0% found this document useful (0 votes)
48 views11 pages

Paper 17573

Uploaded by

publishlambert
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
48 views11 pages

Paper 17573

Uploaded by

publishlambert
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 11

ISSN (Online) 2581-9429

IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

Advancements and Trends in CNN-Based


Handwritten Text Recognition: A Comprehensive
Survey
More Meghana Laxman and Prof. S. G. Joshi
Department of Computer Engineering,
Viswabharti College of Engineering, Ahmednagar, India
[email protected]

Abstract: Handwritten Text Recognition (HTR) is pivotal in transforming handwritten documents into
digital format, enabling efficient search, storage, and information retrieval. In this project, we explore the
application of Convolutional Neural Networks (CNNs) for HTR tasks. We comprehensively analyzed
existing literature surveys to understand the current state-of-the-art techniques, methodologies, and
challenges in HTR using CNNs. The survey encompassed various aspects including network architectures,
dataset compositions, preprocessing techniques, and evaluation metrics. Our findings reveal the evolution
of CNN-based HTR systems and highlight key trends in research, such as the integration of attention
mechanisms and recurrent neural networks to enhance recognition accuracy and contextual understanding.
Through this analysis, we provide insights into the advancements and future directions in CNN-based HTR
methodologies.

Keywords: Handwritten Characters, CNN, Literature Reviews.

I. INTRODUCTION
The field of handwritten text recognition has come a long way in the last 30 years; it helps transform human-readable
handwriting into machine-readable text codes. Many scientific and commercial fields are interested in handwritten text
recognition because it can automatically handle massive amounts of handwritten data. The fundamental issue with
handwritten writing has always been that different depictions of the same character differ. Individual variances in
writing style, as well as contextual variations in writing style, add another layer of complexity to the task of deciphering
English handwritten characters. Also, the writer's mood and the setting in which they write may impact their style.
The intricacy of optical pattern recognition only becomes obvious when attempts are made to construct a computer
system that might interpret handwriting. The best approach to developing systems that can recognise handwriting is
supposedly using artificial neural networks. For more accurate handwriting recognition, neural networks are a great
model for how the human brain functions. Automated systems can now read handwriting as well as, if not better than,
humans. People utilise many writing styles, some very challenging to decipher. Reading handwriting may sometimes be
challenging and time-consuming, especially when dealing with many papers from different people. The proposed
system would be most effective with neural networks integrated because of their superior ability to decipher complex
data and identify patterns that would be difficult to detect using traditional approaches. This project aims to develop a
model that can recognize handwritten numbers and characters from an image using the Convolution Neural Network
idea. We constructed a basic Convolutional Neural Network (CNN) system trained on the NIST dataset.
Handwritten Text Recognition (HTR) is a fundamental task in document analysis and understanding, facilitating the
conversion of handwritten documents into digital format for various applications such as archival, search, and
translation. Over the years, Convolutional Neural Networks (CNNs) have emerged as a powerful tool in image
recognition and understanding, demonstrating remarkable success in tasks ranging from object detection to scene
understanding. Leveraging the spatial hierarchy and local connectivity patterns present in handwritten text images,
CNNs offer a promising avenue for automating the process of text recognition from scanned or digitized documents.

Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 449


www.ijarsct.co.in
ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

The application of CNNs in HTR tasks has garnered significant attention from academia and industry, owing to their
ability to learn hierarchical representations directly from raw pixel intensities, alleviating the need for handcrafted
feature extraction. Using convolutional layers, pooling operations, and non-linear activation functions, CNNs can
effectively capture intricate spatial patterns and structural dependencies inherent in handwritten text images, enabling
accurate recognition even in the presence of variations in writing styles, sizes, and orientations.
This review research paper aims to provide a comprehensive overview of the state-of-the-art methodologies,
techniques, and challenges in Handwritten Text Recognition using Convolutional Neural Networks. We survey a wide
range of literature encompassing various aspects of CNN-based HTR, including network architectures, dataset
compositions, preprocessing techniques, evaluation metrics, and benchmarking protocols. By synthesizing insights
from existing studies, we aim to shed light on the evolution of CNN-based HTR systems, identify key trends and
advancements, and delineate avenues for future research and development in this burgeoning field.
Through this review, we endeavor to provide researchers, practitioners, and enthusiasts with a comprehensive
understanding of the current landscape of HTR using CNNs, offering valuable insights and perspectives to inspire
further innovation and progress in the field.

