM classifiers with different for Arabic handwritten text image recognition and writer identifica-
types of features,’’ Int. J. Image Graph., vol. 19, no. 01, Jan. 2019, tion,’’ in Proc. Int. Conf. Frontiers Handwriting Recognit., Sep. 2012,
Art. no. 1950001. pp. 399–402.
[147] S. Naz, K. Hayat, M. I. Razzak, M. W. Anwar, S. A. Madani, and [169] M. Khayyat, L. Lam, and C. Y. Suen, ‘‘Learning-based word spotting
S. U. Khan, ‘‘The optical character recognition of Urdu-like cursive system for Arabic handwritten documents,’’ Pattern Recognit., vol. 47,
scripts,’’ Pattern Recognit., vol. 47, no. 3, pp. 1229–1248, Mar. 2014. no. 3, pp. 1021–1030, Mar. 2014.

142666 VOLUME 8, 2020

J. Memon et al.: Handwritten OCR: A Comprehensive SLR

[170] M. Lutf, X. You, Y.-M. Cheung, and C. L. P. Chen, ‘‘Arabic font recog- [193] Y.-C. Wu, F. Yin, and C.-L. Liu, ‘‘Improving handwritten Chinese text
nition based on diacritics features,’’ Pattern Recognit., vol. 47, no. 2, recognition using neural network language models and convolutional
pp. 672–684, Feb. 2014. neural network shape models,’’ Pattern Recognit., vol. 65, pp. 251–264,
[171] Y. Elarian, I. Ahmad, S. Awaida, W. G. Al-Khatib, and A. Zidouri, May 2017.
‘‘An Arabic handwriting synthesis system,’’ Pattern Recognit., vol. 48, [194] C. Shi, Y. Wang, F. Jia, K. He, C. Wang, and B. Xiao, ‘‘Fisher vector
no. 3, pp. 849–861, Mar. 2015. for scene character recognition: A comprehensive evaluation,’’ Pattern
[172] E. M. Hicham, H. Akram, and S. Khalid, ‘‘Using features of local den- Recognit., vol. 72, pp. 1–14, Dec. 2017.
sities, statistics and HMM toolkit (HTK) for offline Arabic handwriting [195] Z. Feng, Z. Yang, L. Jin, S. Huang, and J. Sun, ‘‘Robust shared feature
text recognition,’’ J. Electr. Syst. Inf. Technol., vol. 4, no. 3, pp. 387–396, learning for script and handwritten/machine-printed identification,’’ Pat-
Dec. 2017. tern Recognit. Lett., vol. 100, pp. 6–13, Dec. 2017.
[173] M. Elleuch, R. Maalej, and M. Kherallah, ‘‘A new design based-SVM
[196] X. Feng, H. Yao, and S. Zhang, ‘‘Focal CTC loss for Chinese optical
of the CNN classifier architecture with dropout for offline Arabic hand-
character recognition on unbalanced datasets,’’ Complexity, vol. 2019,
written recognition,’’ Procedia Comput. Sci., vol. 80, pp. 1712–1723,
Jan. 2019, Art. no. 9345861.
Jan. 2016.
[174] C. Boufenar, A. Kerboua, and M. Batouche, ‘‘Investigation on deep [197] L. Xu, Y. Wang, X. Li, and M. Pan, ‘‘Recognition of handwritten
learning for off-line handwritten Arabic character recognition,’’ Cognit. Chinese characters based on concept learning,’’ IEEE Access, vol. 7,
Syst. Res., vol. 50, pp. 180–195, Aug. 2018. pp. 102039–102053, 2019.
[175] N. A. Jebril, H. R. Al-Zoubi, and Q. A. Al-Haija, ‘‘Recognition of hand- [198] B. B. Chaudhuri and C. Adak, ‘‘An approach for detecting and cleaning
written Arabic characters using histograms of oriented gradient (HOG),’’ of struck-out handwritten text,’’ Pattern Recognit., vol. 61, pp. 282–294,
Pattern Recognit. Image Anal., vol. 28, no. 2, pp. 321–345, Apr. 2018. Jan. 2017.
[176] R. A. Khan, A. Meyer, H. Konik, and S. Bouakaz, ‘‘Pain detection through [199] B. Su and S. Lu, ‘‘Accurate recognition of words in scenes without char-
shape and appearance features,’’ in Proc. IEEE Int. Conf. Multimedia acter segmentation using recurrent neural network,’’ Pattern Recognit.,
