A

Synopsis

on

Handwritten Digit Recognition

Using CNN

in partial fulfillment of the requirement for the degree

of

Bachelor of Technology
In
COMPUTER SCIENCE AND ENGINEERING
(Artificial Intelligence & Machine Learning)

Submitted by
Prasoon Bhatt (2101331530095)
Shashank Shekhar (2101331530118)

Under the supervision of

Ms. Meghali Das
(Assistant Professor)

NOIDA INSTITUTE OF ENGINEERING AND TECHNOLOGY

GREATER NOIDA
 Index

S. No. Topics Page No.

1. Introduction
2. Existing Systems
3. Problem Statement
4. Proposed Methodology
5. Feasibility Study
6. Facilities Required
7. Proposed System
8. Software Design Diagram
9. Conclusion
10. References

Supervisor Sign:
(Ms. Meghali Das)
 Introduction

Handwritten digit recognition is an essential task in the field of computer

vision and artificial intelligence, where the objective is to automatically
identify numerical digits from images. This technology has been widely
used in applications such as postal mail sorting, bank check processing, and
form digitization. Traditional methods for digit recognition relied on
handcrafted feature extraction and machine learning algorithms such as
Support Vector Machines (SVM) and K-Nearest Neighbors (KNN).
However, with the advancement of deep learning, Convolutional Neural
Networks (CNNs) have become the preferred approach due to their
superior accuracy and ability to learn hierarchical features directly from
images.

This project focuses on building a CNN-based model to recognize

handwritten digits using the MNIST dataset, a benchmark dataset in the
field of image classification. The MNIST dataset consists of 60,000
training images and 10,000 test images, each representing a digit from 0 to
9 in grayscale format with a resolution of 28×28 pixels. Unlike traditional
feature-based approaches, CNNs automatically extract spatial and
hierarchical features from the images, significantly improving recognition
performance.

The project follows a structured approach to developing the digit

recognition system:

Data Preprocessing: The dataset is first normalized by scaling pixel

values to a range of 0-1 to improve training efficiency. Additionally,
images are reshaped into a format suitable for input into the CNN model.

CNN Model Development: A deep learning model is designed with

convolutional layers for feature extraction, pooling layers for
dimensionality reduction, and fully connected layers for classification.
Training the Model: The model is trained using supervised learning,
where labeled images help the neural network learn distinguishing patterns
for each digit. Techniques such as data augmentation and dropout are used
to prevent overfitting.

Model Evaluation: The trained model is tested on unseen data to assess its
accuracy. Performance metrics such as accuracy, precision, recall, and
confusion matrix are used for evaluation.

Prediction and Real-World Application: Once trained, the model can be

deployed in real-world scenarios, such as recognizing handwritten digits in
scanned documents or mobile applications.

CNNs are particularly powerful for digit recognition because they preserve
spatial relationships within the image and can recognize digits even when
they vary in size, rotation, or handwriting style.

With advancements in deep learning, CNN-based digit recognition models

now achieve over 99% accuracy on the MNIST dataset, making them
highly reliable for practical applications.

This project demonstrates the effectiveness of CNNs in image classification

tasks and highlights the potential of deep learning in automating digit
recognition with minimal manual intervention.
 Existing Systems

Handwritten digit recognition has evolved significantly over the years,

from simple rule-based methods to sophisticated deep-learning models.
Traditional machine learning techniques relied heavily on handcrafted
features and statistical models, while modern approaches leverage deep
learning to automatically extract patterns and improve accuracy. Below is a
detailed theoretical comparison of the existing approaches.

