Two-Stage Stain Reduction in Ancient Japanese Manuscripts via Mask-Guided Diffusion Models

Abstract

Ancient Japanese manuscripts are invaluable cultural heritage, yet their readability is often degraded by stains, bleed-through, and insect damage accumulated over centuries. To address this problem, we propose a two-stage stain reduction method based on Denoising Diffusion Restoration Models (DDRM). The first stage applies binarization to remove stains and extract text regions, while DDRM with noise masking reconstructs character details lost during thresholding. In the second stage, patch-wise restored images are reassembled and refined at the page level to ensure background consistency. Experiments on two representative works, \textit{Tsurezuregusa} (322 pages) and \textit{Isoho Monogatari} (198 pages), demonstrate the effectiveness of the proposed method. For \textit{Tsurezuregusa}, our method has achieved PSNR of 15.97 and SSIM of 0.715, compared with 7.46 and 0.510 for simple binarization. With gamma correction ($\gamma=0.8$), PSNR further improved to 18.03. For \textit{Isoho Monogatari}, we have obtained a PSNR of 18.48 and an SSIM of 0.776, outperforming binarization (9.74, 0.694), and the gamma correction increased PSNR to 20.68. In addition, the proposed method improved contrast (higher brightness variance than ground truth) while keeping kurtosis and entropy in a balanced range. The results demonstrate that our diffusion-based approach achieves both effective stain removal and faithful character preservation, contributing to the long-term accessibility and readability of ancient manuscripts.