HANDWRITTEN DIGIT RECOGNITION

A mini-project dissertation submitted in the partial fulfillment of academic

requirements in
VI – Semester (Sixth Semester)
In
COMPUTER SCIENCE AND ENGINEERING
By
DARA SUSHMITHA SREE(100521733082)

T.PRIYANKA (100521733093)

MANEESHA Y (100521733033)
Under the guidance of
Ms V.SUKANYA
Assistant Professor

Department of Computer Science and Engineering

University College of Engineering (Autonomous)

Osmania University, Hyderabad-500007

1
 University College of Engineering (Autonomous) Osmania
University, Hyderabad- 500007

CERTIFICATE

This is to certify that this project work entitled “ HAND DIGIT RECOGNITION”
is a bonafide work carried out by :
DARA SUSHMITHA SREE 100521733082

T.PRIYANKA 100521733093

MANEESHA Y 100521733033

from the department of “Computer Science and Engineering” in the partial fulfillment

of academic requirements in VI- Semester (Sixth Semester) in “Computer Science

and Engineering” from “University College of Engineering(A), Osmania

University”, Hyderabad.

Project Guide Head of the department

Dr. Sukanya Prof.P.V.Sudha
Assistant Professor Head of the department
Dept. of CSE, OU Dept. of CSE, OU

2
 ACKNOWLEDGEMENT

I would like to express my deep sense of gratitude and whole-hearted thanks to my project guide
V SUKANYA , Assistant professor, Department of Computer Science and Engineering,
University College of Engineering Osmania University, for giving me the privilege of working
under his guidance, with tremendous support and cogent discussion, constructive criticism and
throughout this dissertation work carrying out the project work.

I also thank Prof P.V Sudha, Head Department of Computer Science and Engineering for her
support from the Department and allowing all the resources available to us students.
I also extend my thanks to the entire faculty of the Department of Computer Science
and Engineering, University College of Engineering ,Osmania University who encourages
us throughout the course of our Bachelor degree and allowing us to use the many resource
present in the department.

Our sincere thanks to our parents and friends for their valuable suggestions, morals, strength and
support for the completion of our project.

MANEESHA Y 100521733033

DARA SUSHMITHA SREE 100521733082

T.PRIYANKA 100521733093

3
 Table of Contents
S. No Title PAGE NUMBERS

1 Introduction 5 -8

2 Literature Survey 9

3 Problem Statement 10

4 System Design 13-15

5 System Requirements 15
Specifications

6 Technologies Used 16

7 Implementation 17 -20

8 Results 21

9 Testing 22

10 Conclusion And Future

Enhancements
23-25

11 Bibliography and References 26-28

4
INTRODUCTION

➢ In the realm of machine learning and artificial intelligence, the MNIST dataset stands
as a cornerstone, representing a fundamental challenge: classifying handwritten digits.

➢ MNIST, short for Modified National Institute of Standards and Technology,

comprises 60,000 training images and 10,000 testing images of handwritten digits,
each labelled with their corresponding numbers from 0 to 9.

➢ This dataset has served as a benchmark for evaluating the effiency of various machine
learning algorithms, particularly neural networks.

➢ Within this context, the endeavor of MNIST digit classification with neural networks
encapsulates a pivotal exploration into the capabilities and intricacies of artificial
intelligence. By leveraging the power of neural networks, we embark on a journey to

5
 develop models capable of accurately discerning handwritten digits with high
precision and efficiency.

➢ This project delves into the construction, training, and optimization of neural
networks, aiming to achieve superior performance in digit recognition tasks.

➢ Through meticulous data preprocessing, innovative model architectures, and advanced

optimization techniques, we strive to push the boundaries of accuracy and robustness
in MNIST digit classification.

➢ Moreover, this project not only serves as an educational endeavor but also as a
practical demonstration of the practical applications of neural networks in real-world
scenarios.

➢ The skills honed and knowledge gained through this project lay a solid foundation for
tackling more complex machine learning challenges and pave the way for future
innovations in the field.

In this introduction, we lay the groundwork for a comprehensive exploration of MNIST digit
classification with neural networks, setting the stage for an engaging and enlightening
journey into the realm of artificial intelligence and machine learning.

6
Overview of the project:

➢ The project involves building a neural network to classify handwritten digits from the
MNIST dataset using Python.

➢ The workflow includes data collection, image preprocessing (resizing, converting to

grayscale, normalization), model building, training, and evaluation.

➢ The neural network is created using Keras, featuring layers like input, dense, and output
with specific activation functions.

➢ The model is trained with the Adam optimizer and sparse categorical cross-entropy
loss, ensuring reproducible results by setting a random state.