1. Traditional Machine Learning Approaches

a) Template Matching
Template Matching is one of the earliest approaches used for handwritten
digit recognition. It works by comparing an input image with a set of
predefined templates representing each digit. The closest match is selected
as the predicted digit.
Advantages:
 Simple and easy to implement.
 Works well when the handwriting style is uniform.
Disadvantages:
 Fails when there are variations in handwriting.
 Computationally expensive for large datasets.

b) K-Nearest Neighbours (KNN)

KNN is a non-parametric classification algorithm that classifies an input
digit based on its similarity to the k-nearest training samples. The most
common digit among the neighbours determines the classification.
Advantages:
 Simple and effective for small datasets.
 No training is required.
Disadvantages:
 Computationally expensive for large datasets.
 Performance degrades with high-dimensional data.
c) Support Vector Machines (SVM)
SVM is a supervised learning model that separates digit classes using a
hyperplane in a high-dimensional space. It works well with small to
medium-sized datasets and requires feature extraction before classification.
Advantages:
 Works well with linearly separable data.
 Provides good accuracy with proper feature engineering.
Disadvantages:
 Requires manual feature extraction.
 Computationally intensive for large datasets.

2. Deep Learning Approaches

a) Artificial Neural Networks (ANN)

ANN consists of multiple layers of neurons, where each layer extracts and
processes features from input images. However, ANNs have limitations in
recognizing spatial relationships within images.
Advantages:
 Can learn from data without manual feature extraction.
 Performs better than traditional machine learning techniques.
Disadvantages:
 Struggles with spatial relationships in images.
 Requires a large number of training samples.

b) Convolutional Neural Networks (CNN)

CNNs are specifically designed for image-based tasks and have become the
state-of-the-art method for handwritten digit recognition. A CNN consists
of convolutional layers, pooling layers, and fully connected layers that
automatically extract spatial features and classify digits effectively.
Advantages:
 Automatically learns features without manual intervention.
 Achieves high accuracy (>99%) on digit recognition tasks.
 Robust to variations in handwriting style.
Disadvantages:
 Requires a large amount of labelled training data.
 Computationally expensive compared to traditional methods.
c) Recurrent Neural Networks (RNN) & Long Short-Term Memory
(LSTM)
RNNs and LSTMs are commonly used for sequential data, making them
effective for recognizing handwritten text rather than individual digits.

Advantages:
 Effective for recognizing sequences of handwritten characters or
words.
 Can learn temporal dependencies in handwriting.
Disadvantages:
 Not ideal for single-digit classification.
 Requires high computational resources.

3. Comparison of Different Approaches

Approach Description Advantages Disadvantages

Template Compares the input Simple to
Fails with variations in
Matching digit with stored implement; works
handwriting;
templates of digits. well with fixed fonts.
computationally
expensive.
K-Nearest Classifies digits based Non-parametric; Slow for large datasets;
Neighbors (KNN) on their similarity to works well with memory-intensive.
stored samples. small datasets.

Support Vector Uses hyperplanes to Works well with Requires manual

Machines (SVM) separate digit classes small datasets and feature extraction; not
in a high-dimensional well-defined scalable for complex
space. features. data.
Artificial Neural A simple multi-layer Can learn features Limited ability to
Networks (ANN) perceptron (MLP) from data; better capture spatial features;
model forthan traditional not ideal for image-
classification. methods. based tasks.
Convolutional Deep learning model Achieves high Requires large amounts
Neural Networks with convolutionalaccuracy; effective of training data and
(CNN) layers that extract for large datasets; computational power.
features automatically.learns hierarchical
features.
Recurrent Neural Captures sequential Useful for Computationally
Networks (RNN) dependencies in handwritten text expensive; not ideal for
& LSTMs handwritten text. recognition rather single-digit
than isolated digits. classification.
4. Why CNNs are the Best Choice?

Among all the existing approaches, CNNs provide the best results for
handwritten digit recognition due to the following reasons:

Feature Extraction Automation: Unlike traditional methods, CNNs

automatically learn and extract relevant features from images, making them
more efficient.
High Accuracy: CNN-based models achieve more than 99% accuracy on
the MNIST dataset.
Robust to Variations: CNNs can handle differences in handwriting styles,
rotation, and noise in images.
Efficient for Large Datasets: CNNs scale well with large amounts of
training data, making them suitable for real-world applications.

Hence, CNNs have become the industry standard for tasks like handwritten
digit recognition, license plate recognition, and even facial recognition.
 Problem Statement

Handwritten digit recognition is a crucial task in the field of pattern

recognition and artificial intelligence, with applications in postal mail
sorting, bank check processing, and automated form reading. Traditional
methods, such as template matching and statistical classifiers, struggle with
variations in handwriting styles, distortions, and noise in images.

