Handwritten Digit Recognition

Implementation of a Convolutional Neural Network (CNN) to classify handwritten digits using the MNIST dataset.

Project Structure

.
├── digit.png
├── LICENSE
├── model.keras
├── predict.py
├── README.md
├── requirements.txt
├── statistics
│   ├── diagnostics.png
│   └── summary.png
└── train.py

Usage

1. Clone the repository and cd into it

git clone https://github.com/eko-071/digit-recognition.git
cd digit-recognition/

2. Save the image being classified in the folder as digit.png.

Note: Input images should be grayscale with a black background and a white digit, similar to MNIST

3. Install required packages

pip install -r requirements.txt

4. Run the script

python predict.py

Training the Model

To train the model from scratch and generate evaluation plots:

python train.py

The script:

• Trains a CNN on the MNIST dataset
• Uses 5-fold cross-validation
• Saves training statistics in statistics/
• Saves the trained model as model.keras

Results

• Achieves ~99% accuracy on the MNIST dataset.
• diagnostics.png shows two plots: cross-entropy loss vs. epochs (top) and classification accuracy vs. epochs (bottom), with training and validation curves for each cross-validation fold.
• summary.png shows a box plot of classification accuracy values obtained from all cross-validation folds.

Note: Training data is blue, and testing data is orange.

Requirements

• numpy
• matplotlib
• tensorflow
• scikit-learn

References

• But what is a neural network? | Deep learning chapter 1 - 3Blue1Brown
• How to Develop a CNN for MNIST Handwritten Digit Classification - Jason Brownlee