Two-Stage Stains Reduction with Diffusion Models for Ancient Japanese Manuscripts
Yuya Yoshizu, Hayata Kaneko, Ryuto Ishibashi, Lin Meng, Mingcong Deng · 2024
Ancient Japanese manuscripts are a valuable cultural heritage that allows us to understand the culture and values of the past. However, due to preservation over hundreds of years, stains result in antique manuscripts, and they cause low readability and loss of contents to the manuscripts. Considering preserving these assets through virtual restoration, we propose a two-stage stain reduction based on Denoising Diffusion Restoration Models (DDRM). The stain reduction proceeds by combining removing background by the binarization and reconstruction of before-processing images by DDRM. In addition, the sequential conducting that contains patches of book pages and entire them leads to preserving local information and ensuring consistency across the pages. Generation results are qualitatively improved and more realistic than traditional binarization. Also, the proposal method is evaluated by two aspects, the stain reduction metrics and preserving characters metrics. The evaluation result shows our proposal is more effective and achieves higher restoration quality than conventional binarization.