BemnetMussa/DeepLearning---Amharic-handwriting-detection-Project

A deep learning system for recognizing handwritten Amharic (Ethiopic) characters, featuring a custom CNN architecture, Grad-CAM explainability, and an interactive 3D neural network visualization built with Three.js. ፊ AHCNN — Amharic Handwriting Character Recognition A deep learning system for recognizing handwritten Amharic (Ethiopic) characters, featuring a custom CNN architecture, Grad-CAM explainability, and an interactive 3D neural network visualization built with Three.js. Features • Architecture • Dataset • Getting Started • Usage • Project Structure • Results • Contributing • License --- ## Overview The Amharic script (Ge'ez script) is used by over 50 million speakers of Amharic, Tigrinya, and other Ethiopian languages. Despite being one of the most widely used writing systems in Africa, OCR tooling for Amharic remains scarce compared to Latin or CJK scripts. **AHCNN** (Amharic Handwriting CNN) is a lightweight convolutional neural network trained to classify **237 individual Amharic characters** from 28×28 grayscale images. The project pairs the model with: - A **FastAPI** inference server exposing prediction endpoints - A **React + Three.js** frontend that renders the neural network in 3D and animates neuron activations in real time as the model processes each character - **Grad-CAM** heatmaps that explain which regions of the input the model focuses on ## Features - **237-class character recognition** covering the core Ethiopic syllabary - **Custom CNN architecture** (4 convolutional blocks → global average pooling → classifier) - **Grad-CAM explainability** — visualize what the model "sees" in each input - **Real-time 3D visualization** — a Three.js scene showing neurons firing, connections lighting up, and the predicted character highlighted in a grid of all 237 classes - **Demo / auto-pilot mode** — automatically samples random images from the dataset and animates predictions - **Drawing canvas** — draw an Amharic character directly in the browser and get a live prediction *(canvas UI is included but commented out by default; see Usage to enable)* - **FastAPI backend** with `/predict` (upload image) and `/predict_random` (random dataset samp …