Reproduce “An Empirical Study of Scaling Law for OCR” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Miao Rang, Zhenni Bi, Chuanjian Liu, Yunhe Wang, Kai Han
This empirical study investigates scaling laws for Optical Character Recognition (OCR), specifically text recognition, by systematically varying model size, data volume, and computation. The research established predictable power-law relationships, developed the large-scale REBU-Syn dataset, and achieved a new state-of-the-art average accuracy of 97.42% on six benchmarks with the CLIP4STR-L* model, also demonstrating improved performance in large multimodal models.
Key insights
- Smooth power-law relationships exist in OCR, demonstrating predictable performance scaling with model size, data volume, and computational resources.
- Larger OCR models utilize training data more efficiently and converge faster, while an optimal synthetic-to-real data ratio (e.g., 1:3 for PARSeq-S) and diverse synthetic data are crucial for performance.
- Task-related pre-training (e.g., OCR-specific data) is more effective for OCR models than general visual pre-training (e.g., ImageNet), due to the importance of visual-textual connections.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.