➢ Model performance is evaluated through accuracy metrics and a confusion matrix, and
predictions are made on new, preprocessed images.

➢ The project's goal includes not only training the model but also evaluating its
performance on test data using metrics like accuracy and confusion matrices, and
deploying it for predictions on new images

7
Representation of handwritten images with confusion matrix

8
Objectives:

1. Understand Data Collection:

Learn how to source and collect appropriate datasets for image recognition tasks,
specifically using the MNIST dataset.

2. Preprocess Image Data:

Perform essential preprocessing steps, such as resizing images, converting RGB images to
grayscale, and normalizing pixel values.

3. Build a Neural Network Model:

Construct a neural network using Keras with appropriate layers, including input, dense
and output layers, to handle image classification.

4. Implement Activation Functions:

Apply activation functions like ReLU and softmax to enhance model performance and
interpretability.

5. Train the Model:

Train the neural network using the Adam optimizer and sparse categorical cross-entropy
loss function to achieve high accuracy in digit classification.

6. Ensure Reproducibility:
Set a random state to ensure the reproducibility of results, making it easier to replicate and
validate findings.

7. Evaluate Model Performance:

Assess the trained model's performance using accuracy metrics and confusion matrices
to understand its effectiveness on both training and test datasets.

8. Visualize Results:
Use visualization tools like heat maps to interpret the confusion matrix and gain insights
into the model's prediction capabilities.

8. Deploy the Model for Predictions:

Develop a pipeline to preprocess new images and use the trained model to predict
handwritten digits accurately.
Literature Survey on MNIST Digit Classification with Neural
Network

The MNIST dataset, created by Yann LeCun and colleagues in the 1990s, is a collection of
70,000 images of handwritten digits from 0 to 9. This dataset has become a standard
benchmark for testing and comparing machine learning models, especially those used in the
recognition of images.

Early Neural Network Models

➢ In the early days, simple neural networks called feedforward networks were used for
digit classification. However, these models had limited success due to their inability to
effectively handle the spatial structure of images.

Convolutional Neural Networks (CNNs)

➢ The introduction of Convolutional Neural Networks (CNNs) by Yann LeCun

revolutionized image recognition tasks. LeCun's model, LeNet-5, was specifically
designed for recognizing handwritten digits and showed much better performance than
previous models. CNNs work well with images because they can capture spatial
hierarchies and patterns within the data.

Improvements in CNNs

Over the years, researchers have made various improvements to CNNs to enhance their
performance on the MNIST dataset. Some of these improvements include:

➢ Deeper Networks: Building deeper networks with more layers, such as VGGNet and
ResNet, helps in capturing more complex features.

➢ Activation Functions: Using advanced activation functions like ReLU (Rectified

10
 Linear Unit) to introduce nonlinearity into the model.

➢ Regularization Techniques: Methods like Dropout and Batch Normalization help

prevent overfitting, ensuring that the model generalizes well to new, unseen data.

Data Augmentation and Preprocessing

➢ To make the models more robust, data augmentation techniques are used. These involves
creating additional training examples by rotating, shifting, or scaling the original images.
Preprocessing steps like normalizing pixel values (scaling them to a range between 0 and 1)
also improve model performance.

Modern Techniques

Recently, advanced techniques like Transfer Learning and ensemble methods have been
applied to MNIST digit classification.

➢ -Transfer Learning: This involves using a model pre-trained on a large dataset and
finetuning it on the MNIST dataset. This approach leverages features learned from more
complex tasks.
➢ -Ensemble Methods: Combining predictions from multiple models to achieve better
accuracy and robustness.

Evaluation and Performance

➢ The performance of neural network models on the MNIST dataset is usually measured in
accuracy, which is the percentage of correctly classified digits.

➢ Modern models have achieved accuracy rates above 99.7%, making the task nearly solved.

➢ However, MNIST remains a valuable benchmark for testing new machine learning ideas
and techniques.

11
 PROBLEM STATEMENT

➢ The objective of this project is to develop a robust and accurate neural network model
capable of classifying handwritten digits (0-9) from the MNIST dataset.

➢ The MNIST dataset consists of 60,000 training images and 10,000 test images of
handwritten digits, each represented as a 28x28 pixel grayscale image.

Key challenges

• Image Recognition:
Effectively recognizing and classifying digits despite variations in handwritten styles,
thickness, orientation, and noise in the images.
• Model Complexity:
Designing an appropriate neural network architecture that balances complexity and
performance, ensuring the model is neither too simple to underfit nor too complex to overfit
the data.
• Generalization:
Ensuring the trained model can generalize well to unseen data, maintains high accuracy on the
test set.
• Training Efficiency:
Optimizing the training process to achieve high accuracy within a reasonable computational
power and resource usage.
• Regularization:
Implementing techniques to prevent overfitting and ensure the model maintains high
performance on new, unseen images.

