Yonatanhaile2026/tigrinya-trocrhandwritten

TrOCR-Handwritten for Tigrinya OCR

Tigrinya TrOCR — Handwritten Variant

Adapting TrOCR for Printed Tigrinya Text Recognition: Word-Aware Loss Weighting for Cross-Script Transfer Learning

A fine-tuned TrOCR model for printed Tigrinya line-level text recognition. This is the handwritten pre-training variant, fine-tuned from microsoft/trocr-base-handwritten using vocabulary extension and Word-Aware Loss Weighting to resolve word-boundary failures caused by BPE space-marker conventions.

---

Model Details

Field	Value
Model name	Yonatanhaile2026/tigrinya-trocr-handwritten
Base model	microsoft/trocr-base-handwritten
Task	Tigrinya OCR (image-to-text)
Language	Tigrinya (ti)
Script	Ge'ez
Model type	VisionEncoderDecoderModel
Vocabulary	Extended from 50,265 → 50,495 tokens (230 Ge'ez characters added)
Training data	GLOCR Tigrinya News text-line images (synthetic)

---

Performance

Evaluated on a held-out test set of 5,000 synthetic Tigrinya text-line images.

Metric	Value
Character Error Rate (CER)	0.38%
Word Error Rate (WER)	1.15%
Exact Match Accuracy	96.86%

Bootstrap 95% Confidence Intervals (1,000 iterations, TrOCR-Printed)

Metric	Point Estimate	95% CI
CER	0.20%	[0.17%, 0.24%]
WER	0.76%	[0.64%, 0.90%]
Accuracy	97.44%	[97.02%, 97.84%]

Bootstrap intervals were computed on the TrOCR-Printed variant; see the printed model card for details.

Comparison (same dataset and split)

Model	CER	WER	Accuracy
TrOCR-Handwritten (fine-tuned)	0.38%	1.15%	96.86%
TrOCR-Printed (fine-tuned)	0.22%	0.87%	97.20%
CRNN-CTC Baseline	0.12%	0.57%	98.20%

---

Training Details

Hyperparameter	Value
Optimizer	AdamW
Learning rate	4e-5
LR scheduler	Linear decay (no warmup)
Epochs	10
Per-device batch size	2
Gradient accumulation steps	4
Effective batch size	8
Mixed precision	FP16
Boundary loss weight	2.0
Random seed	42
Training duration	~2h 40m
Hardware	NVIDIA RTX 5060 Laptop (8 GB GDDR7)

---

How to Use

from transformers import VisionEncoderDecoderModel, TrOCRProcessor
from PIL import Image

processor = TrOCRProcessor.from_pretrained("Yonatanhaile2026/tigrinya-trocrhandwritten")
model = VisionEncoderDecoderModel.from_pretrained("Yonatanhaile2026/tigrinya-trocrhandwritten")

Load your text-line image
image = Image.open("your_tigrinya_text_line.png").convert("RGB")

pixel_values = processor(images=image, return_tensors="pt").pixel_values
generated_ids = model.generate(pixel_values, num_beams=5, max_length=128)
prediction = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(prediction)

---

Intended Use

Suitable for:

• Tigrinya OCR research on synthetic or clean text-line images
• Baseline comparison against printed-specific and CTC-based OCR models
• Research on cross-script transfer learning and BPE tokenizer adaptation

Not suitable for:

• Production OCR without validation on real scanned or handwritten documents
• Scenarios where the printed variant would be preferable
• Documents with heavy degradation, low resolution, or non-text noise

---

Limitations

• Trained and evaluated exclusively on synthetic printed data from a single domain (newspaper text lines)
• Performance on real-world scanned or genuinely handwritten Tigrinya documents is not validated
• Underperforms the printed TrOCR variant on this synthetic printed corpus
• Results reflect a single training run on one hardware configuration

---

Related Resources

• Printed variant: Yonatanhaile2026/tigrinya-trocr-printed
• Code repository: github.com/YoHa2024NKU/Tigrinya_TrOCR_Printed
• Dataset: GLOCR — Harvard Dataverse

---

Citation

If you use this model, please cite the associated paper and repository:

@misc{medhanie2026adaptingtrocrprintedtigrinya,
      title={Adapting TrOCR for Printed Tigrinya Text Recognition: Word-Aware Loss Weighting for Cross-Script Transfer Learning}, 
      author={Yonatan Haile Medhanie and Yuanhua Ni},
      year={2026},
      eprint={2604.20813},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.20813}, 
}

---

License

MIT