Automated Text Line Segmentation For Ottoman Manuscript Transcription

Text line segmentation is essential for analyzing historical manuscripts. This taskbecomes challenging with documents handwritten documents, especially Ottomanmanuscripts, which often contain complex writing styles and overlapping lines. Thisthesis presents various deep learning-based approaches for automatic text line segmentationin Ottoman manuscripts. The U-Net-based approach relies on binary segmentationfollowed by a connected component post-processing. The YOLO-basedmethods include a single-stage (instance segmentation) and a two-stage (combiningobject detection and segmentation) approach. Models initially trained on Arabicdatasets, are tested on a 40-page Ottoman manuscript dataset containing bothstraight and angled text lines, using various segmentation metrics. The YOLO approaches,which showed promising results, were further evaluated for their effect onOCR performance using an Ottoman dataset with 199 text lines. The best performancewas obtained with the YOLO OBB & Segmentation method, achievinga precision score of 99.3% on straight lines and 95% on angled lines at a 75% IoUthreshold. Furthermore, this approach yielded 13.1% CER and 34.7% WER onthe OCR evaluation dataset. With this best-performing method, an automaticallysegmented Ottoman text line segmentation dataset is also generated, resulting in aprecision score of 96.5% evaluated on a 20-page subset. The dataset collected in thiswork, consisting of 40 manually labeled and 110 automatically labeled pages, is publiclyshared to contribute to research on line segmentation of Ottoman manuscripts.