Scholarlink

<p>Ancient Japanese manuscripts are invaluable cultural assets but are often degraded by stains, fading, and bleed-through. We propose a restoration framework integrating the Diffusion Denoising Restoration Model (DDRM) with a two-stage mask generation process combining Automatic Color Equalization (ACE) and Gaussian Mixture Model clustering across multiple color spaces. A key novelty is the use of binarized text masks as guidance signals for DDRM through noise masking, which constrains denoising to character shapes and suppresses background noise. A high-resolution patch-based strategy with feather blending further enables seamless reconstruction of both text and background. Experiments on synthetic degradations of The Pillow Book and real manuscripts such as Tsurezuregusa demonstrate significant improvements over binarization-based DDRM, with an average PSNR gain of 10 dB, SSIM increase of 0.17, and LPIPS reduction from 0.49 to 0.28. The framework enhances textual clarity, preserves red annotations, and offers a scalable solution for AI-driven cultural heritage preservation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Restoration of ancient Japanese manuscripts via the diffusion denoising restoration model and color space-based masking

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月6日

查看全文
  • Yuya Yoshizu,
  • Hayata Kaneko,
  • Ryuto Ishibashi,
  • Lin Meng

摘要

Ancient Japanese manuscripts are invaluable cultural assets but are often degraded by stains, fading, and bleed-through. We propose a restoration framework integrating the Diffusion Denoising Restoration Model (DDRM) with a two-stage mask generation process combining Automatic Color Equalization (ACE) and Gaussian Mixture Model clustering across multiple color spaces. A key novelty is the use of binarized text masks as guidance signals for DDRM through noise masking, which constrains denoising to character shapes and suppresses background noise. A high-resolution patch-based strategy with feather blending further enables seamless reconstruction of both text and background. Experiments on synthetic degradations of The Pillow Book and real manuscripts such as Tsurezuregusa demonstrate significant improvements over binarization-based DDRM, with an average PSNR gain of 10 dB, SSIM increase of 0.17, and LPIPS reduction from 0.49 to 0.28. The framework enhances textual clarity, preserves red annotations, and offers a scalable solution for AI-driven cultural heritage preservation.