MNIST Handwritten Digit Recognition with PyTorch and Streamlit

This project implements a handwritten digit recognition system using a Convolutional Neural Network (CNN) trained on the MNIST dataset. The model is developed in PyTorch and deployed as a Streamlit web app where users can draw digits and get real-time predictions.

Motivation

Inspired by handwriting recognition features on smartphones, this project allows users to draw numbers on a web-based canvas, and a deep learning model predicts the digit.

Features

• CNN-based Digit Recognition:

  • A PyTorch-trained CNN classifies handwritten digits from the MNIST dataset.
  • The trained model is deployed using Streamlit.

• Interactive Web Interface:

  • Users can draw digits using a mouse or touchscreen.
  • Predictions are displayed in real-time.

Getting Started

Prerequisites

Ensure you have the following installed:

• Python 3.x
• PyTorch & torchvision
• Streamlit
• OpenCV & NumPy (for image processing)

Installation

1. Clone the Repository

git clone https://github.com/your-username/mnist-digit-recognition.git
cd mnist-digit-recognition

2. Train the Model (Optional)

Run the Jupyter Notebook (draft.ipynb) to train the model:

jupyter notebook draft.ipynb

This will:

• Load the MNIST dataset.
• Train a CNN model using PyTorch.
• Save the trained model as model.pth.

3. Install Dependencies

pip install -r requirements.txt

4. Run the Streamlit App

streamlit run app.py

Visit http://localhost:8501 in your browser.

How It Works

1. CNN Training:

   • A convolutional neural network (CNN) is trained on MNIST using PyTorch.
   • The trained model is saved as model.pth.

2. Web-Based Prediction:

   • Users draw digits on a canvas.
   • The image is resized, normalized, and fed to the model.
   • The CNN predicts the digit.

3. Displaying Predictions:

   • The predicted number is shown below the canvas.

To-Do

• Improve accuracy by training on augmented datasets.
• Deploy the app online (e.g., Streamlit Sharing, Hugging Face Spaces).
• Optimize frontend performance.

License

This project is licensed under the MIT License. See the LICENSE file for details.