12
The project will involve the following steps:

Data Preprocessing:

Normalizing and augmenting the dataset to improve model robustness.

Model Design:

Constructing a neural network, likely a Convolutional Neural Network (CNN), to exploit

spa al hierarchies in the image data.

Training and Validation:

Training the model on the training set, validating performance using a validation on set,
and tuning hyperparameters to optimize accuracy.

Evaluation:

Assessing the model's performance on the test set using metrics such as accuracy,
precision, recall, and confusion matrix.

Deployment:
Preparing the model for deployment in real-world applications where handwritten digit
recognition is required.

The successful completion on of this project will result in a neural network model that can
accurately classify handwritten digits, demonstrating the practical application of neural
networks in image recognition tasks.

SYSTEM DESIGN

1. Data Collection and Preprocessing

Data Source: The MNIST dataset, which contains 60,000 training images and 10,000
13
 test images of handwritten digits.
Preprocessing: Preparing the data for the model by normalizing pixel values (scaling
them between 0 and 1) and reshaping images as needed.

2. Model Selection

Choosing a Model: A Convolutional Neural Network (CNN) is often chosen because

it is very effective for image recognition tasks.

Model Architecture:
Input Layer: Takes in the preprocessed image data.

Convolutional Layers: Extract features from the images using filters.

Fully Connected Layers: Flatten the feature maps and connect them to output neurons.

14
 Output Layer: Consists of 10 neurons (one for each digit from 0 to 9) with a softmax

activation function to output probabilities for each digit.

3. Training the Model

Training Process: The model is trained using the training dataset. The model learns by

adjusting its weights through backpropagation to minimize the error in its predictions.

Optimizer: An algorithm (like Adam or SGD) that updates the model’s weights to

minimize the loss function

Data Flow Diagram

Data set image processing splitting train

Neural network model

15
 Input image neural network model out put

SYSTEM REQUIREMENTS AND SPECIFICATION

HARDWARE REQUIREMENTS

TECHNOLOGIES USED

Backend Technologies

Python: The primary programming language used for developing the backend.

Seaborn: A Python visualization library based on Matplotlib. It provides a high-level

interface for drawing attractive and informative statistical graphics.

TensorFlow: Deep learning frameworks used to develop, train, and deploy the neural
network model.

Data Processing and Model Training

16
 NumPy: A library for numerical computations, useful for handling and processing data.

Pandas: A library for data manipulation and analysis.

OpenCV: Libraries for image processing, used to preprocess the input images before

feeding them into the neural network.

Tools

Google Colab: A cloud service that can be used for training the neural network model,
offering free access to GPUs.

IMPLEMENTATION
Firstly we import the dependenies like numpy,pyplot (from
matplotlib),seaborn,cv2,cv2_imshow and confusion_matrix,etc as shown in below image .

Now loading the MNIST data from keras datasets by using load_data() method.

17
Here the x_train and y_train are stored in form of arrays .the below picture prints the data of
x_train[10] value.

And x_test and y_test stores values.(0-9)

Now the data is compressed to fit the neural network.this is obtained by dividing the data by
255 .By doing this the data is compress in between 0-1(before it was 0-255).

Data before regularization Data after regularization

Here the data is divided into the test and train data as shown in below figure

18
 After preparation of the data for neural networks We should build the layers of the neural
networks. For this we used layers module from keras along with Flatten() and Dense()
methods.

Here we have taken a image of a digit from mnist dataset through x_test by using imshow() and show() from

plt module for testing the model

19
The input image is to be resized (to 28,28) after compilation and testing the accuracy of the
neural network

Now this resized image is given as input to trained model. This model will recognize the
number and prints as shown in below image
20
ACCURACY
Accuracy is a key indicator of a model's success in recognizing handwritten digits. A
high accuracy means the model is reliable and effective in identifying digits correctly
we have used below measures for accuracy improvement.
Model Choice: Use robust models like Convolutional Neural Networks (CNNs), which
are well-suited for image recognition tasks

Regularization: Use techniques to prevent overfitting, ensuring the model generalizes

well to new data.

We got below accuracy of 97% for our model after taking suitable measures

21
CONCLUSION

• The MNIST Digit Classification project demonstrates the power and efficiency of
neural networks in recognizing handwritten digits.

22
 • By leveraging a deep learning model, we successfully developed a system that can
accurately classify handwritten digits, providing quick and reliable results.

• Through this project, we have gained valuable insights into the application of machine
learning and neural networks in real-world tasks.

