Journal of Theoretical and Applied Information Technology

15th March 2024. Vol.102. No 5

© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

HANDWRITTEN DIGIT RECOGNITION USING MACHINE

LEARNING

1
NIRASE FATHIMA ABUBACKER, 2MAFAS RAHEEM
1
Assoc. Professor, School of Computing, Asia Pacific University of Technology & Innovation, Malaysia.
2
Senior Lecturer, School of Computing, Asia Pacific University of Technology & Innovation, Malaysia.
E-mail: [Link]@[Link], 2raheem@[Link]

ABSTRACT

The handwritten digit recognition system is a popular research topic, and much research has been done
throughout the years. The implementation of this system will be beneficial for many sectors in today’s world.
Various types of algorithms can be used to develop a solution for this system. However, the accuracy of the
results plays an important role in determining the best solution for the handwritten digit recognition system.
In this project, selected machine learning and deep learning algorithms were used to build models to find the
most suitable model with the best possible accuracy. According to the results, the CNN model performed
better than the other models with an accuracy of 99.25% and 0.99 for each Precision, Recall and F1 Score
compared to all the other models.
Keywords: Digit Recognition, Handwritten, Recognition Model; Machine Learning; Deep Learning

1. INTRODUCTION data analytics algorithms and techniques. Machine

learning is a subset of Artificial Intelligence AI and
Handwritten recognition of characters has computer science. Algorithms are used to improve
been around since the 1980s. According to the the accuracy of the results extracted from data that
Collins dictionary, a digit refers to a written symbol has been collected. A handwritten digit recognition
of any number between 0-9. Digits play a vital role system requires high-accuracy results and reliability.
in the daily routine of life. Industries such as banks, This is important as the system converts handwritten
healthcare, and insurance have a high dependency on digits into machine-readable format. There are many
digits. In a bank, starting from creating an account to challenges in handwritten digit recognition as there
cash withdrawal needs the right digits or numbers. are various styles of handwriting of humans.
Every bank process starts with clients filling up a Additionally, there are different symbols for digits in
form by writing on it. As an important detail, their different languages. However, it’s not simple to find
account number, identification number, or even the right algorithm to build the system. Selectively
phone number are requested. Digits in those forms several machine learning algorithms must be
are then identified by the bank officers manually or researched and trained to find the most accurate
scanned by computers. Whereas in healthcare method. Poor quality of human handwritten digits
industries; patients’ details in forms, doctor’s notes compared to printed copies or typed digits becomes
on patient’s diseases, and even medicine-consuming a hurdle for machines to recognize the digits
guidelines need correct interpretation of the digits. accurately.
Forms that are filled up using handwriting like tax
forms have their focus on numerical entries too. Not This is a huge issue in many industries like
just that, online handwriting recognition on tablets healthcare, insurance and banking, especially where
and zip code recognition for postal mail sorting uses one wrongly interpreted digit from a customer’s
digit [1]. handwriting results in a big problem. For example,
due to poor handwriting, a zero and eight may look
Handwritten digit recognition system is a similar or even digit nine. In this case, a computer
popular research topic among technologists. This system may interpret it as a different digit instead of
system is used to recognize characters, especially the correct digit which ends up wrong. Thus, a
handwritten recognition system requires a more
digits written by humans. This topic involves various
accurate, intelligent and reliable algorithm solution.

2172
 Journal of Theoretical and Applied Information Technology
15th March 2024. Vol.102. No 5
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

