Handwritten Digit Recognition

A PyTorch-based CNN for recognizing handwritten digits (0–9), with an interactive desktop UI for real-time inference via drawing, webcam, and image upload.

Features

• Drawing canvas — draw digits directly and get instant predictions.
• Webcam input — recognize digits from a live camera feed.
• Image upload — load and classify digit images from disk.
• Multi-digit recognition — detects and classifies sequences of digits.
• Confidence filtering — predictions below 80% confidence are discarded.
• Data augmentation — expanded ~1,000 samples to ~52,000 using ±25° rotations.

Installation

Requires Python 3.8+.

pip install -r requirements.txt

Usage

Launch the desktop app:

python -m ui.main_app

Keyboard Shortcuts:

Shortcut	Action
Ctrl+S / Enter	Run prediction
Ctrl+Z	Undo last stroke
Ctrl+O	Upload image
Delete / Backspace	Clear input
Ctrl+Q	Quit

Training

1. Place source images in data/raw/ or data/augmented/.
2. Generate train/validation splits:

   python prepare_dataset.py

3. Augment the dataset (optional):

   python augment_data.py

4. Train the model:

   python train.py

5. Evaluate performance:

   python evaluate.py

Project Structure

Digits_Recognition/
├── data/                   # Raw and processed datasets
├── images/                 # Training curves, confusion matrix
├── inference/              # Prediction and preprocessing logic
│   ├── predictor.py
│   ├── preprocessor.py
│   └── webcam_stream.py
├── models/                 # CNN architecture and saved checkpoints
│   ├── cnn_model.py
│   └── checkpoints/
├── training/               # Training pipeline
│   ├── augmentation.py
│   ├── dataset_loader.py
│   ├── metrics.py
│   └── trainer.py
├── ui/                     # Desktop interface (Tkinter)
│   ├── main_app.py
│   ├── canvas_panel.py
│   ├── result_display.py
│   ├── upload_panel.py
│   └── webcam_panel.py
├── utils/                  # Logging and visualization helpers
│   ├── logger.py
│   └── visualizer.py
├── augment_data.py         # Standalone augmentation script
├── evaluate.py             # Model evaluation script
├── prepare_dataset.py      # Dataset split script
├── train.py                # Training entry point
└── requirements.txt

Results

Training Curves

The model reaches over 99% training accuracy and nearly 100% validation accuracy with no signs of overfitting.

Confusion Matrix

Near-perfect classification across all 10 digits with minimal misclassifications.