This project aims to develop a robust and efficient Convolutional Neural

Network (CNN) model for recognizing handwritten digits from images.
The CNN will automatically extract spatial hierarchies of features, making
it more accurate than traditional machine learning approaches. The system
will be trained on a dataset of labelled handwritten digits (such as the
MNIST dataset) and will be evaluated based on its classification accuracy.

Recognizing handwritten digits presents several challenges that must be

addressed to build an accurate and efficient system:

Variability in Handwriting Styles: Different people write the same digit

in unique ways, making it difficult to create a universal recognition model.

Distortions and Noise: Handwritten digits may be distorted due to poor

handwriting, ink smudges, or low-quality image acquisition methods like
scanning or camera capture.

Overlapping and Connected Digits: In some scenarios, digits may be

joined or overlapped, making segmentation and classification more
difficult.

Feature Extraction Complexity: Traditional methods require manual

feature extraction, which is time-consuming and may not generalize well
across different datasets.

Real-Time Processing Requirements: Applications such as postal

services and bank check verification require high-speed processing without
compromising accuracy.
Scalability for Large Datasets – Efficient training and inference on large-
scale datasets are crucial for real-world deployment.

Multilingual and Multi-Script Recognition – Expanding the model to

recognize handwritten characters from different languages and scripts for
broader usability.
Robustness Against Adversarial Attacks – Ensuring the model is
resistant to manipulated inputs that could deceive recognition accuracy.

Energy-Efficient AI Processing – Optimizing the model for low-power

edge devices to enable offline recognition in embedded systems.

Adaptive Learning for New Handwriting Styles – Implementing

techniques such as transfer learning and continual learning to improve
performance across diverse handwriting samples.

CNN-based models effectively tackle these challenges by learning complex

hierarchical patterns directly from image data, eliminating the need for
manual feature selection by creating a robust and efficient handwritten digit
recognition system that can be applied to automated form processing,
financial transactions, and other digit-based classification tasks.
 Proposed Methodology

The proposed methodology involves developing a Convolutional Neural

Network (CNN)-based model to accurately classify handwritten digits. The
approach consists of several key steps, including data preprocessing, model
design, training, evaluation, and deployment.

1. Data Collection and Preprocessin

Dataset Selection: The model will be trained on the MNIST dataset, which
contains 60,000 training images and 10,000 test images of handwritten
digits (0-9).
Data Normalization: The pixel values (0-255) are normalized to a range
of 0 to 1 to improve the training efficiency.
Reshaping: The dataset images (28x28 pixels) are reshaped to fit the CNN
input layer.
Augmentation (if needed): Techniques like rotation, shifting, and
zooming may be applied to increase data diversity and improve model
generalization.

2. CNN Model Design

Convolutional Layers: Extract spatial features (edges, curves, and shapes)

from the input image using multiple filters.
Pooling Layers: Reduce dimensionality while retaining essential
information to improve computational efficiency.
Flattening Layer: Converts feature maps into a 1D vector for
classification.
Fully Connected Layers: Classifies digits based on extracted features.
SoftMax Activation Layer: Assigns probability scores to each digit (0-9)
and selects the most probable class.

3. Model Training

Loss Function: Categorical Cross-Entropy is used to measure the

difference between predicted and actual labels.
Optimization Algorithm: Adam optimizer is used for efficient weight
updates.
Batch Processing: Training is done in mini-batches to improve
convergence speed.
Validation Set: A subset of the dataset is used to monitor model
performance and prevent overfitting.

4. Model Evaluation

Accuracy: Measures the percentage of correctly classified digits.

Precision and Recall: Analyses how well the model distinguishes between
similar digits.
Confusion Matrix: Visualizes correct and incorrect predictions.

5. Model Deployment and Applications

Real-Time Classification: The trained model can be integrated into an

application for digit recognition in banking, postal services, and document
automation.
Integration with Optical Character Recognition (OCR): The CNN
model can be combined with OCR systems for handwritten text
recognition.
Edge Deployment: The model can be optimized for low-power devices for
mobile or embedded applications.