There is much research that has been done previously accuracy of handwritten digit recognition. The two
but achieving 100% accuracy seems to be almost authors have used the MNIST dataset for a
impossible. The problem context behind this comprehensive data analysis on the topic. The results
challenge is that even 1% inaccuracy can lead to were tabulated by parameters of the confusion matrix
many wrong interpretations in recognizing the digits. and precision, Recall and F1 score were calculated
Therefore, this paper aims to study and analyse the using that table for all three models built using the
various methods present in building the models and algorithms. As for the observations, KNN relatively
developing the handwritten digit recognition system. had less accuracy compared to SVM and CNN.
Whereas CNN was crowned as the algorithm with
the highest accuracy of 99.4% for the training data
2. LITERATURE REVIEW and 98.4% for the test data. It was also discussed
that; the number of epochs has impacted the accuracy
The research about recognition started with of CNN where more the number of epochs the higher
optical character recognition (OCR) as physicist the accuracy achieved.
Emanuel Goldberg invented a machine that could
convert characters into telegraph code and in the In 2020, research on handwritten digit
1920’s he went a step further by creating the first recognition was done by Savita Ahlawat and Amit
electronic document retrieval system. Later, OCR Choudhary suggested that a combination of
technology proliferated, and businesses started to convolutional neural network (CNN) and support
rely on it to convert data from paper documents [2]. vector machine (SVM) achieved a higher accuracy
Since then, the research on recognition has become of 99.28% for the training data compared to SVM
in-depth and the understanding of algorithms has alone. The proposal stated that handwritten digit
become better. The most suitable research papers are recognition involves automatic feature generation
reviewed in this section as data analytics methods using CNN and the SVM is implemented in
have technologically improved a lot. These research predicting the output. Various parameters such as
papers cover the latest information and results that gamma and degree and decision function are altered
can be helpful for this project. and go through several stages of testing. Finally, the
maximum accuracy was achieved when the gamma
Reference [3] studied the effectiveness of is 0.1, degree 5 and the decision function with one-
some machine learning algorithms in handwritten on-one [5]. Implementation of hybrid algorithms is
digit recognition systems. The paper presented new for the handwritten digit recognition field and
Multilayer Perceptron (MLP), SVM, Naïve Bayes, still in the early stages.
Bayes Net, Random Forest, J48 and Random Tree as
the approaches used for offline handwritten digit The research was conducted on handwritten
recognition. Simulations were done based on digit recognition, particularly for baking systems
accuracy, time consumption and different errors. using the CNN method. The research is about using
Based on these parameters, MLP has been found to CNN to develop an automatic banking deposit
have the highest accuracy with 90.37% overall number recognition system for the cash deposit
compared to other machine learning algorithms. process at the bank counter. The authors stated that
However, this RESEARCH WAS DONE AS AN CNN is the best algorithm to implement for their
INITIAL ATTEMPT WITHOUT USING ANY proposed idea as it has the highest accuracy of all the
standard classification techniques. This research algorithms present [6]. Similarly, another research
focused only on machine learning algorithms and did was done using the CNN approach to analyze the
not attempt to analyse algorithms from other deep learning algorithm in proving the effectiveness
domains of data analytics. of for handwritten digit recognition system. As
shown in Table 1, CNN achieved its maximum
Reference [4] proposed an automatic accuracy of 99.87% in the MNIST dataset hence it is
handwritten digit recognition on document images proved that CNN is a great approach from deep
using SVM, Artificial Neural Network (ANN) and learning algorithm for handwritten digit recognition
CNN. It resulted that, CNN achieved the highest system by the authors [7].
accuracy with 71% performance compared to ANN
and SVM models. Feature extraction and classification are
machine learning technology's two primary
Unlike the other research works, the K- functions. The CNN design does away with the
Nearest Neighbour (KNN) algorithm was analysed necessity for a separate feature extraction method by
along with SVM, and CNN was analysed to study the

