trocr-traditional-chinese-baseline

TrOCR Traditional Chinese Baseline

Model Description

This is a TrOCR model trained from scratch on 4.1M synthetic Traditional Chinese OCR dataset for historical document recognition.

• Base Model: microsoft/trocr-base-stage1
• Training Data: 4.1M synthetic Traditional Chinese images
• Training Steps: 100k steps
• Task: Optical Character Recognition (OCR) for Traditional Chinese

Intended Use

This baseline model is designed for:

• Traditional Chinese text recognition
• Historical document digitization
• Vertical and horizontal text layout support

Training Details

Training Data

• Dataset: ZihCiLin/traditional-chinese-ocr-synthetic
• Size: 4.1 million image-text pairs
• Layout: 50% horizontal, 50% vertical
• Character Set: 13,172 Traditional Chinese characters (CNS11643)

Training Hyperparameters

• Batch size: 64 (per device)
• Learning rate: 5e-5
• Optimizer: AdamW
• LR scheduler: Cosine
• Training steps: 100k
• Warmup ratio: 0.06

Usage

from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image

# Load model and processor
processor = TrOCRProcessor.from_pretrained("ZihCiLin/trocr-traditional-chinese-baseline")
model = VisionEncoderDecoderModel.from_pretrained("ZihCiLin/trocr-traditional-chinese-baseline")

# Load image
image = Image.open("document.jpg").convert("RGB")

# Generate text
pixel_values = processor(image, return_tensors="pt").pixel_values
generated_ids = model.generate(pixel_values)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

print(generated_text)

Limitations

• Trained on synthetic data; performance may degrade on real historical documents with severe degradation
• Best performance on clean, well-printed text
• For historical manuscripts, use the finetuned variant: ZihCiLin/trocr-traditional-chinese-historical-finetune

Citation

If you use this model, please cite:

@inproceedings{lin2026decoding,
  title={Decoding-Time Fusion of OCR and Large Language Models for Traditional Chinese Historical Document Recognition},
  author={Lin, Zih-Ci and Liao, Wen-Hung},
  booktitle={International Conference on Pattern Recognition (ICPR)},
  year={2026}
}

Related Resources

• Paper: Coming soon (ICPR 2026)
• Code: GitHub Repository
• Synthetic Generator: ocr-synth-generator
• Annotation System: document-ocr-annotation-system