End-to-End Text Line Detection and Ordering
Researchers have developed Orli, an end-to-end model for text line detection and ordering in historical documents. Orli treats the entire process as a single image-to-sequence problem, directly generating text lines in reading order from a page image. Trained on a large, diverse corpus, Orli achieves state-of-the-art performance on line detection and reading order benchmarks, even without dataset-specific training, and can adapt to specialized layouts with fine-tuning. The model's code and weights are publicly available. AI
IMPACT This model could improve automated analysis of historical documents by handling complex layouts and marginalia more effectively.