2173
 Journal of Theoretical and Applied Information Technology
15th March 2024. Vol.102. No 5
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

combining classification and feature extraction understanding to identify crucial factors for optimal
techniques into a single model [12]. SVM [13] is a handwritten digit recognition. The MNIST dataset,
sophisticated classification method with strong consisting of 60,000 grayscale images of single
generalization capabilities that are based on the digits, underwent data understanding, cleaning, and
principles of structural risk reduction and statistical exploration. Noise reduction techniques were
learning theory through manual feature extraction. applied to enhance image quality. Selected
algorithms were trained and tested with appropriate
Table 1: Comparison of models parameters. Evaluation criteria included accuracy
Algorithms CNN MLP SVM percentage and precision score. The models'
deployment was emphasized for practical use, with
a focus on organizing and presenting knowledge
Accuracy 71.00 [4] 90.37 [3] 98.35 [5] gained for production purposes. The data analysis
(%) 99.87 [7] 39.00 [4] process involved collecting, modelling, and
97.83 [14] analyzing data for decision-making. Data
exploration examined the MNIST dataset's
Parameters GPU: Execution Gamma = structure, revealing 10 rows of digits (0-9) and 785
CPU = time = 0.1
columns in both training and test sets. Data cleaning
30:1 2:32 min
Degree =
confirmed the absence of missing values or outliers.
15000 Number 5 Normalization was performed to scale pixel values
iterations of epochs (0-255) to a range of 0 to 1 and reshape transformed
= 30 C = 1.0, pixel arrays into a (28,28,1) matrix for improved
Lr = 0.01, deep learning model implementation.
Penalty =
Decay = 12, 3.1 Data Visualization
1e-6,
max_iter
Momentu = 1000,
m = 0.9
tol = 1e-4

The review has shown various methods and

approaches. Each approach has different results in
different environments of testing. The application of
machine learning and deep learning shows great
potential for developing a handwritten digit
recognition system with a high level of accuracy.

Traditional handwritten digit recognition

system includes two stages: feature extraction and
classification. Most of them use shallow structures to
deal with computing units and limited kinds of
samples, such as Support Vector Machines (SVM)
[14]. Faced with complex classification problems, Figure 1: Sample of digits
the generalization ability and performance of SVM
are insufficient when the samples have rich Figure 1 shows that there are different types
meanings. Recently, Convolutional Neural Network of writing styles for each digit. For example, number
(CNN) [15] has been widely used in the field of 8 has five different styles within this sample number
computer vision because of its excellent and there could be more styles in the dataset and the
real world.
performance and has made outstanding
achievements in the accuracy of various machine
learning tasks.
3.2 Data Preparation
The ‘label’ column (numbers) is encoded
3. METHODOLOGY into a hot vector for each class and there are 10
classes (0-9). The training data is segmented into two
The study employed the CRISP-DM data exclusives which are a train and validation sets.
mining methodology, starting with business Train data is used for the model training whereas

2174
 Journal of Theoretical and Applied Information Technology
15th March 2024. Vol.102. No 5
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