Expo (ICME), Jul. 2013, pp. 1–6. vol. 63, pp. 397–405, Mar. 2017.
[177] A. T. Sahlol, M. A. Elaziz, M. A. A. Al-Qaness, and S. Kim, ‘‘Handwrit- [200] R. Graef and M. M. N. Morsy, ‘‘A novel hybrid optical character recogni-
ten Arabic optical character recognition approach based on hybrid whale tion approach for digitizing text in forms,’’ in Proc. Int. Conf. Design Sci.
optimization algorithm with neighborhood rough set,’’ IEEE Access, Res. Inf. Syst. Technol. Cham, Switzerland: Springer, 2019, pp. 206–220.
vol. 8, pp. 23011–23021, 2020. [201] M. Ahmed and A. I. Abidi, ‘‘Performance comparison of ANN and
[178] K. Dutta, P. Krishnan, M. Mathew, and C. Jawahar, ‘‘Towards accurate template matching on English character recognition,’’ Int. J. Advance
handwritten word recognition for Hindi and Bangla,’’ in Proc. Nat. Res., Ideas Innov. Technol., vol. 5, no. 4, pp. 367–372, 2019.
Conf. Comput. Vis., Pattern Recognit., Image Process., Graph. Singapore: [202] M. Yashodha, S. Niranjan, and V. N. M. Aradhya, ‘‘Deep learning for
Springer, 2019, pp. 367–372. trilingual character recognition,’’ Int. J. Natural Comput. Res., vol. 8,
[179] B. M. Sagar, G. Shobha, and R. P. Kumar, ‘‘OCR for printed Kannada no. 1, pp. 52–58, Jan. 2019.
text to machine editable format using database approach,’’ WSEAS Trans. [203] A. Yousaf, M. J. Khan, M. J. Khan, N. Javed, H. Ibrahim, K. Khurshid,
Comput., vol. 7, no. 6, pp. 766–769, 2008. and K. Khurshid, ‘‘Size invariant handwritten character recognition using
[180] G. Lehal and N. Bhatt, ‘‘A recognition system for Devnagri and English
single layer feedforward backpropagation neural networks,’’ in Proc. 2nd
handwritten numerals,’’ in Advances in Multimodal Interfaces. Berlin,
Int. Conf. Comput., Math. Eng. Technol. (iCoMET), Jan. 2019, pp. 1–7.
Germany: Springer, 2000, pp. 442–449.
[181] F. Kimura and M. Shridhar, ‘‘Handwritten numerical recognition based [204] S. Yousfi, S.-A. Berrani, and C. Garcia, ‘‘Contribution of recur-
on multiple algorithms,’’ Pattern Recognit., vol. 24, no. 10, pp. 969–983, rent connectionist language models in improving LSTM-based Arabic
Jan. 1991. text recognition in videos,’’ Pattern Recognit., vol. 64, pp. 245–254,
[182] S. B. Patil and N. V. Subbareddy, ‘‘Neural network based system for script Apr. 2017.
identification in Indian documents,’’ Sadhana, vol. 27, no. 1, pp. 83–97, [205] R. Elanwar, W. Qin, and M. Betke, ‘‘Making scanned Arabic documents
Feb. 2002. machine accessible using an ensemble of SVM classifiers,’’ Int. J. Docu-
[183] M. Hanmandlu, J. Grover, V. K. Madasu, and S. Vasikarla, ‘‘Input fuzzy ment Anal. Recognit., vol. 21, nos. 1–2, pp. 59–75, Jun. 2018.
modeling for the recognition of handwritten Hindi numerals,’’ in Proc. [206] I. A. Doush, F. Alkhateeb, and A. H. Gharaibeh, ‘‘A novel Arabic OCR
4th Int. Conf. Inf. Technol. (ITNG), Apr. 2007, pp. 208–213. post-processing using rule-based and word context techniques,’’ Int. J.
[184] N. K. Garg, D. L. Kaur, and D. M. Kumar, ‘‘Segmentation of handwritten Document Anal. Recognit., vol. 21, nos. 1–2, pp. 77–89, 2018.
hindi text,’’ Int. J. Comput. Appl., vol. 1, no. 4, pp. 22–26, Feb. 2010. [207] S. S. R. Rizvi, A. Sagheer, K. Adnan, and A. Muhammad, ‘‘Optical
[185] N. K. Garg, L. Kaur, and M. K. Jindal, ‘‘A new method for line segmenta- character recognition system for nastalique Urdu-like script languages
