sanoyz/ethiopic-handwriting-api

Ethiopic Handwriting Recognition API # 🖋️ Ethiopic Handwriting Recognition API **Live API Endpoint:** ethiopic-handwriting-recogn… --- ## 📖 Overview This API serves a research prototype for **Online Ethiopic Handwriting Recognition**. Unlike traditional OCR which processes static images, this system captures pen strokes as a **temporal signal** (x/y coordinates, timestamps, and inter-stroke dynamics). This project introduces the first comprehensive study of online Ethiopic handwriting recognition. It is an **independent, unfunded research initiative** built to demonstrate the feasibility of using memory-augmented transformers and temporal features for low-resource script recognition. > **⚠️ Important:** This project is **distinct** from my Master's thesis and uses a smaller, private dataset. The full thesis is currently under a formal university publication embargo. --- ## 🚀 Key Technical Contributions ### 1. Rich Feature Engineering We extract a **156-dimensional stroke feature representation** per stroke, combining: - **Resampled point trajectories** (x, y, and curvature). - **Per-stroke statistics** (length, bounding box, total duration). - **Inter-stroke gap features** encoding the spatial, temporal, and structural relationships between consecutive strokes. ### 2. Memory-Augmented Temporal Encoder The core architecture is built around a **position-aware memory bank** of learnable stroke-pattern prototypes. This memory is fused with the input stroke sequence through **multi-head cross-attention**, allowing the model to dynamically retrieve relevant stroke patterns during inference. ### 3. Dual Ablation Methodology We employ a rigorous dual ablation strategy: - **Retraining from scratch** without certain components. - **Lesioning** (removing components at inference time) from a converged checkpoint. Key findings from this methodology (with the current dataset): - The **memory bank is structurally necessary**. Removing it degrades performance significantly. - …