validation data is used for cross-validation of the metric visualization. It is fitted with 15 epochs and
model’s accuracy and to see how well the model is a batch size of 112, completing training in 20
generalized for the data other than training data. seconds due to its simple architecture. The highest
training accuracy achieved is 97.55%, and
validation accuracy is 97.17%. Training loss
4. MODELLING & ANALYSIS gradually decreases, while validation loss varies
4.1 Convolutional Neural Network (CNN) across epochs.
A CNN model was constructed using the
LeNet-5 architecture. The model featured two 4.3 Support Vector Machine (SVM)
Conv2D layers with 32 and 64 filters, respectively. The SVM model-building process began
The kernel size was set to 5x5 for the first layer and with both linear and non-linear models using default
reduced to 3x3 for the second layer. Rectified linear hyperparameters for comparison. The linear SVM
unit (ReLU) activation, same padding, and model yielded an accuracy of approximately
MaxPool2D with a (2,2) pool size and strides were 90.56%, as shown in the confusion matrix. The non-
applied. A Dropout layer with a 25% dropout rate linear SVM model, employing the 'rbf' kernel with
was included. The dense layer had a value of 255, default gamma and C values, demonstrated an
representing the maximum pixel value in the training increased accuracy of 94.39%.
dataset. The final layer was a softmax classifier
related to cross-entropy loss. The RMSProp Considering the higher accuracy of the non-
optimizer with default values for learning rate, rho, linear model, hyperparameter tuning was performed
epsilon, and decay was employed for effective and specifically for it. The optimization focused on
faster model convergence. Learning rate annealing finding the optimal C and gamma values. C controls
was implemented using ReduceLROnPlateau to error, while gamma influences the curvature in the
monitor and adjust the learning rate during training non-linear SVM model. A 5-fold cross-validation
epochs, contributing to achieving high accuracy. The with grid search was employed for this purpose,
training utilized CUDA to enhance speed, with 15 taking nearly 40-50 minutes. The results indicated
epochs and a batch size of 112. The training process that at higher gamma values (0.01), the model
took approximately 40 minutes, revealing that the exhibited overfitting, with 100% training accuracy
training accuracy consistently surpassed the but less than 80% test accuracy. The analysis
validation accuracy. The training and validation revealed that a gamma value of 0.001, combined
losses exhibited minimal differences. The model with C=15, provided the highest test accuracy
achieved the highest validation accuracy of 99.32% around 94% while avoiding overfitting.
and a training accuracy of 99.5%. The final hyperparameters selected were C
= 10 and gamma = 0.001, and the SVM model was
fitted accordingly.
4.2 Multilayer Perceptron (MLP)
The Multilayer Perceptron (MLP) model
has a simple architecture with four sequential layers 5. RESULTS AND DISCUSSION
in the Keras model. The first layer is a flattened
layer that converts a 2D image matrix into a 1D A model’s performance is evaluated based
vector to suit MLP's requirement for 1D input. The on different metrics such as accuracy, loss, error
second layer is a dense layer with 128 neurons and value and precision score. The different evaluation
ReLU activation, facilitating quick convergence. metrics will allow data analysts to understand a
The third hidden layer mirrors the second in terms model’s performance based on its strengths and
of units and activation. The fourth layer is the output weaknesses. For this project, three models, CNN,
layer with softmax activation, representing the MLP and SVM will be evaluated based on certain
network's guess for each digit. To control metrics and a decision will be made to choose the
underfitting or overfitting, a kernel_regularizer best algorithm for handwritten digit recognition.
parameter of 0.002 is applied, like dropout in the
CNN model, with 10 units in this layer.
5.1 CNN
The model summary provides an overview The Convolutional Neural Network (CNN)
of the layers, their order, output shapes, parameters model achieved an accuracy of approximately
per layer, and the total model parameters. The 99.31% with a loss of 0.0326 in the validation test.
optimizer used is the same as in the CNN model. This high accuracy indicates the model's strong
The MLP model is compiled with defined predictive capabilities.
parameters, including a TensorBoard callback for

2175
 Journal of Theoretical and Applied Information Technology
15th March 2024. Vol.102. No 5
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

Figure 2: Loss graph - CNN

Figure 2 illustrates the loss for the training

and validation sets over epochs, showing an initial Figure 4: Classification Report - CNN
disparity that converges as epochs progress, with a
slight difference at the end where the validation set
loss is slightly higher. Despite this, the model
exhibits a good fit, as the loss consistently decreases
with each epoch.

Figure 5: Prediction - CNN

The predictions are illustrated in Figure 5,

Figure 3: Accuracy graph - CNN
which reveals instances where predicted values
Figure 3 displays the accuracy for both deviate from true values. A specific example with
sets, indicating that training accuracy surpasses index number 2853 from the test data is highlighted,
validation accuracy due to the model's familiarity showcasing an accurate prediction by the CNN
with the training data. model.

Metrics such as recall and F1 score further 5.2 MLP

The provided text discusses the
validate the model's positive performance. Despite
performance evaluation of a Multilayer Perceptron
not achieving 100% accuracy, the CNN model is
(MLP) model based on certain figures. The model
effective in predicting most images as per the achieved an accuracy of approximately 97% and a
classification report shown in Figure 4. loss of 0.1767 on the validation test. The text
emphasizes the importance of validation accuracy in
assessing the model's predictive capability and
suggests that hyperparameters can be optimized for
better results.