tion of handwritten hindi text,’’ in Proc. 7th Int. Conf. Inf. Technol., New using supervised learning,’’ Int. J. Pattern Recognit. Artif. Intell., vol. 33,
Generat. (ITNG), 2010, pp. 392–397. no. 10, Sep. 2019, Art. no. 1953004.
[186] Y. Perwej and A. Chaturvedi, ‘‘Machine recognition of hand written [208] K. U. U. Rehman and Y. D. Khan, ‘‘A scale and rotation invariant Urdu
characters using neural networks,’’ 2012, arXiv:1205.3964. [Online]. Nastalique ligature recognition using cascade forward back propagation
Available: http://arxiv.org/abs/1205.3964 neural network,’’ IEEE Access, vol. 7, pp. 120648–120669, 2019.
[187] S. Karthik and K. S. Murthy, ‘‘Deep belief network based approach to [209] R. Sarkhel, N. Das, A. Das, M. Kundu, and M. Nasipuri, ‘‘A multi-scale
recognize handwritten Kannada characters using distributed average of deep quad tree based feature extraction method for the recognition of iso-
gradients,’’ Cluster Comput., vol. 22, no. S2, pp. 4673–4681, Mar. 2019. lated handwritten characters of popular indic scripts,’’ Pattern Recognit.,
[188] S. Kowsalya and P. Periasamy, ‘‘Recognition of tamil handwritten charac-
vol. 71, pp. 78–93, Nov. 2017.
ter using modified neural network with aid of elephant herding optimiza-
[210] F. Sarvaramini, A. Nasrollahzadeh, and M. Soryani, ‘‘Persian handwrit-
tion,’’ Multimedia Tools Appl., vol. 78, no. 17, pp. 25043–25061, 2019.
[189] S. Naz, A. I. Umar, R. Ahmad, I. Siddiqi, S. B. Ahmed, M. I. Razzak, ten character recognition using convolutional neural network,’’ in Proc.
and F. Shafait, ‘‘Urdu nastaliq recognition using convolutional–recursive Iranian Conf. Electr. Eng. (ICEE), 2018, pp. 1676–1680.
deep learning,’’ Neurocomputing, vol. 243, pp. 80–87, Jun. 2017. [211] S. Valikhani, F. Abdali-Mohammadi, and A. Fathi, ‘‘Online continu-
[190] M. Al-Ayyoub, A. Nuseir, K. Alsmearat, Y. Jararweh, and B. Gupta, ous multi-stroke Persian/Arabic character recognition by novel spatio-
‘‘Deep learning for Arabic NLP: A survey,’’ J. Comput. Sci., vol. 26, temporal features for digitizer pen devices,’’ Neural Comput. Appl.,
pp. 522–531, May 2018. vol. 32, pp. 3853–3872, May 2019.
[191] M. Kumar, M. K. Jindal, R. K. Sharma, and S. R. Jindal, ‘‘Offline [212] P. Kiaei, M. Javaheripi, and H. Mohammadzade, ‘‘High accuracy farsi
handwritten numeral recognition using combination of different feature language character segmentation and recognition,’’ in Proc. 27th Iranian
extraction techniques,’’ Nat. Acad. Sci. Lett., vol. 41, no. 1, pp. 29–33, Conf. Electr. Eng. (ICEE), Apr. 2019, pp. 1692–1698.
Feb. 2018. [213] S. Long, X. He, and C. Yao, ‘‘Scene text detection and recognition:
[192] N. D. Cilia, C. D. Stefano, F. Fontanella, and A. S. di Freca, The deep learning era,’’ 2018, arXiv:1811.04256. [Online]. Available:
‘‘A ranking-based feature selection approach for handwritten character https://arxiv.org/abs/1811.04256
recognition,’’ Pattern Recognit. Lett., vol. 121, pp. 77–86, Apr. 2019. [214] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang,
[Online]. Available: http://graphonomicsfor e-citizens: e-health, e- A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei,
society, e-education and http://www.sciencedirect.com/science/article/ ‘‘ImageNet large scale visual recognition challenge,’’ Int. J. Comput. Vis.,
pii/S0167865518301272 vol. 115, no. 3, pp. 211–252, Dec. 2015.