6. Future Enhancements

Transfer Learning: Using pre-trained deep learning models to improve

performance.
Advanced Architectures: Exploring deeper CNN models or hybrid
approaches (CNN + Recurrent Neural Networks) for sequence-based
handwriting recognition.
Custom Datasets: Training the model on diverse datasets beyond MNIST
to handle real-world handwriting variations.
Use Case Diagram

A Use Case Diagram is a graphical representation used in Unified

Modelling Language (UML) to depict the interactions between a system
and its actors (users or external systems). It describes functional
requirements by showing different use cases (functionalities) and their
relationships with actors.
ER Diagram

An Entity-Relationship (ER) Diagram is a graphical representation of a

database structure, showing how entities (objects) relate to each other in a
system. It is widely used in database design to visualize data relationships,
attributes, and constraints.
Data Flow Diagrams (0 level)

A Level 0 Data Flow Diagram (DFD), also known as a Context Diagram, is

a high-level graphical representation of a system that shows how it interacts
with external entities (users, systems, or databases). It depicts the entire
system as a single process, with inputs and outputs flowing between the
system and its external entities.
Data Flow Diagrams (1 level)

A Level 1 Data Flow Diagram (DFD) provides a detailed breakdown of the

main process from Level 0 DFD. It decomposes the single high-level
process into multiple sub-processes, showing data flow between them,
external entities, and data stores. This level helps in understanding internal
functionalities while maintaining an overall system perspective.
Flow Charts

A Flowchart is a graphical representation of the step-by-step workflow of a

system, process, or algorithm using different symbols. It visually illustrates
the sequence of operations in the Handwritten Digit Recognition System,
from data input (handwritten digit) to model processing and final
prediction. Flowcharts help in understanding, designing, and debugging the
system logic efficiently.
 Feasibility Study

1. Feasibility of the Project

The feasibility of this project is evaluated based on its technical, economic,

operational, and legal aspects, ensuring its practicality and success.

Technical Feasibility: The project is viable with modern deep learning

frameworks like TensorFlow/Keras and publicly available datasets like
MNIST. With GPU acceleration, training and deployment are efficient.
Economic Feasibility: The project requires minimal investment, using
open-source tools and datasets. Cloud-based deployment can reduce
hardware costs, making it cost-effective for businesses.
Operational Feasibility: The system is easy to integrate into real-world
applications such as banking, postal automation, and OCR systems. It
enhances efficiency and reduces manual errors.
Legal & Ethical Feasibility: The project follows data privacy guidelines
by using public datasets. It can be made bias-free by training on diverse
handwriting styles.

2. Need for Handwritten Digit Recognition

Automation & Efficiency: Many industries, such as finance and postal

services, rely on handwritten data. Automating digit recognition reduces
manual effort and speeds up processing.
Error Reduction: Manual data entry is prone to human errors. A CNN-
based model improves accuracy, ensuring reliable data extraction.
Scalability: Traditional methods struggle with handwriting variations.
CNN models adapt and scale for different handwriting styles and datasets.
Real-Time Processing: Applications such as automated bank check
verification and digital form processing require fast recognition, which
deep learning enables.
3. Significance of the Project

Automation of Manual Processes – Reduces human effort in digit

recognition tasks such as postal sorting, bank check processing, and
automated form reading.
Enhanced Accuracy and Efficiency – Uses Convolutional Neural
Networks (CNNs) to improve recognition accuracy compared to traditional
methods.
Real-World Applications – Beneficial for financial institutions,
healthcare, and education by enabling efficient data entry and verification.
Advancement in AI and Machine Learning – Contributes to the growth of
AI-powered recognition systems and computer vision applications.
Scalability and Adaptability – Can be extended to recognize different
languages, symbols, and handwritten text styles for broader usability.
Time and Cost Efficiency – Reduces the need for manual verification,
leading to faster processing times and lower operational costs.
Improved User Experience – Enables seamless interaction with
touchscreen devices, smart applications, and assistive technologies.
Security and Fraud Detection – Helps in detecting forged handwriting or
unauthorized document alterations in sensitive sectors like banking.
Data-Driven Insights – Collecting and analysing digit recognition data can
enhance AI models and improve predictive analytics.
Support for Digital Transformation – Encourages the adoption of
paperless processes, reducing reliance on physical documents and
promoting digitization.
 Facilities required