2176
 Journal of Theoretical and Applied Information Technology
15th March 2024. Vol.102. No 5
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

Figure 6: Loss graph - MLP

Figure 8: Classification report - MLP

While the MLP model exhibits high

accuracy, the text acknowledges that no model
achieves 100% accuracy. For practical purposes,
important errors are examined, revealing instances
where predicted values deviate from true values.

Figure 7: Accuracy graph - MLP

Figure 6 and Figure 7 illustrate the model's

loss and accuracy for both training and validation
sets across epochs respectively. The text notes an
initial disparity in loss scores between the sets, which
converges as epochs progress, with a slight
divergence towards the end. Despite this, the model
is deemed to have a good fit, given the consistent
Figure 9: Prediction - MLP
decrease in loss scores.
Figure 8 demonstrates accurate predictions Figure 9 specifically shows accurate
for all classes, indicating good model performance. predictions for test data using the MLP model.
The classification report further assesses precision, Overall, the model is considered to perform well,
recall, and F1-score, highlighting high scores across with detailed evaluations of accuracy, precision,
most classes, with minor errors in precision for one recall, and important errors.
class (class 7) as shown in Fig. 9.

5.3 SVM
The Figure 10 presents a classification
report depicting the accuracy results of an SVM
model. This model was constructed using optimal
hyperparameters, specifically C=10 and
gamma=0.001, resulting in a maximum accuracy of
95.31%. Diagram 56 illustrates precision scores
ranging from 0.92 to 0.98 for different classes, while
recall scores fall within the range of 0.93 to 0.98.
Additionally, the f1 scores span from 0.95 to 0.98.

2177
 Journal of Theoretical and Applied Information Technology
15th March 2024. Vol.102. No 5
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

These scores collectively indicate a high level of Table 3, which likely summarizes previous research
performance for the model, with values considered on handwritten digit recognition using CNN models.
to be quite good. The accuracies reported in these studies range from
98% to 99.87%. However, it notes that these studies
may lack clear evaluation steps, implying that the
methodologies used to assess model performance
might not have been sufficiently rigorous or
transparent. The CNN model developed in the
current project achieves a notably higher accuracy. It
comes close to the highest recorded accuracy of
99.31% with only a marginal difference of 0.06%.
This implies that the model developed in this project
performs exceptionally well in recognizing
handwritten digits.

Importantly, the excerpt highlights that the

current project's CNN model was evaluated using
rigorous cross-validation steps to eliminate
Figure 10: Confusion matrix - SVM overfitting. Cross-validation is a technique used to
assess how well a model will generalize to an
independent dataset. By employing this technique,
5.4 Model Comparison the project ensures that the reported accuracy is
Table 2 reveals that the CNN model reliable and not inflated due to overfitting. The
outperformed other models in terms of accuracy, comparison suggests that the CNN model developed
achieving 99.31%. Despite taking 30 minutes for in the current project outperforms other existing
training, it demonstrated superior performance. In models in handwritten digit recognition. Despite the
comparison, the MLP model completed training in marginal difference in accuracy, the rigorous
just 20 seconds with 97% accuracy. The suggestion evaluation process enhances confidence in the
is that, with improved hyperparameter tuning, both performance of the current model. This indicates that
CNN and MLP could potentially achieve even the project's model may represent a significant
higher accuracies. On the other hand, the SVM advancement in the field of handwritten digit
model exhibited the lowest accuracy among the recognition.
deep learning models and took 40 minutes for
training. In conclusion, the CNN model is deemed Overall, the excerpt provides insights into
the most suitable algorithm for a handwritten digit the performance of CNN models in handwritten digit
recognition system. recognition tasks, emphasizing the importance of
rigorous evaluation methods and highlighting the
Table 2: Model comparison superior performance of the CNN model developed
Model / Model Comparison Summary in the current project.
Metrics CNN MLP SVM
Test 99.31% 96.89% 95.31% Table 3: Comparison of CNN results with existing
Accuracy studies
Precision 0.99 0.97 0.95 Test
(avg) Research Paper Accuracy
Recall (avg) 0.99 0.97 0.95 (%)
F1 score 0.99 0.97 0.95 Hybrid CNN-SVM Classifier for
(avg) Handwritten Digit Recognition (CNN- 99.28
Training 30 min 20 sec 40 min SVM) [5]
time Handwritten Digit Recognition Using
99.87
Deep Learning (CNN) [7]
Handwritten Digit Recognition System
This excerpt discusses the performance of using Convolutional Neural Network 98.00
convolutional neural network (CNN) models in (CNN) [8]
handwritten digit recognition tasks, comparing them Implementation of Handwritten Digit
to existing studies and highlighting the accuracy 99.15
Recognizer using CNN [9]
achieved in a current project. The excerpt mentions