VOLUME 8, 2020 142667

 J. Memon et al.: Handwritten OCR: A Comprehensive SLR

[215] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ‘‘ImageNet classifi- MAIRA SAMI received the B.E. degree in the field
cation with deep convolutional neural networks,’’ in Proc. 25th Int. of computer and information systems engineering,
Conf. Neural Inf. Process. Syst. (NIPS), vol. 1. Red Hook, NY, and the master’s degree in data engineering and
USA: Curran Associates, 2012, pp. 1097–1105. [Online]. Available: information management from the NED Univer-
http://dl.acm.org/citation.cfm?id=2999134.2999257 sity of Engineering and Technology (NEDUET).
[216] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, She is currently working as a Faculty Member with
V. Vanhoucke, and A. Rabinovich, ‘‘Going deeper with convolutions,’’ SZABIST, Karachi, Pakistan.
in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2015,
pp. 1–9.
[217] K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for
image recognition,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit.
(CVPR), Jun. 2016, pp. 770–778.
[218] T.-L. Yuan, Z. Zhu, K. Xu, C.-J. Li, T.-J. Mu, and S.-M. Hu, ‘‘A large RIZWAN AHMED KHAN received the Ph.D.
Chinese text dataset in the wild,’’ J. Comput. Sci. Technol., vol. 34, no. 3, degree in computer science from Université
pp. 509–521, May 2019, doi: 10.1007/s11390-019-1923-y. Claude Bernard Lyon 1, France, in 2013. He has
[219] N. Nayef, Y. Patel, M. Busta, P. N. Chowdhury, D. Karatzas, worked as a Postdoctoral Research Associate
W. Khlif, J. Matas, U. Pal, J.-C. Burie, C.-L. Liu, and J.-M. Ogier, with the Laboratoire d’InfoRmatique en Image et
‘‘ICDAR2019 robust reading challenge on multi-lingual scene text Systemes d’information (LIRIS), Lyon, France.
detection and recognition—RRC-MLT-2019,’’ 2019, arXiv:1907.00945. He is currently working as a Professor with
[Online]. Available: https://arxiv.org/abs/1907.00945 Barrett Hodgson University, Karachi, Pakistan.
[220] G. D. Markman, D. S. Siegel, and M. Wright, ‘‘Research and technology His research interests include artificial intelli-
commercialization,’’ J. Manage. Stud., vol. 45, no. 8, pp. 1401–1423, gence, computer vision, machine learning, and
2008. [Online]. Available: https://onlinelibrary.wiley.com/ doi/abs/10. human perception.
1111/j.1467-6486.2008.00803.x

MUEEN UDDIN received the Bachelor of Sci-

ence and Master of Science degrees in computer
sciences from Isra University, Pakistan, and the
Ph.D. degree from Universiti Teknologi Malaysia
(UTM). He is currently working as an Associate
JAMSHED MEMON received the Ph.D. degree Professor with Ilma University, Karachi, Pakistan.
in information systems from Universiti Teknologi He possesses diverse education and research back-
Malaysia, in 2015. He is currently working as ground. He also possesses very strong research and
an Associate Professor with Quest International publication background with over 85 international
University Perak (QIUP), Malaysia. He has exten- publications to his name. He has strong networks
sive research and industry experience. He has and security related background, where he has developed many algorithms
authored over 20 international research articles. and techniques to secure networks and cloud related applications. He is also
His research interests include green information working on blockchain technology to provide and enable healthcare related
technology, information security, and technology solutions.
entrepreneurship.

142668 VOLUME 8, 2020

View publication stats