Digit Detection Using MNIST Dataset

Author: Yash Patel

📌 Dataset

This project uses the MNIST dataset, an open-source collection of handwritten digit images (0–9).

• 60,000 training images
• 10,000 testing images
• Each image is 28 × 28 pixels, grayscale
• Each image is labeled with the digit it represents (0–9)

🎯 Objective

The goal of this project is to build a Convolutional Neural Network (CNN) that can recognize and classify handwritten digits with high accuracy.

🛠 Preprocessing

• Normalization: Pixel values were scaled to the range [0,1] for faster convergence.
• One-hot encoding: Labels were converted to vectors so that the model outputs probabilities across 10 classes.

🏗 Model Architecture

The model is a CNN, consisting of the following layers:

1. Conv2D (32 filters, 3×3, ReLU) ×2

   • Extracts local features (edges, corners, textures).
   • ReLU introduces non-linearity.

2. MaxPooling2D (2×2)

   • Reduces spatial dimensions, speeds up computation, and reduces overfitting.

3. Dropout (25%)

   • Randomly drops neurons during training to prevent overfitting.

4. Conv2D (128 filters, 3×3, ReLU) ×2

   • Learns deeper, more detailed patterns.

5. MaxPooling2D (2×2) + Dropout (25%)

6. Conv2D (256 filters, 3×3, ReLU)

   • Detects high-level abstract features.

7. MaxPooling2D (2×2) + Dropout (25%)

8. Flatten

   • Converts 2D feature maps into a 1D vector.

9. Dense Layers

   • 128 → 100 → 80 → 40 neurons (ReLU)
   • Fully connected layers for feature combination.

10. Dense (10, Softmax)

• Outputs probabilities for digits 0–9.
• The highest probability corresponds to the predicted digit.

⚙️ Training & Testing

• Optimizer: Adam (adaptive learning rate)
• Loss Function: Categorical Cross-Entropy (multi-class classification)
• Metric: Accuracy

The model was trained for 10 epochs with a batch size of 128.

📊 Results

• Test Accuracy: 99.36% on the 10,000 test images
• The model demonstrates excellent performance in handwritten digit recognition.

---