2178
 Journal of Theoretical and Applied Information Technology
15th March 2024. Vol.102. No 5
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

Test was able to extract valuable insights and best

Research Paper Accuracy practices. These insights likely informed decisions
(%) regarding parameter tuning, feature selection, model
Recognition of Handwritten Digit using architecture, and other crucial aspects of the model-
Convolutional Neural Network (CNN) 99.15 building process. Ultimately, the project's approach
[10] underscores the importance of marrying theoretical
Handwritten Digit Recognition using understanding with empirical evidence to iteratively
Machine and Deep Learning Algorithms 99.31 refine and enhance machine learning models.
(CNN) [11]
Hand Written Digit Recognition Using The accuracy of the models significantly
95.00 impacts the speed of digit recognition in the era of
Machine Learning (CNN) [16]
Hand Written Digit Recognition Using electronic data. Despite the commendable
98.40 achievements, there is still room for enhancement,
Machine Learning (CNN) [17]
This study (CNN) with proper particularly through fine-tuning hyperparameters
99.31 such as learning rate, optimizers, model depth, and
validation
epochs. Future research, especially focused on CNN,
holds the potential for achieving nearly 100%
accuracy, with a specific emphasis on model and
6. CONCLUSION
pixel depth tuning for further development.
The conclusion drawn from the project is
that Convolutional Neural Networks (CNNs) have
proven to be the most efficient and accurate model
for the task of recognizing handwritten digits. REFERENCES
Through rigorous cross-validation procedures, a
suitable architecture for the CNN model was [1] G. Jain, and J. Ko, “Handwrittien Digits
identified, resulting in an impressive accuracy rate of Recognition”, Multimedia Systems|Project
99.25%. This accuracy rate signifies the Report, 2008, pp. 1-14.
effectiveness of the CNN model in accurately [2] S. Britton, “Optical character recognition
identifying and classifying handwritten digits,
(OCR) technology: A brief history”, 2022,
showcasing its potential for various applications in
[Link]
fields such as character recognition, optical character
recognition (OCR), and digitized document opinion/optical-character-recognition-ocr-
processing. The thoroughness of the cross-validation technology-a-brief-history/ (Accessed 2023
steps ensures that the chosen model architecture is November 10)
robust and reliable, offering confidence in its [3] S. M. Shamim, M. B. A. Miah, A. Sarker, M.
performance across different datasets and real-world Rana, and A. Jobair, “Handwritten Digit
scenarios. Therefore, the project's findings highlight Recognition using Machine Learning
the superiority of CNNs for handwritten digit Algorithms”, Indonesian Journal of Science
recognition tasks and emphasize their significance in and Technology, Vol. 3, No. 1, 2018, pp. 29-
advancing machine learning technologies for image 39.
classification tasks. [4] A. Challa, “Automatic Handwritten Digit
The project embarked on an exploration of Recognition On Document Images Using
Multilayer Perceptron (MLP) and Support Vector Machine Learning Methods”, 2019, pp. 50.
Machine (SVM) algorithms, recognizing their [5] A. Savita, and C. Amit, “Hybrid CNN-SVM
potential but also acknowledging the necessity for Classifier for Handwritten Digit Recognition”,
refinement. This recognition underscores the Procedia Computer Science, Vol. 167, 2020,
importance of comprehending the nuanced pp. 2554-2560.
capabilities and limitations inherent in each
algorithm. It underscores the significance of crafting [6] V. Gopalakrishan, R. Arun, L. Sasikumar, and
bespoke solutions that align with the unique K. Abhirami, “Handwritten Digit Recognition
attributes of the problem at hand. In the pursuit of for Banking System”, International Journal of
refining these algorithms, the project leaned heavily Engineering Research & Technology (IJERT),
on a foundation of extensive research. This research 2021, pp. 313-314.
encompassed a wide array of sources, including [7] J. S. Gaganashree, and D. K. Padmashali,
scholarly journal articles, authoritative books, and “Handwritten Digit Recognition Using Deep
pertinent online resources. This comprehensive Learning”, International Journal of Research
approach to gathering insights played a pivotal role in Engineering, Science and Management, Vol.
in shaping the process of model construction. By 4, No. 7, 2021, pp. 182-185.
delving into a diverse range of literature, the project

