Recognizing paragraphs in handwritten documents poses challenges due to layout variations. Accurate paragraph segmentation relies on spatial information, encompassing the relationships between text elements. The study introduces the first Russian dataset at the paragraph level, comprising approximately 6293 paragraph images with PAGE XML-encoded ground truth, it is prepared from the Russian school notebooks’ dataset at word level, which contains approximately 1857 images detailed in JSON format. The VAN model is fine-tuned for comprehensive paragraph recognition, and its performance is compared with alternative non-layout-aware approaches.
Мохаммед С. (науч. рук. Тесля Н.Н.) Advancing handwritten paragraph recognition : A spatial approach with the Russian notebooks dataset // Сборник тезисов докладов конгресса молодых ученых. Электронное издание. – СПб: Университет ИТМО, [2024]. URL: https://kmu.itmo.ru/digests/article/12639