This notebook implements a convolutional neural network (CNN) for handwritten digit recognition on MNIST dataset images resized to 128x128 pixels. The key parts of the notebook are:
Importing TensorFlow/Keras and image processing libraries.
Loading MNIST digits and resizing them from 28x28 to 128x128.
Preprocessing data by normalizing pixels and one-hot encoding labels.
Visualizing sample training images.
Building a CNN model with two convolutional layers, max pooling, dropout, and dense layers.
Training the model for 5 epochs with batch size 64 and 20% validation split.
Saving the trained model as mnist_cnn_128x128_model.h5.
Evaluating test accuracy around 98.5% and printing a classification report.
Plotting training/validation accuracy and loss curves.
Displaying a confusion matrix and sample predictions visually.
The model achieves strong performance on enlarged MNIST dataset images, demonstrating effective CNN training and digit classification.