2179
 Journal of Theoretical and Applied Information Technology
15th March 2024. Vol.102. No 5
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

[8] K. Y. Goh, and A. S. Ghafar, “Handwritten

Digit Recognition System using Convolutional
Neural Network (CNN)”, PEAT, Vol. 2, No. 1,
2021, pp. 578-592.
[9] B. M. Vinjit, M. K. Bhojak, S. Kumar, and G.
Nikam, “Implementation of Handwritten Digit
Recognizer using CNN”, Workshop on
Advances in Computational Intelligence, 2021.
[10] M. M. Ali, and M. A. Hossain, “Recognition of
Handwritten Digit using Convolutional Neural
Network (CNN)”, Global Journal of Computer
Science and Technology: Neural & Artificial
Intelligence, Vol. 19, No. 2, 2019, pp. 27-33.
[11] D. Ritik, K. Rishika and P. Samay,
“Handwritten Digit Recognition using Machine
and Deep Learning Algorithm”, International
Journal of Computer Applications, Vol. 176,
No. 42, 2021, pp. 27-33.
[12] P. Preethia, R. Asokanb, N. Thillaiarasuc and T.
Saravanan, “An effective digit recognition
model using enhanced convolutional neural
network based chaotic grey wolf optimization”,
Journal of Intelligent & Fuzzy Systems, Vol. 41,
2021, pp. 3727-3737.
[13] C. Li, C. Ting-Xuan, X. Ying-Qing Xia, Xia
Cao, and Ling Wu, “Ensemble Learning
Network for Handwritten Digit Recognition
Based on Fusion Optimized CNN”, Journal of
Computers, Vol. 34, No. 3, 2023, pp. 137-150.
[14] M. H. Ahamed, S. M. I. Alam, and M. Islam,
“SVM Based Real Time Hand-Written Digit
Recognition System”, International
Conference on Engineering Research and
Education, 2019.
[15] R. Payel, C. Somnath, H. Chayan, S.
Shibaprasad, S. M. Obaidullah, and R. Kaushik,
“Comparative study on the performance of the
state-of-the-art CNN models for handwritten
Bangla character recognition”, Multimedia
Tools and Applications, 2023, Vol. 82,
pp.16929–16950.
[16] A. V. S. R. Praneeth, L. R. Pappu, G. P. Kumar,
A. S. Bhatta, B. R. Chowdary, and G. S.
Shishwan, “Hand Written Digit Recognition
Using Machine Learning”, International
Journal of Research in Engineering, Science
and Management, Vol. 5, No. 1, 2022, pp.59 -
62.
[17] M. Sharma, P. S. Sindal, and M. Baskar, “Hand
Written Digit Recognition Using Machine
Learning”, Proceedings of Data Analytics and
Management. Lecture Notes in Networks and
Systems, Vol. 572, 2023, pp.31 - 43.

2180