aderajewYeshiwendm/Amharic_alphabetical_family_recognition

# Amharic Alphabet Family Recognition Using Deep Learning ## 📌 Project Overview This project focuses on recognizing **Amharic alphabetical family characters (ሀ to በ)** using deep learning techniques. The goal is to build an end-to-end deep learning system starting from **custom data collection** to **model design, training, and evaluation**. The system processes scanned handwritten worksheets containing a 10×7 grid of characters, automatically extracts individual character images, applies data augmentation, and trains a Convolutional Neural Network (CNN) to classify characters into 10 alphabet families. The trained model achieves **~98% accuracy** on the test set. --- ## 🎯 Objectives - Collect a **custom dataset** of Amharic alphabetical family characters (ሀ to በ) - Extract characters from scanned worksheets using perspective transformation - Preprocess and augment the dataset (6 variations per character) - Design and train a **PyTorch CNN model** for character recognition - Evaluate model performance using accuracy, confusion matrix, and loss curves --- ## 🔠 Alphabet Classes The project focuses on the following **10 Amharic alphabet family characters**: - **ሀ** (ha) - **ለ** (le) - **ሐ** (hha) - **መ** (me) - **ሠ** (se) - **ረ** (re) - **ሰ** (sa) - **ሸ** (sha) - **ቀ** (qe) - **በ** (be) --- ## 📊 Dataset Collection - The dataset was **collected manually** by scanning handwritten worksheets - **80 scanned sheets** (JPG images) containing 10×7 grids of characters - Each sheet contains 70 character cells (10 rows × 7 columns) - Characters were written by multiple individuals to ensure diversity - Images were captured using phone cameras and stored in `scanned_sheets/` directory - No open-source or publicly available datasets were used **Dataset Statistics:** - Original scanned sheets: **80 images** - Extracted character images (after augmentation): **~22,500 images** - Training set: **~15,800 images** (70%) - Validation set: **~3,300 images** (15%) - Test set: ** …