Urdu Handwritten Alphabets Recognition

Handwritten Urdu alphabet recognition using DNN and CNN on student-collected images.

Overview

This repository contains code, dataset samples, and reports for a course project that builds models to recognize handwritten Urdu alphabets captured with a mobile phone camera.

Features

• DNN baseline (2 dense hidden layers)
• CNN baseline (2 convolutional layers)
• Preprocessing and augmentation (rotation, translation, zoom, contrast, gaussian noise)
• Experiments with multiple input sizes (28×28, 64×64, 128×128)
• Evaluation: accuracy, precision/recall/F1, confusion matrices

Technologies

• Python 3.11
• TensorFlow / Keras
• Pillow (PIL), scikit-learn, matplotlib, seaborn
• FPDF for report generation

Quick Start

1. Create a Python virtual environment and install dependencies (example):

python -m venv .venv
.\.venv\Scripts\Activate
pip install -r requirements.txt

2. Prepare dataset (if not present): place unaugmented submission images in 2023cs607(A)/alphabets_128x128 or use dataset/ structure shown in the repo.

3. Train models (example, 128×128):

python run_all.py --dataset dataset --img-size 128 --batch-size 32 --epochs 30

4. Generated outputs (models, reports) are saved under outputs* directories.

Dataset

• The repository includes a small student-collected dataset and a submission folder 2023cs607(A) with single-sample images (pre-augmentation).
• Image sizes included: 128×128 (also stored: 64×64, 28×28 outputs).
• Note: Teacher requested 34–35 submission images; this project contains 40 images in 2023cs607(A)/alphabets_128x128. Confirm with your instructor if needed.

Results

• See report.pdf and report.md for evaluation details and comparative analysis (best reported model: CNN at 128×128 in this project).

Files of interest

• run_all.py — training and evaluation script
• source_code_bundle.txt — all source code in single text file (submission requirement)
• report.pdf — 3–4 page project report (course requirement)
• 2023cs607(A)/alphabets_128x128 — submission images (pre-augmentation)

License

This project is released under the MIT License — see LICENSE.

Contact

Submitted by: Muhammad Rateeb (2023-CS-607) Email: (add your contact email here)

ML CCP - Handwritten Urdu Alphabet Recognition

What this project does

• Trains two models (DNN + CNN) on 40-class Urdu alphabet dataset
• Applies preprocessing and 5 data augmentations
• Saves metrics, confusion matrices, and trained models
• Generates a short report

Run training

.venv311\\Scripts\\python.exe run_all.py --img-size 128 --epochs 30

Outputs are written to the outputs/ folder.

Generate report

.venv311\\Scripts\\python.exe generate_report.py

This creates report.md and report.pdf.