Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Sep 4·5 min read
How Far Can Synthetic Data Take Thai OCR?
TL;DR
Wayu-Paxa-OCR-Zero, a Thai OCR model, was developed using synthetic data without real OCR labels, achieving a median character error rate reduction from 6.64% to 1.24% on printed pages.
✦ Why It Matters
Consider using synthetic data for training OCR models to enhance performance without relying on extensive real-world datasets.
Key Takeaways
- Utilize synthetic data to train OCR models effectively.
- Focus on typeface diversity to improve model performance.
- Incorporate real handwriting glyphs for better transferability.
- Evaluate training granularity for optimal results in OCR tasks.
How It Works
Wayu-Paxa-OCR-Zero was developed by analyzing the impact of various factors on OCR performance, including typeface diversity and spatial structure. A controlled document-reconstruction pipeline was used to evaluate the effects of these factors under different training conditions, such as page-level and crop-level training.