II. LITERATURE REVIEWS


An increasing number of individuals can understand the MODI script, yet its use is on the decline. Information that is
culturally important, historically noteworthy, and administratively relevant were all included in the saved archaic
historical MODI manuscripts. Thousands of handwritten MODI papers maintained in official and public sectors provide
information essential to the contemporary period. Among the 3,350 instances included in the MODI-HHDoc Dataset
are photographs of old handwritten MODI documents. Using this dataset, one may create a system to digitize,
recognize, transliterate, and extract information from old MODI documents written by hand. [1]
The MODI script is used by hundreds, yet few understand it. Handwritten MODI records include culturally,
historically, and administratively valuable information. Many of the thousands of handwritten MODI documents
provide vital information on current public and governmental affairs. The MODI-HHDoc Dataset comprises 3,350
handwritten mediaeval MODI documents. This dataset may be used to digitise, recognise, transcribe, and translate
historic MODI documents to get MODI script information. The system needed data to handle multiple approaches. [2]
Researchers performed a comprehensive survey to assess Modi script character recognition. This paper covers all key
methods, strategies, and advances for Modi script character identification. The authors may employ machine learning,
deep learning, and image processing. This article provides a complete review of Modi script character recognition
procedures and highlights key trends, challenges, and research opportunities. The survey is reliable and necessary for
Modi script character recognition academics, practitioners, and enthusiasts, according to the IEEE publication
platform.[3]

Fig.1 Handwritten script


The researchers develop an effective Modi script character identification method in their published work. This work
aims to improve Modi script character recognition by developing and deploying novel methods, including machine

Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 450


www.ijarsct.co.in
ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

learning and image processing. The article summarises the suggested technique, including its essential components and
procedures, to enhance recognition results. IJCA publication validates the work, which sheds light on Modi script
character identification for scholars and practitioners and significantly advances the field.4]
The researchers thoroughly evaluated Modi script character identification systems. This review will include all
computational methods used by researchers up to the publishing date, including machine learning and image
processing. The authors want to provide researchers, experts, and fans a complete history, present, and future of Modi
script character identification. The International Journal of Computer-Aided Character Recognition (IJCA) may provide
legitimacy to the review and provide a resource for Modi script character identification researchers.[5]
Character segmentation utilising old handwritten MODI Script papers is the subject of this study. According to the
Vertical Projection Profile (VPP) method, a zero-pixel column ought to appropriately split two adjacent letters into
individual ones on a single line. Updates to the VPP output are necessary as this script prints characters on a single line
without raising a hand. In this study, the authors provide a novel approach to character segmentation error reduction by
line separation utilising dual thresholding criteria. It was born out of the author's extensive research on the subject. Any
researcher would benefit from the study's methodology since it is easy, quick, and inexpensive.[6]
The main objective of this study is to review the existing literature on MODI handwritten character recognition.
Because every writer has a distinct style, character identification from handwriting has become more difficult in recent
years. Computers may make previously illegible handwriting intelligible using character recognition software.
Numerous historical efforts to understand Indian languages have left their mark on the written record. Here, the MODI
script especially deserves more respect. To ensure the preservation of these priceless archives, several language-specific
identification algorithms have been developed. The variety and complexity of writing styles seen in every Handwritten
Optical Character Recognition (HOCR) dataset keeps academics interested in Indian scripts.[7]
How to handle the many different writing styles included in HOCR datasets has become a big problem in the field of
Handwritten Optical Character Recognition (HOCR) in the last several decades. Indian culture and geography combine
to form a singular identity as a place of unity in diversity; this, in turn, has resulted in a rich tapestry of spoken and
written languages spoken and written throughout the nation. A lot of work has gone into creating HOCR systems in
India that can handle different types of handwriting. There has been a dramatic increase in the study of character
recognition systems due to the complexity and diversity of Indian characters. The studies that used MODI for
handwritten character recognition are the primary topic of the previous research that is covered in this publication.[8]
Creating a system that can identify handwritten characters is the main goal of this project. While the title implies that
the tactics used to identify handwritten characters were carefully researched, the presented data does not specify
particular methodologies or procedures. Due to the importance and prevalence of handwritten writing, especially in
many languages and scripts, this study enhances optical character recognition systems. The ongoing efforts to improve
algorithms for handwritten content identification are shown by their usage in character recognition.[9]
run a battery of tests to ensure that several algorithms can identify Modi characters written by hand, such as Naive
Bayes with a maximisation function, Euclidean distance, and Minimum distance with an improvement and
minimization function. Automation in feature vector binarization and techniques for extracting concavity, structural,
and gradient features are also proposed. To validate the offered intelligent recognition approaches, the invariant Modi
handwritten dataset is used. Combining concavity characteristics, a correlation coefficient, and maximum voting
increases the recognition rate for handwritten Modi numerals. Meanwhile, GSC features do a good job of describing
Modi's vowels. The correlation coefficient becomes more efficient as classes with distinct characteristics increase.
Although it is challenging, a minimal distance classifier with a minimization algorithm can recognise both free-form
cursive and stylish handwritten Modi characters. The GSC feature extraction approach can characterise the most
baffling Modi aspects by effectively extracting geometrical, structural, and concavity information. Our best recognition
rate of 83.19% was achieved by combining the GSC feature extraction method with the Minimum distance classifier
and its reduction function. The identification rate is validated using 95,000 Modi handwritten characters from 55
classes. The proposed system has an invariant quality that has been shown. The enormous number of character symbols
is a significant obstacle when producing training data for the ingenious but dated Modi document recognition system.
one way to circumvent this is to construct an annotated Modi character dataset from the original handwritten Modi
manuscripts. [10]
Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 451
www.ijarsct.co.in
ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

