Handwritten Character Recognition (A–Z)

A desktop-based handwritten character recognition system using a CNN trained on self-collected data. The project includes a zipped self-drawn dataset for immediate training and testing.

Features

• Draw characters using mouse
• Real-time prediction with confidence score
• Save new samples for training
• Retrain model directly from UI
• Undo strokes (Backspace / Ctrl+Z)
• Keyboard-based prediction (Enter key)
• Single unified desktop interface

Tech Stack

• Python 3.10+
• PyTorch
• OpenCV
• NumPy
• Tkinter

Project Structure

handwriting_recognition/
│
├── app/
│   └── unified_whiteboard.py
│
├── model/
│   ├── cnn.py
│   └── train.py
│
├── data/
│   ├── self_data.zip
│   └── self_data/
│       ├── A/
│       ├── B/
│       └── ...
│
├── inference/
│   ├── preprocess.py
│   └── predict.py
│
├── utils/
│   ├── config.py
│   └── label_map.py
│
└── README.md

Setup

Create Virtual Environment

python -m venv venv

Activate Environment

venv\Scripts\activate

Install Dependencies

pip install torch torchvision numpy opencv-python pillow

Dataset Setup (Required)

The repository includes a zipped self-collected dataset.

Extract Dataset

Expand-Archive data\self_data.zip data\

After extraction, the structure should be:

data/self_data/A
data/self_data/B
...

Training

Train the model using the provided self-data:

python -m model.train

The trained model is saved to:

model/model.pth

Run Application

Start the unified desktop UI:

python -m app.unified_whiteboard

Usage

1. Draw a character on the board
2. Press Enter or click Predict
3. Press Backspace or Ctrl+Z to undo
4. Enter a label and click Save Sample to add new training data
5. Click Retrain Model to update the model

Model Architecture

• Input: 1 × 28 × 28 grayscale image
• Convolutional layers with ReLU and max pooling
• Fully connected layers
• Output: 26 classes (A–Z)

Notes

• The included dataset matches the whiteboard input format
• Adding more samples and retraining improves accuracy
• Public datasets were intentionally avoided due to input mismatch