1. Software Requirements

Programming Language: Python (Recommended: Version 3.x)

Deep Learning Framework: TensorFlow / Keras
Libraries & Dependencies: NumPy, OpenCV, Matplotlib, Pandas, Scikit-
learn
Dataset: MNIST (or other handwritten digit datasets)
Development Environment: Jupyter Notebook, Google Colab, PyCharm,
VS Code
Operating System: Windows, Linux, or macOS

2. Hardware Requirements

Processor (CPU): Minimum Intel i3 or equivalent, Recommended Intel

i7 / Ryzen 7 or better
Graphics Card (GPU): Minimum Integrated Graphics, Recommended
NVIDIA GPU (e.g., GTX 1650, RTX 3060)
RAM: Minimum 4GB, Recommended 8GB or more
Storage: Minimum 20GB free space, Recommended SSD for faster
processing.

3. Additional Requirements for Deployment

Cloud Services : Google Colab, AWS, or Azure for cloud-based training

and deployment
Edge Deployment : Raspberry Pi or Jetson Nano for running the model on
embedded systems

This setup ensures efficient model training, testing, and deployment for
handwritten digit recognition.
 Proposed System

The proposed system aims to develop an automated handwritten digit

recognition model using Convolutional Neural Networks (CNNs). This
deep learning-based approach effectively handles variations in handwriting,
distortions, and noise, ensuring high accuracy in digit classification.

A. Functional Requirements

The functional requirements define the key operations and features that the
Handwritten Digit Recognition System must support.

1. Data Processing & Preprocessing

 The system should accept handwritten digit images as input.

 It should convert images to grayscale and normalize pixel values for
uniform processing.
 Noise removal and image resizing should be performed to enhance
recognition accuracy.

2. Model Training & Learning

 The system should use a Convolutional Neural Network (CNN) for

feature extraction and classification.
 The model should be trained using a labelled dataset (e.g., MNIST) to
learn digit patterns.
 It should apply data augmentation (rotation, scaling, flipping) to
improve generalization.

3. Handwritten Digit Recognition

 The system should classify input images into digits (0-9) based on
trained features.
 It should return a confidence score (probability) for each prediction.
 Real-time classification should be supported for fast processing of new
inputs.
4. Performance Evaluation

 The system should calculate accuracy, precision, recall, and F1-score

for model evaluation.
 It should allow comparison with other models (e.g., SVM, Random
Forest) for benchmarking.

5. User Interface & Interaction

 A graphical user interface (GUI) or API should be provided for users

to upload and test digit images.
 The system should display the recognized digit and confidence score
after processing.

6. Deployment & Integration

 The trained model should be deployable on cloud, web, or mobile

platforms.
 The system should be scalable to handle a large number of digit
recognition requests.
 It should support integration with OCR systems or IoT devices for
real-world applications.

B. Non-Functional Requirements

Non-functional requirements define the quality attributes and performance

expectations of the Handwritten Digit Recognition System beyond its
core functionalities.

1. Performance Requirements

 The system should achieve at least 98% accuracy on the MNIST

dataset.
 Predictions should be made in real-time (within milliseconds) for quick
response.
 The model should be optimized for low latency and minimal
processing overhead.
2. Scalability

 The system should handle large datasets without performance

degradation.
 It should be capable of scaling across cloud platforms, edge devices,
and web applications.

3. Security

 User-uploaded images should be processed securely without

unauthorized access.
 The system should prevent data leaks and model tampering by
implementing authentication and encryption mechanisms.

4. Usability

 The user interface (UI) should be simple and intuitive, allowing easy
digit uploads and recognition.
 The system should provide clear output with confidence scores for
better interpretability.

5. Maintainability & Upgradability

 The system should support model retraining with new datasets to

improve accuracy.
 The codebase should be modular, well-documented, and easy to update
with minimal effort.