2.1 MACHINE LEARNING AND DEEP LEARNING


The Maratha government adopted the "MODI lipi" script as their official writing system between the 17th and the
middle of the 20th century. Also included in this script are religious texts from the Indian state of Maharashtra. Being a
"MODI manuscript," it provides cultural significance by describing the period in which it was written. Not many people
are aware of this "lipi," which is a shame since it has the potential to educate and motivate today's youth. with the field
of handwritten character identification, there has been discussion of developing a MODI recognition system to aid with
learning. Character recognition is only one of several pattern recognition applications that use deep learning methods.
The use of transfer learning algorithms has recently improved recognition problems. To categorise MODI characters,
the proposed approach uses CNN and Vgg16 algorithms.[11]
For optical character recognition (OCR), MODI is only one of several difficult languages. The MODI script is difficult
to decipher since it is largely cursive and has few outwardly comparable letters; deep learning algorithms like
InceptionV3 and RestNet have yet to be applied to the Modi script. For offline deep learning handwritten character
recognition, it is recommended to use Residual in conjunction with the InceptionV3 framework. About twenty-five
individuals had their handwriting samples included in the Modi Barakhadi collection. Examination of the dataset,
including 7,721,001 items. Separate the characters using the Otsu binarization method toprepare them for further
processing. Benchmark the performance of each approach on pre-processed data using a real-world, multi-author
database of handwritten characters. A model precision of 0.86 and a recognition accuracy of 94.552 percent were
produced by applying RestNet50 to processed pictures. Test results show that InceptionV3's processed-image
recognition is 93.923 percent accurate, with a model precision of 0.843 percent.[12]

Fig.2 CNN Modules for handwritten character


The suggested experiment in this study employed the deep learning model Alexnet to correctly identify characters in a
script that is thought to be from ancient India. to assess the efficacy of the deep learning technique in this particular
experimental environment, we just used the specific characters from the old MODI script. Alexnet, a MATLAB model,
was used to accomplish an experimental setup success rate of 89.72%. All fifty-eight letters of the handwritten MODI
script were used in this experiment. The final output considered one hundred distinct visual versions for each character.
The success of the trials is detailed in this presentation, which also describes the present status of the planned research
approach. Consequently, the simple and effective implementation using MATLAB programming is shown via the use
of the deep learning model in the first phase of the experimental research activity.in [13]
The objective of creating a practical Character Identification System was to keep identification accuracy high without
compromising the quality of the input documents. Utilising robust Convolutional Neural Network (CNN) methods, the
Modi Script Character Recognizer System was developed and put into operation (MSCR). team has thoroughly
evaluated all 48 letters that make up the MODI script, which includes vowels, consonants, and 10 numbers. In order to
aid our system's training, we collected a large number of handwritten samples from various locations. This literature
review explores the fascinating path to build MSR, a character recognition system, for the ancient MODI script by
digging into the specifics of its development and performance. [14]
Deep learning can solve the problem of recognising handwritten characters in a scanned image. Making written
Gujarati resources digital is of the utmost importance. Findings from this research point to the importance of developing
a technology to scan handwritten text digitally. For this, can employ deep learning techniques, which make working
with photos a breeze. [15]
Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 452
www.ijarsct.co.in
ISSN (Online) 2581-9429
2581
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access,
Access, Double-Blind, Peer-Reviewed,
Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

