Cyril's Lexicon Layout Analysis Through Deep Learning

Outstanding achievements in several vision and image processing tasks through deep learning have led the scientific community to adopt these techniques for digitizing manuscripts. Such a strategy is widely known as document layout analysis and concerns detecting and classifying the writings' elements, e.g., lines, images, and text. As one of humanity's priorities is preserving cultural heritage, this paper offers a deep learning approach for layout analysis employed on Cyril's Lexicon, an extensive Greek dictionary compiled between the end of the fifth and the beginning of the sixth century AD. The proposed pipeline, built upon YOLOv9m, is trained over the manual annotated Lexicon scanned pages to detect and classify its elements, particularly in lemma, definition, or abbreviation. Through an extensive experimental protocol, we prove that our system can reach high performance.