• Upon the completion of this project not only showcases the capabilities of AI in digit
classification but also lays the groundwork for future projects involving image
recognition and classification tasks.

• Overall, the MNIST Digit Classification project is a testament to how advanced

technologies can simplify and automate tasks, making processes more efficient and
accessible.

FUTURE SCOPE

❖ The MNIST Digit Classification project opens up many exciting possibilities for future
work and improvements.

❖ Enhancements can involve using advanced neural network architectures like

Convolutional Neural Networks (CNNs) for improved accuracy.

❖ Integration with real-world applications, such as automated data entry or sorting systems,
can increase practical utility

23
The other features that includes are:

1.Improving Accuracy:

• Use more advanced neural network architectures like convolutional neural

networks (CNNs) to improve classification accuracy.
• Implement techniques like data augmentation to create more training data and enhance
model performance.

2. Expanding to Other Datasets:

• Extend the model to recognize handwritten letters or symbols from other datasets.
• Develop models for more complex image recognition tasks beyond digits.

3. Real-Time Classification:

• Optimize the system to classify digits in real-time using live camera feeds or
• Continuous input from drawing interfaces.

4. Mobile and Offline Applications:

• Develop mobile applications that can classify digits offline, providing a wider
range of accessibility.
• Create desktop applications that can run without an internet connection.

5. User Feedback Loop:

24
• Implement a feedback system where users can correct wrong classifications, helping the
model learn and improve over time.

• Use this feedback to continuously update and refine the model.

6.Split the input image to recognize the multiple digits :

• Implement the improved model that will be able to recognize the multiple digits in
given input image

7.Integration with Other Systems:

• Integrate the digit classification system with educational tools for teaching and practicing
handwriting.
• Apply the technology in other fields, such as digitizing handwritten forms , checks in
banking.

8.Security Enhancements:
• Ensure secure handling of user data and inputs, especially in applications
that require sensitive information.
• Implement robust input validation to prevent misuse or exploitation of the
system.

9.Scalability:

• Enhance the system to handle a larger number of simultaneous users efficiently.

• Deploy the application on cloud platforms to ensure scalability and reliability.

25
 BIBILOGRAPHY & REFERENCES

BIBLIOGRAPHY
Books

➢ Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. This book

provides a comprehensive introduction to deep learning techniques, including neural

networks and their applications in image recognition

➢ Geron, A. (2017). Hands-On Machine Learning with Scikit-Learn and TensorFlow:

Concepts, Tools, and Techniques to Build Intelligent Systems. O'Reilly Media. This book

covers machine learning and deep learning with practical examples, including the use of the

MNIST dataset.

Online Resources
➢ MNIST Dataset:

➢ Handwritten Digits. Available at:

➢ http://yann.lecun.com/exdb/mnist/. This website provides access to the

➢ MNIST dataset and detailed information about its composition and usage.

➢ Keras Documentation.

26
➢ Available at: https://keras.io/. The official documentation for Keras, which provides

guidelines and examples for building and training neural networks, including applications

to the MNIST dataset.

➢ Seaborn Documentation. Available at: https://seaborn.pydata.org/.

➢ The official documentation for Seaborn, a visualization library used for creating

informative and attractive statistical graphics

➢ Matplotlib Documentation. Available at: https://matplotlib.org/. The official documentation

for Matplotlib, a plotting library in Python, used as the foundation for Seaborn.

REFERENCES

➢ Patel, A., Kumar, P., & Mehta, R. (2022). Enhanced Handwritten Digit

Recognition Using Optimized Convolutional Neural Networks. Journal of Artificial

Intelligence and Data Science, 4(2), 123-131.

➢ This study uses of optimized CNN architectures for improving the accuracy of digit

recognition on the MNIST dataset.

27
➢ Ghosh, A., Sen, S., & Roy, A. (2023). Transfer Learning Approaches for Handwritten Digit

Recognition. Applied Soft Computing, 119, 108793.

➢ This paper examines the application of transfer learning techniques to enhance the

performance of models trained on the MNIST dataset.

➢ Khan, A., Sohail, A., Zahoora, U., & Qureshi, A. S. (2020). A Survey of the recent

Architectures of Deep Convolutional Neural Networks.

➢ Artificial Intelligence Review , 53(8), 5455-5516. This paper reviews recent

advancements in CNN architectures, which are crucial for improving performance on

datasets like MNIST.

➢ Rawat, W., & Wang, Z. (2017). Deep Convolutional Neural Networks for the

Image Classification: A Comprehensive Review. Neural Computation, 29(9),

2352-2449.

➢ This paper reviews the use of deep CNNs in image classification tasks,

including MNIST.

28
29
30
23