Abhishek Jha
abhishekjha1008
AI & ML interests
Historical handwriting recognition (HTR) & multilingual document AI. Building specialist recognisers that stay faithful to the page where general VLMs hallucinate; cross-lingual transfer for low-resource scripts.
Recent Activity
liked a model 10 days ago
abhishekjha1008/yolov8m-historical-line-segmenter upvoted a collection 10 days ago
Historical Danish & Swedish HTR posted an update 10 days ago
Reading order: the quiet problem after you've found the lines 📜
On a historical page with more than one column, finding the text lines is largely solved. Deciding the order to read them in is not, and that order is what a recogniser is ultimately given. On multi-column pages a perfectly recognised page can still come out in the wrong order.
I built a small, learning-free step that recovers the order, and benchmarked it against htrflow, Microsoft's Table Transformer and PaddleOCR on 198 real handwritten tables (the USS Jeannette Arctic logbook).
A few honest findings:
- On dense handwritten tables, most detectors barely find the lines at all, so detection, not ordering, is the harder part there.
- Given the lines, the simple step is level with much heavier table models, and they all reach the same ceiling. A shared open problem, not solved by any one of them.
Try it: https://huggingface.co/spaces/abhishekjha1008/historical-reading-order-demo
Full study: https://valhtrdata01.z1.web.core.windows.net/
Dataset and code: https://huggingface.co/datasets/abhishekjha1008/historical-table-reading-order
Comments and pointers welcome.
#HTR #OCR #DigitalHumanities #HistoricalDocuments #LayoutAnalysisOrganizations
None yet
Upload folder using huggingface_hub
#1 opened about 1 month ago
by abhishekjha1008