The developers of this software set out to solve a common problem: how to make it simple for everyone, anywhere to
transpose handwritten notes into text. For those who have trouble reading digital text, there is also an in-built
in Text-to-
Speech feature. e. Scaled up, it has the potential to benefit a great many more individuals. The results of using the
handwriting recognition model have been promising. [16]
The Maratha dynasty promoted Marathi as a sign of strength and culture in the 17th century using M MODI, a script that
is akin to Devanagari. Identification and translation of MODI scripts may be difficult due to its shrinking user
population, script databases that aren't dependable, and a lack of relevant literature. In this innovative research, the
writers
ters test the hypothesis that CNN architecture can recognise MODI numeric and alphabetic letters. After applying a
conventional ML classifier to the data, compare Random Forest with XGBoost; the result is a letter recognition
accuracy of 92% and a number recognition
ecognition accuracy of 93.33%.[17]
Integrating deep learning methods results in an upgraded paradigm. In the first part of the article, the author discusses
the difficulties of reading the old Modi script and how complex methods are required to solve these problems. Data
preparation, model design, and training procedures are likely covered in great depth in the methodology section, which
pertains to the application of deep learning. The results and discussions highlight the improved model's performance
measures
ures compared to more conventional approaches. Expect the conclusion to highlight their deep learninglearning-based
method's success, providing insights into real-world
real world consequences and potential avenues for future study at the
crossroads of historical texts and state-of--the-art
art technology. To better understand the recent advances in Modi script
character identification, it is recommended that you read the whole work. [18]
The course will begin with an overview of the Modi script and the growing significance of automated automated recognition
systems. To address these challenges, the authors may discuss the limitations of existing approaches and the potential
benefits of machine learning. The technique section will likely detail the character recognition methods used. Accura
Accurate
information about the data collection, preparation, and ML algorithms used should be provided. They could review the
training process and any changes to boost the model's precision. The outcomes and talks show their recommended
approach. Learn as much ass you can about the model's accuracy, precision, and recall. Reviewing the model's
performance, the authors may look at how it stacks up against previous research or industry norms. In the last section,
they would review the findings, stress the importance of their work, and propose areas for further research. In addition
to recommendations for extending the approach to other scripts or languages, keep an eye out for the real real-world
implications of the findings. For a more in-depth
in understanding of the specifics,
fics, methodology, and findings of this
inquiry into Modi script character identification, it is recommended to read the whole paper. [19]

Fig.3.. DNN Technique of Character Recognition


Recogn
Since the introduction of digital computers, machine reading simulations have piqued the curiosity of computer scienc
science
researchers. "character recognition" means to recognise printed or handwritten characters to extract machine
machine-readable
text from document pictures. Latin, English, and Chinese are foreign scripts that have proven useful for character
identification. When working with scripts employed in Indian languages, the complicated character identification
Copyright to IJARSCT DOI: 10.48175/
10.48175/IJARSCT-17573 453
www.ijarsct.co.in
ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

approach presents an exceptionally formidable challenge. Devanagari, an abbreviation for the ancient Indian MODI
script, was the initial alphabet used to write Marathi. The historical significance of the script outweighs its lack of
formal usage. It isn't easy to apply MODI character detection on texts that require word-stopping signals, have
comparable character forms, and have a wide variety of writing styles. Innovations in machine learning have been
crucial to the success of character recognition systems. Examine and contrast the most recent feature extraction and
classification approaches to MODI script recognition. [20]

2.2 Transfer Learning


This article details the process of building a supervised Transfer Learning (TL) classification system and compiling an
image dataset of MODI handwritten characters. The network takes weights from an existing Alexnet model and uses
them to retrain a Deep Convolutional Neural Network (DCNN). This network pulls features from every level it can
reach in its role as a feature extractor. In order to train the classifier models, Support Vector Machines (SVMs) with
activation characteristics are used. improve your knowledge of feature analysis and the accuracy of recognition by
using these models. Deep feature discriminating attributes are chosen based on both subjective and objective criteria.
Handwritten MODI characters achieved a recognition accuracy of 92.32% and handwritten Devnagari characters a
recognition accuracy of 97.25%.[21]
First, there will be an introduction explaining the purpose of the study and the specific challenges of detecting
handwritten Modi script. In order to prove that their work is unique, the writers should use Transfer Learning, a
powerful method for improving performance in related jobs by transferring information learnt from one. Extensive
insights into the detailed use of Transfer Learning to improve the recognition model are provided in the methodology
section, which covers a wide range of topics including data preparation, model creation, and knowledge transfer from
pre-trained models. Discussions and outcomes will likely focus on the performance metrics of the recognition model,
comparing it to more traditional methods to emphasise its efficacy.[22]
This method covers data preprocessing and the subtleties of the chosen architecture. Regarding processing speed and
accuracy, among other metrics, the findings and comments provide a thorough assessment of the approach compared to
existing OCR systems. Highlighting the importance of neural networks in the optical character recognition (OCR) of
the Modi script, the conclusion offers indications of possible real-world applications and directions for further research.
The paper delves into how historical scripts and current machine-learning techniques work together to digitize and
archive cultural artifacts. [23]

