✍️ Handwritten Digit Recognizer

A Convolutional Neural Network (CNN) trained on the MNIST dataset that achieves ~99.4% test accuracy on handwritten digit recognition (0–9).
Includes an interactive Gradio web demo where you can draw a digit and get an instant prediction.

---

🖼️ Demo

Draw on canvas	Upload an image
Draw any digit with your mouse or touchscreen	Upload a 28×28 (or any size) grayscale image

---

📁 Project Structure

handwritten-digit-recognizer/
├── train.py          # Train the CNN on MNIST and save model.h5
├── predict.py        # CLI tool to run inference on any image
├── app.py            # Interactive Gradio web demo
├── requirements.txt  # Python dependencies
└── README.md

---

🚀 Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Train the Model

python train.py

This will:

• Automatically download the MNIST dataset (~11 MB)
• Train the CNN for up to 30 epochs with early stopping
• Save the best model to model.h5
• Generate training_curves.png

Expected output:

Test Accuracy : 99.40%
✅ Model saved to model.h5

3. Launch the Web Demo

python app.py

Opens a Gradio interface in your browser at http://localhost:7860.

4. CLI Prediction

python predict.py --image my_digit.png
python predict.py --image my_digit.png --show   # also display a visualisation

---

🧠 Model Architecture

Input (28×28×1)
  │
  ├─ Conv2D(32, 3×3) + BatchNorm → Conv2D(32, 3×3) + BatchNorm
  ├─ MaxPool(2×2) → Dropout(0.25)
  │
  ├─ Conv2D(64, 3×3) + BatchNorm → Conv2D(64, 3×3) + BatchNorm
  ├─ MaxPool(2×2) → Dropout(0.25)
  │
  ├─ Flatten → Dense(256) + BatchNorm → Dropout(0.5)
  └─ Dense(10, softmax)

Training details:

Parameter	Value
Optimizer	Adam (lr = 1e-3)
Loss	Categorical cross-entropy
Batch size	128
Max epochs	30 (early stopping)
Data augmentation	Random rotation ±10°, zoom ±10%, translation ±10%
Test accuracy	~99.4%

---

📊 Results

Metric	Value
Training accuracy	~99.8%
Validation accuracy	~99.5%
Test accuracy	~99.4%
Parameters	~430,000

---

🛠️ Requirements

• Python 3.9+
• TensorFlow / Keras 2.12+
• NumPy
• Pillow
• Matplotlib
• Gradio 4+

---

📜 License

MIT License — see LICENSE.