joearid/Amharic-OCR-CRNN-CTC-for-Printed-Ge-ez-Script-Recognition

Amharic OCR: CRNN-CTC pipeline for printed Ge'ez script. 96.3% accuracy, 0.64% CER. CNN+BiLSTM+CTC with weighted sampling for 280 glyphs. LLM-augmented correction outperforms ChatGPT OCR. 167 FPS on GPU. # Amharic OCR: CRNN-CTC for Printed Ge'ez Script Recognition A complete OCR pipeline for printed Amharic text using a Convolutional Recurrent Neural Network (CRNN) with Connectionist Temporal Classification (CTC) loss. Achieves **96.3% line-level accuracy** and **0.64% Character Error Rate** on held-out real printed data. **Key Innovation:** LLM-augmented OCR pipeline that combines our specialized CRNN with large language model post-processing, outperforming general-purpose vision models on Amharic text. --- ## 🎯 Problem Amharic, written in the Ge'ez script (ግዕዝ), is a liturgical and literary language used by the Ethiopian Orthodox Church and millions of speakers. Despite its cultural importance, **no mainstream OCR solution exists** for Amharic comparable to English OCR systems. The script presents unique challenges: - **280 unique glyphs** (abugida system: consonant + vowel combinations) - **Complex morphology** with subtle diacritic distinctions (e.g., ሰ vs ሠ, ሰ vs ስ) - **Severe class imbalance** (Zipf's law: 40 characters appear ≤10 times) - **Limited digitized training data** compared to Latin scripts --- ## 🏗️ Architecture We implement a **CRNN-CTC** architecture inspired by Shi et al. (2017): ``` Input (1×64×256) ↓ CNN Backbone (7 conv layers, VGG-style) ↓ Feature Map (512×4×64) → Reshape → Sequence (64×2048) ↓ 2-Layer Bidirectional LSTM (hidden=256) → (64×512) ↓ Linear Classifier → (64×280 classes) ↓ CTC Loss (alignment-free training) ``` | Component | Details | |-----------|---------| | **CNN** | 7 convolutional layers with BatchNorm, ReLU, progressive max-pooling. Output stride: 4× in width, 16× in height | | **RNN** | 2-layer BiLSTM, 256 hidden units, bidirectional | | **Output** | 280 classes (278 characters + space + CTC blank) | | **Parameters** | ~7.5M total | | **Loss** | CTC Loss (Graves et al., 2006) — no character-level annotations needed | ### Design Choices | Choice | Justification | |--------|---------------| | **CNN backbo …