JumaRubea/mnist_handwritten_recognition

lenet-5_architecture_model

Handwritten Digit Recognition

This project implements the LeNet-5 neural network architecture to recognize handwritten digits using the MNIST dataset.

Table of Contents

• Introduction
• Prerequisites
• Installation
• Model Architecture
• Results
• Contributing

Introduction

LeNet-5 is a classic convolutional neural network (CNN) architecture designed by Yann LeCun,learn more, primarily for handwritten digit classification. This project uses TensorFlow and Keras to build and train the LeNet-5 model on the MNIST dataset. The dataset is already included in this project.

Prerequisites

Make sure you have the following installed:

• [Python_3.6+] (https://www.python.org/downloads/)
• [TensorFlow_2.x] (https://www.tensorflow.org/install)
• [NumPy] (https://numpy.org/install/)
• Matplotlib
• Pandas (optional)

Installation

Clone this repository:

git clone https://github.com/jumarubea/lenet-5_architecture_model.git
cd lenet-5-digit-recognition

Model Architecture

LeNet-5 consists of the following layers:

• Convolutional Layer: 6 filters of size 5x5, activation function: tanh
• Average Pooling Layer: pool size 2x2
• Convolutional Layer: 16 filters of size 5x5, activation function: tanh
• Average Pooling Layer: pool size 2x2
• Convolutional Layer: 120 filters of size 5x5, activation function: tanh
• Flatten Layer
• Dense Layer: 84 units, activation function: tanh
• Output Layer: 10 units, activation function: softmax

Note: for the purpose of accuracy measure, i implement relu activation instead of tanh except for the 84 dense layer.

Results

The trained LeNet-5 model achieves a test accuracy of approximately 98% on the MNIST dataset.

Contributing

If you want to contribute to this project, please fork the repository and submit a pull request.