Blind Text Image Super-resolution for Enhancing the Readability of Fragments Hidden in Book Bindings

Fragments reused in bookbindings are a crucial source for studying pre-modern book culture, representing books that were no longer read, and thus repurposed for their material value. The texts on these fragments are often hidden beneath pastedowns, scraped off, or otherwise obscured. Recovering their content is essential for paleographic and book historical analysis. Imaging methods such as Hyperspectral Imaging (HSI), Multispectral Imaging (MSI), and Ultraviolet Reflected (UVR) photography, combined with post-processing, are commonly used to reveal hidden text. However, the resulting images are often blurry and of low resolution, limiting their readability by both Optical Character Recognition (OCR) systems and human experts. This paper explores the use of deep generative models for super-resolving Latin text in fragment images, with a focus on fragments from the Herlufsholm Collection at the University of Southern Denmark. This collection contains numerous books with fragments—a valuable but understudied resource for investigating Scandinavian book history. We evaluate the effectiveness of text-specific super-resolution models in enhancing legibility and demonstrate their potential to support fragmentological research by making unreadable text accessible for scholarly analysis. The relevant material is available at: https://github.com/bhrnprhmd/Text_image_super_res_folatin_CHR2025.