6. Compatibility

 The system should be compatible with Windows, Linux, and macOS.

 It should support deployment on cloud platforms (AWS, Google
Cloud, Azure) and embedded devices (Raspberry Pi, Jetson Nano,
etc.).
7. Reliability & Availability

 The model should be highly reliable, ensuring minimal incorrect

predictions.
 The system should have backup mechanisms to restore previous
versions in case of failure.

8. Efficiency

 The model should be optimized to use minimal computational

resources while maintaining accuracy.
 It should efficiently utilize GPU/TPU acceleration for faster
processing.

9. Legal & Ethical Considerations

 The system should comply with data privacy laws (GDPR, CCPA)
when handling user data.
 It should ensure fairness and bias mitigation in digit recognition to
avoid discrimination against specific handwriting styles.

C. Major Modules and Their Functionalities

The Handwritten Digit Recognition System is divided into several key

modules, each handling a specific aspect of digit classification.

1. Data Preprocessing Module

 Loads handwritten digit images.

 Converts images to grayscale and resizes them to a standard dimension
(e.g., 28×28 pixels).
 Normalizes pixel values for improved model training.
 Removes noise and enhances image quality.
2. Feature Extraction Module

 Uses Convolutional Neural Networks (CNNs) to automatically extract

important features.
 Identifies patterns such as edges, curves, and shapes in handwritten
digits.
 Ensures efficient representation of digit characteristics for
classification.

3. Model Training Module

 Trains the CNN model using a labelled dataset (e.g., MNIST).

 Implements backpropagation and optimization algorithms (e.g., Adam,
SGD).
 Uses data augmentation techniques to improve generalization.
 Saves the trained model for future predictions.

4. Digit Recognition Module

 Takes user-uploaded images as input.

 Passes them through the trained CNN model for classification.
 Outputs the predicted digit along with a confidence score.

5. User Interface (UI) Module

 Provides a web-based or desktop interface for users to upload images.

 Displays recognized digits and confidence levels.
 Supports real-time handwriting input for instant recognition.

6. Performance Evaluation Module

 Measures model accuracy, precision, recall, and F1-score.

 Compares CNN performance with traditional models (e.g., SVM,
KNN).
 Generates reports for performance analysis and improvement.
 Software Design Diagrams

Class Diagram

A Class Diagram is a structural diagram in UML that represents the

blueprint of a system by showing its classes, attributes, methods, and
relationships between objects. It helps in object-oriented design by
depicting how different components of a system interact, making it
essential for software development and database modeling.
Sequence Diagram

A Sequence Diagram is a UML interaction diagram that illustrates the

order of interactions between system components or objects over time. It
represents how messages are exchanged between entities in a sequential
manner, helping in understanding the workflow, execution order, and
system behavior for a given scenario.
Database Design (DFD/ER Diagram)

DFD (level 0)

A Level 0 Data Flow Diagram (DFD), also known as a Context Diagram, is

a high-level graphical representation of a system that shows how it interacts
with external entities (users, systems, or databases). It depicts the entire
system as a single process, with inputs and outputs flowing between the
system and its external entities.
DFD (level 1)

A Level 1 Data Flow Diagram (DFD) provides a detailed breakdown of the

main process from Level 0 DFD. It decomposes the single high-level
process into multiple sub-processes, showing data flow between them,
external entities, and data stores. This level helps in understanding internal
functionalities while maintaining an overall system perspective.
Gantt Chart

A Gantt Chart is a visual project management tool that represents a timeline

of tasks in a project. It shows task durations, dependencies, and deadlines,
helping teams track progress, manage resources, and ensure timely
completion. It is widely used for planning, scheduling, and monitoring
project workflows.
Pert Chart

A PERT (Program Evaluation and Review Technique) Chart is a project

management tool used to visualize task dependencies and timelines. It
represents project activities as nodes and their relationships as arrows,
helping in scheduling, identifying the critical path, and optimizing project
duration. It is useful for analysing task sequences and managing
uncertainties in complex projects.
 Conclusion

Handwritten digit recognition using Convolutional Neural Networks

(CNNs) is a powerful and efficient approach for automating digit
classification tasks.

