Handwritten Digit Recognition

A simple neural network project that recognizes handwritten digits (0-9) using TensorFlow and OpenCV.

Features

• Train on MNIST dataset (60,000 training images)
• Save/load trained models
• Predict custom digit images

Requirements

Create requirements.txt:

numpy
opencv-python
matplotlib
tensorflow

Installation

Using uv package manager

uv install

• uv init # to init uv project
• uv venv # to create virtual env
• uv sync # to install dependencies from pyproject.toml
• uv run main.py # to run python file

Windows

python -m venv digit_env
digit_env\Scripts\activate
pip install -r requirements.txt

Linux/macOS

python3 -m venv digit_env
source digit_env/bin/activate
pip install -r requirements.txt

Project Structure

handwritten-digit-recognition/
├── main.py
├── requirements.txt
├── README.md
├── digits/
│   ├── digit1.png
│   └── digit2.png
└── handwritten-digit.keras

Usage

1. Train the model:

   python main.py

2. Test with custom images:

   • Create a digits/ folder
   • Add PNG images named digit1.png, digit2.png, etc.
   • You can make using paints in windows for testing purpose.
   • Run the script again

3. Load saved model:

   • Uncomment the line: model = tf.keras.models.load_model("handwritten-digit.model")
   • Comment out the training code

Model Architecture

• Input: 28×28 flattened images (784 features)
• Hidden Layer 1: 128 neurons (ReLU)
• Hidden Layer 2: 128 neurons (ReLU)
• Output: 10 neurons (Softmax)

Performance

• Training Accuracy: ~98%
• Test Accuracy: ~97%
• Training Time: ~2 minutes

---