2.3 CNN
Deep learning algorithms have been successful in various pattern identification tasks, including character recognition.
Convolutional Neural Networks are excellent character recognition deep learning models. CNN can extract features and
recognise characters. This study examines MODI script character identification using a CNN autoencoder. A Support
Vector Machine (SVM) classifies data using gathered features. On-demand data augmentation increases data diversity
and generalizability. Marathi was only written in MODI till 1950. Library and temple collections worldwide have
MODI documents. To get MODI manuscript data from libraries, script character recognition research is needed. The
recommended technique uses a CNN autoencoder for feature extraction and an SVM for the classifier to achieve state-
of-the-art MODI character recognition results. [24]
This study uses cutting-edge technology to recognise MODI script handwriting using a Convolutional Neural Network
(CNN) for feature extraction and a Support Vector Machine (SVM) for the classifier. This work integrates cutting-edge
deep learning algorithms with traditional script analysis approaches to improve Handwritten Character Recognition for
the uncommon and difficult MODI script. The research uses convolutional neural network (CNN)-based feature
extraction to improve identification accuracy in an old screenplay using cutting-edge machine learning methods. [25]
This study uses a CNN and GUI to identify handwritten text. A simple solution for automatic character identification
uses convolutional neural networks (CNNs), which are successful in image-related tasks, and an intuitive user interface.
This study significantly advances the area by introducing a fully functional system that uses state-of-the-art neural
network algorithms for precise identification and prioritizes usability via a graphical user interface. When it comes out

Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 454


www.ijarsct.co.in
ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

in April of 2021, it will be the most cutting-edge answer to the problems that have recently plagued the area of
handwritten character recognition. [26]
The capacity of computers to decipher handwriting will soon be of paramount importance. Handwritten character
recognition technology has the potential to automate several mundane tasks. Convolutional neural networks (CNNs) are
superior to the other two approaches in speed and accuracy, as shown in the study. No matter how much progress CNN
has made in accuracy, the framework required to provide state-of-the-art results remains the same. Enhance precision
by including a spellchecker in the final product and using image-cleaning procedures on inputs. Learning about other
areas that might benefit from the proposed method for automation. [27]

Fig.4 CNN Architecture


A fresh approach to combating the issue of credit card fraud. Addressing the increasing challenges in this subject, the
paper's introduction explains why their analysis is relevant to current techniques. The authors provide a hybrid model
that combines Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs), as detailed in the
techniques section, which deals with data preparation, model construction, and training insights. The model's success
indicators are included in the results and comments. These indicators may include a comparison to more traditional
techniques of fraud detection. The conclusion provides a brief overview of the findings, highlighting the efficacy of
their hybrid CNN-SVM model, and makes recommendations for further study and potential credit card security
applications. It is suggested that you read the whole study to get a better understanding of the novel synergies used for
enhanced fraud detection.[28]
Beginning with a discussion of the field's challenges and the need for more advanced models, the essay will highlight
the distinct benefits of these models by integrating CNNs and SVMs (SVMs). Read the methodology section to learn
about the joint operation of CNNs and SVMs. It should include everything from data pretreatment to the design of the
hybrid classifier and the training method. In the discussion and results sections, you can expect to see performance
metrics for the hybrid model compared to more traditional methods. The conclusion will give a brief overview of the
model's effectiveness, recommendations for further research on handwritten digit recognition, and real-world examples
of its use. If you want to know what innovative methods were employed to improve recognition results, read the whole
article. [29]
The article will probably begin with an introductory section that discusses the difficulties of medical imaging, the
importance of accurate detection methods, and the novelty of their Hybrid CNN-SVM Threshold Segmentation
Approach. Data preparation, model design, training details, threshold segmentation, and the integration of
Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs) are all described in great depth in the
methods section. In the results and discussions sections, you can expect to see performance metrics, comparisons to
other methods, and possible obstacles listed. In the conclusion, you can expect a brief summary of their approach's
effectiveness, along with some suggestions for future medical imaging research and practical applications. You can
learn more about their novel approaches to precise tumour identification and classification by reading the complete
article. [30]
The paper will probably begin with an introduction that describes the script's special problems and stresses using
Convolutional Neural Networks (CNNs) as a contemporary method for character recognition. With its in-depth analysis
of data preparation, CNN architecture, and training complexities, the methodology section provides essential insights
Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 455
www.ijarsct.co.in
ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

