Handwritten Digit Classification

A CNN-based handwritten digit classifier trained on both the MNIST and USPS datasets. This project explores how data augmentation techniques (noise injection and rotation) affect model robustness across different handwriting sources.

Approach

1. Data Loading — MNIST digits via TensorFlow and USPS postal digits via DeepLake
2. Preprocessing — USPS images (16x16) are padded to match MNIST dimensions (28x28)
3. Data Augmentation — Random noise injection and rotation applied at configurable levels
4. Model Training — Four CNN variants trained under different data conditions
5. Evaluation — Models compared across clean, noisy, and rotated test sets using accuracy, precision, recall, and F1 score

Models

Model	Training Data	Purpose
Clean Model	Clean MNIST + USPS	Baseline performance
Mixed Model	Noisy MNIST + USPS	Robustness to noise
Rotated Clean Model	Rotated clean data	Robustness to rotation
Rotated Mixed Model	Rotated noisy data	Combined robustness

Model Architecture

• 3 convolutional layers (32, 64, 64 filters) with ReLU activation and max pooling
• Fully connected dense layer (64 units)
• Output layer with 10 units (softmax) for digit classification
• Optimizer: Adam | Loss: Sparse Categorical Crossentropy

Tech Stack

• Python 3.11
• TensorFlow / Keras
• NumPy, SciPy
• scikit-learn
• Matplotlib
• Pillow
• DeepLake

Project Structure

├── HandwrittenDigitClassification.ipynb   # Full pipeline notebook
├── Models/                                 # Saved trained models (.keras)
├── requirements.txt                        # Python dependencies
└── README.md

Getting Started

git clone https://github.com/yourusername/handwritten-digit-classification.git
cd handwritten-digit-classification
pip install -r requirements.txt
jupyter notebook HandwrittenDigitClassification.ipynb

License

MIT License