Fig. 3: Preprocessing of manuscript images for restoration experiments.

Example of the preprocessing step applied to historical manuscripts prior to segmentation and restoration. (Left) Original high-resolution scan of a double-page spread, captured with a color calibration strip for reference. Images were derived from high-resolution scans of the Tsurezuregusa manuscript (National Institute of Japanese Literature, CODH dataset). (Right) Cropped single-page images automatically generated using LabelImg by manually annotating the four corner points of each page. This preprocessing step standardizes the input dimensions and ensures consistent alignment across samples, facilitating downstream color correction, clustering, and diffusion-based restoration.