into CNN applications. The results and discussions section would highlight performance metrics, maybe comparing the
CNN method to others; the conclusion section would briefly highlight the effectiveness of their CNN-based model for
Modi script character recognition; the end section would hint at potential future research directions and practical
implications related to the preservation of historical scripts through advanced technologies. Delving into the complete
paper guarantees a more profound comprehension of integrating conventional scripts with modern neural network
methods. [31]
The project's introduction likely delves into the unique characteristics of the Modi script and the inherent complexities
in its digital processing. The methodology section will provide a comprehensive overview of the thresholding
techniques employed, detailing their application in image processing and segmentation, with insights into experimental
choices. Through performance measurements and comparative evaluations, these strategies would be shown to work for
Modi script and discussed. The result will likely highlight thresholding strategies and suggest further historical script
digitalization study. Reading the complete study will help you understand the careful examination of Modi script
thresholding strategies. [32]
Handwritten character recognition converts handwritten characters into machine-readable format for processing. You
may use photos, documents, or other character-based sources. Recognising complicated compound handwritten
characters with intricate patterns is difficult. Convolutional neural networks (CNNs) have made great progress in HCR
by learning discriminating characteristics from huge amounts of raw data. This article identifies test dataset characters
using CNNs. CNN's character recognition accuracy and training on an image dataset are the main focus of this work.
CNNs compare and contrast characters' characteristics and forms to identify them. The NIST dataset was used to
evaluate our CNN implementation for handwritten character accuracy. A training batch of 1000 NIST pictures yields
92.91% accuracy on 200 images. [33]
Many past information and bespoke characteristics have been the backbone of traditional handwriting recognition
algorithms. It is not easy to train an optical character recognition (OCR) system using these requirements. In recent
years, researchers in handwriting identification have made great strides by concentrating on deep-learning approaches.
However, further research is essential, and improvements to identification accuracy are required due to the ever-
increasing amounts of handwritten data and the availability of enormous computing power. The best method for
handling handwriting identification challenges is convolutional neural networks (CNNs), which automatically extract
different characteristics by analyzing the structure of handwritten letters and sentences. This proposed study
investigates CNN-based handwritten digit identification using several design parameters such as stride size, receptive
field, kernel size, padding, and dilution. Furthermore, our objective is to assess how well different SGD optimization
techniques enhance the accuracy of digit identification from handwritten input. By using ensemble architecture, the
recognition accuracy of a network is enhanced. Since ensemble topologies increase testing complexity and
computational expense, we aim to achieve equal accuracy using a pure CNN architecture without ensemble
architecture. Therefore, a convolutional neural network (CNN) design is suggested for achieving accuracy that
surpasses ensemble topologies while simultaneously reducing operational complexity and expense. In addition, we
detail the optimal set of learning parameters for CNN construction, which allowed us to break the previous absolute
record for MNIST handwritten digit classification. After a battery of tests, the recognition accuracy for the MNIST
dataset was 99.87%. [34]
New difficulties in feature selection for handwriting identification issues have emerged due to the data expansion,
which has led to large-scale, high-dimensional data. An ideal feature selection algorithm that can improve the
recognition accuracy of a handwriting recognition system while reducing training efforts and computational cost is
required due to practical challenges such as the high degree of individual variation and ambiguity in handwriting styles.
This work proposes a feature selection method based on evolutionary algorithms for handwritten digit recognition,
focusing on the feature selection process. An effective feature set that includes the sample numbers' local and global
properties is obtained by developing a hybrid feature set that combines statistical and geometrical features. The
technique employs a genetic algorithm for feature selection and a k-nearest neighbor for fitness evaluation to find the
unique features of the handwritten digit dataset. Results showed a 66% decrease in the original feature set without
compromising identification accuracy in the studies conducted on the standard The Chars74K handwritten digit dataset.
The findings of the experiment prove that the strategy is successful. [35]
Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 456
www.ijarsct.co.in
ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

These days, computer vision applications rely heavily on Convolutional Neural Networks (CNN). This article presents
the results of an offline study that employed a state-of-the-art CNN to recognise Tamil handwriting. While HTCR has
always relied on human input to extract features, CNNs automate the process. An HP Labs India-developed collection
of isolated handwritten Tamil characters was used. We trained a convolutional neural network (CNN) model using
offline Tamil text and got excellent results on both the training and testing datasets. This was our first attempt at CNN
development. This effort aims to use deep learning methods to establish a standard for offline HTCR. Compared to
more conventional methods, this work's training accuracy of 95.16 percent is far superior. [36]
Optical Character Recognition (OCR) is a scanning technique that can accurately identify printed and handwritten
letters, numbers, and symbols at high electrical speeds. The comprehension of visual information is now known as
Intelligent Character Recognition (ICR). It is possible to transform scanned printed or handwritten characters into
ASCII text using the OCR module's Intelligent Character Recognition (ICR) feature. When exchanging information
digitally, the typical format is ASCII data. Letters, numbers, symbols, and whitespace are all given conventional
numerical values by ASCII. Output Character Recognition (OCR) is the technical name for the process of converting
two-dimensional textual information into machine-encoded text using an electronic equipment. You may use a scanner
or even just take a photo of the text and the recognition algorithm will be able to tell the difference between machine-
written and hand-written text. A Convolutional Neural Network model trained on the NIST dataset, which contains
more than 100,000 pictures, is the subject of this articles' stated objective. In order to determine the likelihood of each
class to which a picture belongs, the network learns from the characteristics retrieved from the pictures. With a loss of
2.53%, we have attained an accuracy of 90.54%. [37]

III. CONCLUSION
This research analyses Convolutional Neural Network-based handwritten text recognition. Synthesised multiple
literature surveys to uncover CNN-based HTR methodology themes and trends. Our investigation shows that CNN
architectures are crucial to HTR task excellence. We also stressed dataset curation, preprocessing, and assessment
criteria in evaluating HTR systems. Advanced neural network topologies, preprocessing methodologies, and assessment
frameworks will improve handwritten text recognition accuracy and efficiency. This study is useful for CNN-based
HTR researchers and practitioners.

REFERENCES
[1]. Hong-Phuong Tran; Andrew Smith; Eric Dimla, “Offline Handwritten Text Recognition using Convolutional
Recurrent Neural Network”, International Conference on Advanced Computing and Applications (ACOMP),
2019
[2]. Manisha Deshmukh, Satish Kolhe, “MODI-HHDoc: Historical MODI Script Handwritten Document
Dataset”, Mendeley Data, V1, 2023
[3]. Aparna Shirkande, Alok Agarwal, “A Review on Various MODI Text Recognition Techniques”, Journal of
Image Processing and Artificial Intelligence, 2023
[4]. Sneha Deshmukh, Dr. Prashant R. Deshmukh, "A Comprehensive Survey on Modi Script Character
Recognition" , IEEE, 2023
[5]. Abhay Verma, Dr. R. K. Gupta, "An Efficient Approach for Modi Script Character Recognition" , IJCA,
2022
[6]. Varsha R. Kurhade, Dr. R. S. Bichkar, "A Review on Modi Script Character Recognition Techniques", IJCA,
2020
[7]. Parag Tamhankar, Krishna Masalkar, and Satish kolhe. A novel approach for character segmentation of
offline handwritten marathi documents written in modi script. Procedia Computer Science, 2020
[8]. Prathmesh Sainath Chidrawar & Vidya Dhamdhere ,A Systematic Review on MODI Script Character
Recognition, Conference paper, (LNNS,volume 727) ,2023
[9]. Sahil Das, Krishna Wankhede, Anand Rituraj, “REVIEW ON MODI HANDWRITTEN CHARACTERS
RECOGNITION”, International Research Journal of Modernization in Engineering Technology and Science,
2022
Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 457
www.ijarsct.co.in
ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

