Scholarlink

<p>As a world cultural heritage, Dunhuang murals face severe damage from erosion, aging, and human activity, necessitating high-quality digital restoration. Unlike prohibited physical renovations, digital methods enable non-destructive recovery and detailed archiving. While diffusion-based image restoration has advanced in other domains, mural restoration still struggles with blurred edges, color inconsistency, and insufficient detail. To address this, we propose a novel diffusion model incorporating an improved denoising network and an Enhanced Multi-head Contextual Attention (EMCA) strategy. This approach better captures complex features and multi-scale information, yielding clearer outputs and enhanced visual effects. Experimental results on three categories of damaged mural datasets (cracks, peeling, fading) demonstrate superior performance over baseline methods, with significant improvements in PSNR, SSIM, and LPIPS, enhancing both restoration accuracy and visual quality.</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

DMEA-Net: a diffusion-based network for mural image restoration using enhanced multi-head contextual attention

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月7日

查看全文
  • Yang Li,
  • Chuanlin Zhang,
  • Yacong Li,
  • Dong Sui,
  • Maozu Guo

摘要

As a world cultural heritage, Dunhuang murals face severe damage from erosion, aging, and human activity, necessitating high-quality digital restoration. Unlike prohibited physical renovations, digital methods enable non-destructive recovery and detailed archiving. While diffusion-based image restoration has advanced in other domains, mural restoration still struggles with blurred edges, color inconsistency, and insufficient detail. To address this, we propose a novel diffusion model incorporating an improved denoising network and an Enhanced Multi-head Contextual Attention (EMCA) strategy. This approach better captures complex features and multi-scale information, yielding clearer outputs and enhanced visual effects. