MNIST Digit Recognition with CNN

A deep learning project using a Convolutional Neural Network (CNN) to classify handwritten digits (0-9) from the MNIST dataset. This project demonstrates image classification with Keras/TensorFlow.

🚀 Overview

This project trains a CNN on the MNIST dataset and provides a script to test custom digit images. The model achieves high accuracy on standard MNIST data and can predict digits from user-provided images.

🏗️ Model Architecture

The CNN includes:

• 2 Convolutional Layers with ReLU activation and max pooling for feature extraction.
• Flatten Layer to transition to dense layers.
• 2 Dense Layers (64 units with ReLU, 10 units with Softmax for classification).
• Trained with Adam optimizer and sparse categorical crossentropy loss.

Note: The model is trained on MNIST data with white digits on a black background. Test images are automatically inverted if they have a white background to match this format.

📁 Files

• MNSITcode.py: Trains the model on MNIST and saves it as mnist_model.keras.
• Test_Model.py: Loads the model and predicts digits from image.png.
• MNIST_Complete.py: Combined script for training and testing.
• mnist_model.keras: Saved trained model (generated after running MNSITcode.py).

🛠️ Setup and Usage

1. Install Dependencies:

   pip install tensorflow numpy matplotlib pillow

2. Train the Model:

   python MNSITcode.py

   This trains for 10 epochs with validation and saves the model.

3. Test on Custom Image:

   • Place a 28x28 grayscale image named image.png in the directory.
   • Run:

     python Test_Model.py

   • The script handles image inversion if needed and outputs the prediction with confidence.

📊 Results

• Training Accuracy: ~90-99% on MNIST test set.
• Custom Prediction: Works best with clear, centered digits. Includes automatic preprocessing for color inversion.

🔧 Notes

• Ensure image.png and image2.png is grayscale; the script resizes and preprocesses it.
• For low-confidence predictions, check image quality or retrain with more epochs.
• Model uses .keras format for compatibility.