This project successfully demonstrates how deep learning techniques can

enhance accuracy, overcome handwriting variations, and improve real-
world applications such as banking, postal services, and document
digitization.

By leveraging public datasets (MNIST), open-source frameworks

(TensorFlow/Keras), and GPU acceleration, the model achieves high
precision and real-time performance. The proposed methodology ensures a
scalable, adaptable, and cost-effective solution, making it feasible for large-
scale automation.

Future improvements, such as transfer learning, custom datasets, and

advanced architectures, can further enhance performance, enabling broader
applications like handwritten text recognition and signature verification.
This project lays a strong foundation for AI-driven automation in pattern
recognition, reducing manual effort and increasing efficiency in various
industries.

Additionally, the system can be integrated with Optical Character

Recognition (OCR) technology to enhance text recognition beyond digits,
making it useful for processing handwritten documents.

Furthermore, with edge deployment on embedded systems (Raspberry Pi,

Jetson Nano, etc.), the model can be optimized for real-time, low-power
applications, extending its usability to mobile and IoT-based solutions.
 References
[1] LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W., & Jackel, L. D.
(1989). "Handwritten Digit Recognition with a Back-Propagation Network." Advances in Neural
Information Processing Systems (NeurIPS).

[2] Deng, L. (2012). "The MNIST database of handwritten digit images for machine learning
research." IEEE Signal Processing Magazine, 29(6), 141-142.

[3] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). "ImageNet classification with deep
convolutional neural networks." Advances in Neural Information Processing Systems (NeurIPS).

[4] Patel, M., & Goyal, S. (2020). "A comparative study on handwritten digit recognition using
machine learning and deep learning techniques." International Journal of Machine Learning and
Computing, 10(2), 75-82.

[5] Zhang, X., Zhao, J., & LeCun, Y. (2015). "Character-level convolutional networks for text
classification." Advances in Neural Information Processing Systems (NeurIPS).

[6] Kottakota, A., Teja, P. S., Chander, R. K., & Hema, D. D. (2023). "NeuroWrite: Predictive
Handwritten Digit Classification using Deep Neural Networks." arXiv preprint arXiv:2311.01022.

[7] Hochuli, A. G., Britto Jr, A. S., Saji, D. A., Saavedra, J. M., Sabourin, R., & Oliveira, L. S.
(2020). "A Comprehensive Comparison of End-to-End Approaches for Handwritten Digit String
Recognition." arXiv preprint arXiv:2010.15904.

[8] Gondere, M. S., Schmidt-Thieme, L., Sharma, D. P., & Scholz, R. (2021). "Multi-script
Handwritten Digit Recognition Using Multi-task Learning." arXiv preprint arXiv:2106.08267.

[9] Bonyani, M., Jahangard, S., & Daneshmand, M. (2020). "Persian Handwritten Digit, Character
and Word Recognition Using Deep Learning." arXiv preprint arXiv:2010.12880.

[10] Hochuli, A. G., Britto Jr, A. S., Saji, D. A., Saavedra, J. M., Sabourin, R., & Oliveira, L. S.
(2020). "A Comprehensive Comparison of End-to-End Approaches for Handwritten Digit String
Recognition." arXiv preprint arXiv:2010.15904.

[11] Gondere, M. S., Schmidt-Thieme, L., Sharma, D. P., & Scholz, R. (2021). "Multi-script
Handwritten Digit Recognition Using Multi-task Learning." arXiv preprint arXiv:2106.08267.

[12] Bonyani, M., Jahangard, S., & Daneshmand, M. (2020). "Persian Handwritten Digit, Character
and Word Recognition Using Deep Learning." arXiv preprint arXiv:2010.12880.

[13] Hochuli, A. G., Britto Jr, A. S., Saji, D. A., Saavedra, J. M., Sabourin, R., & Oliveira, L. S.
(2020). "A Comprehensive Comparison of End-to-End Approaches for Handwritten Digit String
Recognition." arXiv preprint arXiv:2010.15904.