Search papers, labs, and topics across Lattice.
This study addresses the challenges of digitizing complex historical Sanskrit manuscripts by introducing a local traditional OCR pipeline that utilizes iterative fine-tuning at both layout and appearance levels. The method significantly reduces human annotation effort by adapting to the specific manuscript distribution, leading to improved transcription accuracy across multiple pages. Results show substantial quantitative gains in OCR performance, alongside a newly created dataset featuring detailed layout-level and Unicode annotations in PAGE-XML format.
Iterative fine-tuning of OCR can drastically cut down the time and expertise needed for transcribing complex historical manuscripts.
Digitizing the text from handwritten historical manuscripts is required to make them easily accessible, preservable, and to enable historical scholars to study them in new ways. Historical manuscripts, however, often exhibit complex heterogeneous layouts and non-standard appearance due to period-specific writing styles, page textures, camera noise, and other nuisance factors, making them difficult to perform OCR on. To tackle this challenge, we introduce a local traditional OCR pipeline, which can be iteratively fine-tuned on the target manuscript at the layout-level and the appearance-level. By adapting to the target manuscript distribution, the proposed Traditional OCR pipeline makes better predictions on subsequent pages, causing iterative reduction in human annotation effort, which is expensive and time-consuming as it requires historical domain expertise. Using this pipeline, we digitize text from three complex historical Sanskrit manuscripts and introduce a dataset with granular layout-level annotations, along with Unicode annotations in the standard PAGE-XML format. We demonstrate quantitative gains due to iterative fine-tuning of the proposed traditional OCR pipeline, and also benchmark the performance of leading Multi-Modal Large Language Models on the introduced Dataset. Code and dataset are available at: https://github.com/flame-cai/gnn-synthetic-layout-historical/.
