This is not just another MNIST classifier. Most tutorials teach you to train a model, but they fail when you try to use it in real life because users don't draw perfectly centered digits.
This project solves that problem.
I built a custom desktop application with a Smart Preprocessing Engine. It behaves like a human eye: looking at the drawing, cropping the useless whitespace, and centering the digit-just like the data the AI was trained on.
- 🧠 Custom CNN Architecture: A lightweight but powerful Convolutional Neural Network.
- 🎯 Smart Centering Algorithm: Draw in the corner, draw small, draw big—it doesn't matter. The algorithm fixes your input automatically.
- 🖥️ Native GUI: No browser needed. Fast, responsive
tkinterinterface. - ⚙️ On-Device Training: Train the neural network locally on your CPU/GPU with one click.
- 🐧 Cross-Platform: Runs smoothly on Windows, macOS, and Linux.
You need Python 3.10+ installed.
git clone https://github.com/v0id-core/digit-recognizer.git
cd digit-recognizerWindows:
python -m venv venv
venv\Scripts\activatemacOS / Linux:
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txtpython desktop_app.py- Train the model: If you launch the app for the first time, click "Train Model" in the app interface. The app will download the MNIST dataset and train the neural network (~1-2 minutes).
- Draw: Use your mouse to draw a digit (0-9) on the canvas.
- Recognize: Click "Recognize" to see the prediction and confidence score.
- Core: Python 3.10+
- ML Framework: PyTorch, Torchvision
- GUI: Tkinter
- Image Processing: PIL (Pillow), NumPy
- App doesn't start? Ensure that you have Python 3.10+ and all dependencies installed correctly. Check version: python --version.
- Training fails? Check your internet connection. The MNIST dataset requires downloading when you press "Train Model".
- This project is licensed under the MIT License. You can freely use, modify, and distribute the code.
- v0id-core
- GitHub: https://github.com/v0id-core