About This Project
Dataset
This classifier is trained on the MNIST dataset. Each image is 28×28 pixels in grayscale.
Neural Network Architecture
The model uses a Convolutional Neural Network (CNN) with the following structure:
- Input Layer: 28×28×1 grayscale images
- Conv Block 1: 32 filters (3×3) + BatchNorm + MaxPool + Dropout
- Conv Block 2: 64 filters (3×3) + BatchNorm + MaxPool + Dropout
- Conv Block 3: 128 filters (3×3) + BatchNorm + MaxPool + Dropout
- Dense Layers: 256 → 128 neurons with Dropout
- Output Layer: 10 classes (digits 0-9) with Softmax
How It Works
When you draw a digit, the system:
- Captures your drawing from the canvas
- Detects the bounding box of your digit
- Crops, centers, and resizes to 28×28 pixels
- Sends to the CNN model for classification
- Returns probabilities for each digit (0-9)