[10]. Prajjwal Bajpai, Mohd. Shahdil Shuaib, “HANDWRITTEN CHARACTER RECOGNITION SYSTEM”,
International Research Journal of Engineering and Technology (IRJET), 2020
[11]. Manisha S. Deshmukh, Satish R. Kolhe, “Character Recognition of the Unconstrained and Invariant Dataset
of the Handwritten Ancient Modi Script Documents”, SSRN-id4330133, 2023
[12]. Samrudhi Bhalerao; H. D. Gadade, “Modi Handwritten Characters Recognition Using Deep Learning
Algorithm”, Iternational Conference on Machine Learning, Computer Systems and Security (MLCSS), 2023
[13]. Chaitali Chandankhede & Rajneeshkaur Sachdeo , “Offline MODI script character recognition using deep
learning techniques”, Multimedia Tools and Applications volume 82, 2023
[14]. Dr. Kirti Mahajan, Niket Tajne, “An Ancient Indian Handwritten Script Character Recognition by Using
Deep Learning Algorithm”, 1st international conference on emerging Scientific Applications in the Field of
Engineering and Technology, 2021
[15]. Jidnyasa Kondhare, Prof.V.A.Yaduvanshi, Ruchira Patil Radhika Kaldate, “Recognition Of Handwritten
Modi Digits And Characters By Using Deep Learning Algorithm”, Journal of Emerging Technologies and
Innovative Research (JETIR), 2022
[16]. Bhargav Rajyagor, Rajnish Rakhlia, “Handwritten Character Recognition using Deep Learning”,
International Journal of Recent Technology and Engineering (IJRTE), 2020
[17]. Athira M Nair, Chrissie Aldo, Blessil Bose, Alex Joseph, Praseetha V.M, Sr. Elizabeth M J, “Handwritten
Character Recognition using Deep Learning in Android Phones”, International Research Journal of
Engineering and Technology (IRJET), 2021
[18]. Maitreyi Ekbote, Aishwary Jadhav, Dayanand Ambawade, “Implementing a Hybrid Deep Learning Approach
to Achieve Classic Handwritten Alphanumeric MODI Recognition”, International Journal of Engineering and
Advanced Technology Volume-12 Issue-1, October 2022
[19]. Dr. Alok Jain, Dr. Poonam Sinha, "Enhanced Modi Script Character Recognition using Deep Learning”,
IJCSMC, 2023
[20]. Aparna S. Gadge, Prof. C. S. Jadhav, "Recognition of Modi Script Characters using Machine Learning",
IJRASET, 2020
[21]. Jossy P. George “Feature Extraction and Classification Techniques of MODI Script Character Recognition”,
Semanticscholar, Corpus ID: 260481222, 2019
[22]. S. Chandure and V. Inamdar, “Handwritten Modi character recognition using transfer learning with
discriminant feature analysis,” IETE Journal of Research, 2021.
[23]. Dr. R. K. Singh, Dr. Sanjay Kumar Singh, "Handwritten Modi Script Recognition using Transfer Learning",
IJRASET, 2021
[24]. Priyanka Sharma, Dr. S. S. Bhatia, "Modi Script OCR using Neural Networks", IJCST, 2021
[25]. Solley Joseph and Jossy George. “Handwritten character recognition of MODI script using convolutional
neural network based feature extraction method and support vector machine classifier.” ,IEEE 5th
International Conference on Signal and Image Processing (ICSIP), 2020
[26]. Vishal Pawar, Deepak Wadkar, Shivkumar Kashid, Premraj Prakare, Vinayak More, Prof. S.A. Babar,
“Handwritten Character Recognition of MODI Script using Convolutional Neural Network Based Feature
Extraction Method and Support Vector Machine Classifier”, International Research Journal of Engineering
and Technology (IRJET) , 2022
[27]. Sandhya Anpat, Sayali Khetri, Shreejal Dhule, Dr. J.E. Nalavade, “Handwriting Character Recognition using
CNN with GUI”, International Research Journal of Engineering and Technology (IRJET) 2021
[28]. Ajinkya Wani, Unmay Pawar, Yash Gatagat, Dr. Meenakshi Thalor, “Handwritten Character Recognition
Using CNN, KNN and SVM”, International Journal of Technology Engineering Arts Mathematics Science,
2021
[29]. Tesfahun Berhane , Tamiru Melese, Assaye Walelign, and Abdu Mohammed, Hindawi,” A Hybrid
Convolutional Neural Network and Support Vector Machine-Based Credit Card Fraud Detection Model “,
Mathematical Problems in Engineering Volume 2023

Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 458


www.ijarsct.co.in
ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)
International Open-Access, Double-Blind, Peer-Reviewed, Refereed, Multidisciplinary Online Journal
Impact Factor: 7.53 Volume 4, Issue 5, April 2024

[30]. M.O. Khairandish, M. Sharma, V. Jain, J.M. Chatterjee, N.Z. Jhanjhi, “A Hybrid CNN-SVM Threshold
Segmentation Approach for Tumor Detection and Classification of MRI Brain Images”, Elsevier Masson
SAS, 2021
[31]. Prof. Meenakshi Mukhopadhyay, et al, "Modi Script Character Recognition Using Convolutional Neural
Networks", IJIRSET, 2022
[32]. Bhumika Solanki and Maya Ingle. “Performance evaluation of thresholding techniques on modi script”,
International Conference on Advanced Computation and Telecommunication (ICACAT), 2018
[33]. Khandokar, M Hasan Md, F Ernawan, S Islam Md and M N Kabir, “Handwritten character recognition using
convolutional neural network”, Journal of Physics: Conference Series, 2021
[34]. Savita Ahlawat, Amit Choudhary, Anand Nayyar,Saurabh Singh and Byungun Yoon, “Improved Handwritten
Digit Recognition Using Convolutional Neural Networks (CNN)”, Sensors 2020
[35]. Ahlawat, S.; Rishi, R. A genetic algorithm-based feature selection for handwritten digit recognition. Recent
Pat. Comput. Sci. 2019
[36]. Kavitha, B.; Srimathi, C. Benchmarking on offline Handwritten Tamil Character Recognition using
convolutional neural networks. J. King Saud Univ. Comput. Inf. Sci. 2019
[37]. Atman Mishra, A. Sharath Ram, Kavyashree C, “Handwritten Text Recognition Using Convolutional Neural
Network”, arXiv, 2023

Copyright to IJARSCT DOI: 10.48175/IJARSCT-17573 459


www.ijarsct.co.in

You might also like