Text line segmentation of Ottoman manuscripts

Official URL: https://dx.doi.org/10.1109/SIU66497.2025.11112097

Text line segmentation is a key challenge in historical document analysis, especially for Ottoman manuscripts with complex calligraphy and overlapping text. This paper presents a comparative study of various deep-learning approaches for the automatic segmentation of plain text lines in Ottoman manuscripts. In the study based on YOLO and U-Net architectures, a series of adjustments and sub-approaches are used to address the challenges of Ottoman script. The U-Net focuses on fine-grained pixel-by-pixel segmentation, whereas YOLO handles the problem in two strategies: direct instance segmentation and instance segmentation applied after object detection using oriented bounding boxes to simplify the segmentation. Approaches are evaluated on manually labeled 25-paged Ottoman manuscripts. The comparative analysis reveals that the YOLO segmentation approach achieves the highest performance for plain text line segmentation with a pixel-level IoU score of 83.9% and F1-score of 91%.