AlphaReader

A handwritten character recognition system built with PyTorch. Trains a CNN on the EMNIST-Letters dataset and uses it to recognise A–Z characters from real images.

Project Structure

OCR/
├── Scanner.py   # CLI entry point
├── config.py    # Shared constants (device, paths, transforms)
├── model.py     # AlphabetCNN architecture
├── train.py     # Training & evaluation on EMNIST
├── predict.py   # Single-character inference
├── scan.py      # Full-image segmentation & OCR
└── README.md

Setup

Requirements

• Python 3.8+
• PyTorch
• torchvision
• Pillow
• NumPy
• OpenCV (only for the scan command)

Install Dependencies

pip install torch torchvision pillow numpy opencv-python

Usage

All commands are accessed through Scanner.py:

1. Train the Model

python Scanner.py train

Flag	Default	Description
--epochs	5	Number of training epochs
--batch-size	64	Training batch size
--lr	0.001	Learning rate

# Example: train for 10 epochs with a smaller learning rate
python Scanner.py train --epochs 10 --lr 0.0005

The trained weights are saved to alphabet_cnn.pth.

2. Predict a Single Character

python Scanner.py predict <image_path>

python Scanner.py predict letter_a.png
#    Predicted character: A
#    Confidence: 98.3%

3. Scan a Full Image

python Scanner.py scan <image_path>

This segments individual characters from the image using adaptive thresholding and contour detection, groups them into lines, and prints the recognised text.

python Scanner.py scan handwritten_note.png
#    Detected 12 character(s):
#
#    HELLO
#    WORLD

Model Architecture

AlphabetCNN — a lightweight CNN for 28×28 greyscale character images:

Layer	Details
Conv2d	1 → 32, kernel 3×3
Conv2d	32 → 64, kernel 3×3
MaxPool2d	2×2
Dropout	0.25
Linear	9216 → 128
Linear	128 → 26 (A–Z)

How It Works

1. Training — The model is trained on EMNIST-Letters (26 classes, A–Z) with cross-entropy loss and Adam optimiser.
2. Preprocessing — Input images are converted to greyscale, resized to 28×28, transposed (to match EMNIST orientation), and normalised.
3. Scanning — OpenCV applies adaptive thresholding and finds contours to isolate individual characters. Each crop is fed through the CNN for prediction. Characters are grouped into lines based on vertical position, and word boundaries are detected via horizontal gaps.