MNIST Paint Digit Classifier

A simple convolutional neural network (CNN) built with TensorFlow/Keras to classify handwritten digits using the MNIST dataset. It also includes a tool to recognize digits drawn in external software (like Paint) on a black background with a white number.

This project is designed as a beginner-friendly yet slightly more advanced introduction to convolutional neural networks. It goes beyond the typical minimal examples by including:

• Batch normalization
• ELU activations
• Dropout regularization

More importantly, it focuses on real usability: instead of limiting recognition to internal MNIST test images, it enables direct classification of digits drawn externally, making it practical, visual, and more engaging.

Who is this for?

• Students starting with AI or deep learning who want to test models with their own images.
• Teachers looking for a quick, visual demo of digit recognition.
• Anyone who wants to train a CNN and classify real-world inputs without setting up complex pipelines or external datasets.

Project Structure

Mnist_numbers/
├── mnist_cnn.py         # Script to train the CNN model using MNIST
├── test_model.py        # Script to test the model with custom PNG images
├── numberdetector.h5    # Trained model saved in HDF5 format
└── numerospaint/        # Folder containing hand-drawn digit images (PNG format)
    ├── zero.png
    ├── one.png
    ├── ...
    └── nine.png

Features

• Trains a CNN on the MNIST dataset with advanced layers.
• Saves the model as an .h5 file.
• Allows prediction from hand-drawn images (square, grayscale).

Input Digit Samples

These are all the digits used as input examples from the numerospaint/ folder:

Results

• Achieves ~99% accuracy on MNIST test set.
• Achieves 100% accuracy when classifying Paint images with clear digit shapes and correct format.

Requirements

• Python 3.7+
• TensorFlow
• NumPy
• OpenCV
• Matplotlib

If you need help setting up TensorFlow with GPU support, you can follow this guide:
TensorFlow GPU Setup by pelayo-felgueroso

How to Use

1. Train the Model

Run mnist_cnn.py to train the CNN and save the model:

python mnist_cnn.py

This generates a file called numberdetector.h5.

2. Test with Custom Images

Place your hand-drawn digit PNGs (black background, white number, square aspect ratio) into the numerospaint/ folder. Then run:

python test_model.py

The script will:

• Load the specified image
• Preprocess and resize it to 28x28
• Predict the digit and display the confidence

Example Output

The model predicts the image is the number: 4, with confidence: 0.99987

Image Format Guidelines

• Format: PNG
• Aspect Ratio: Square
• Background: Black
• Digit: White (centered and clearly drawn)

Author

Pelayo Felgueroso

---

Feel free to fork, modify, and improve this repository. It's designed for learning and experimentation with computer vision and CNNs using TensorFlow.