Handwritten Paragraph Recognition Using Spatial Information on Russian Notebooks Dataset

Handwritten paragraph recognition is a vital aspect of handwritten document analysis, enhancing accuracy and usability across various applications. However, recognizing paragraphs in handwritten documents is challenging due to layout variations and irregularities. Spatial information, encompassing spatial relationships between text elements, is essential for accurate paragraph segmentation and document comprehension. Recent works in handwritten Russian recognition have primarily focused on character and line-level recognition. This study is the first attempt on paragraph-level recognition for Russian handwriting, utilizing the Vertical Attention Network (VAN) with a hybrid attention method. Key contributions include the preparation of a unique Russian dataset at the paragraph level, containing around 2600 images with PAGE XML-encoded ground truth. The VAN model was fine-tuned for whole paragraph recognition, and comprehensive experiments were conducted, comparing its performance against alternative non-layout-aware approaches. This work advances layout-aware recognition in handwritten Russian documents, addressing an unexplored area in the field.