YoHa2024NKU/Tigrinya_TrOCR_Printed

A fine-tuned TrOCR model for printed Tigrinya text recognition using the Ge'ez script. Adapted from microsoft/trocr-base-printed using vocabulary extension and Word-Aware Loss Weighting to resolve word-boundary failures caused by BPE space-marker conventions.  # Tigrinya TrOCR **Adapting TrOCR for Printed Tigrinya Text Recognition: Word-Aware Loss Weighting for Cross-Script Transfer Learning** -76b900.svg) --- **All experiments in this project are conducted on synthetic printed Tigrinya text line images. Both the TrOCR-Handwritten and TrOCR-Printed models, as well as the CRNN-CTC baseline, are evaluated exclusively on this printed dataset. No real handwritten text is used in this study.** **TigrinyaTrOCR** is a fine-tuned Transformer-based OCR model designed for the **Tigrinya language** (Ge'ez script). It utilizes the **Microsoft TrOCR** architecture (Vision Transformer Encoder + GPT-2 Decoder) to achieve state-of-the-art results on printed Tigrinya text. This repository provides the first Transformer-based benchmark for printed Tigrinya OCR using the Ge'ez script. We fine-tune TrOCR with an extended tokenizer (+230 Ge'ez characters) and **Word-Aware Loss Weighting**, a technique that resolves systematic word-boundary failures caused by applying Latin-centric BPE to a non-Latin script. Both the handwritten and printed TrOCR-base variants are fine-tuned under identical conditions and converge to near-identical performance on printed data, confirming that the adaptation methodology rather than the pre-training domain is the dominant factor. The Tigrinya writing system uses the Ge'ez script (fidel), an abugida comprising 33 base consonants, 7 vowel orders (231 core syllographs), 4 labialized consonant groups, and 8 punctuation marks. ## Results ### Main Results (Printed Synthetic Test Set, n=5,000) | Model | CER (%) | WER (%) | Accuracy (%) | |----------------------|---------|---------|--------------| | TrOCR-Handwritten | 0.38 | 1.15 | 96.86 | | TrOCR-Printed | 0.22 | 0.87 | 97.20 | | CRNN-CTC Baseline | 0.12 | 0.57 | 98.20 | ### Ablation: Word-Aware Loss Weighting | Approach | CER (%) | WER (%) | Accuracy (%) …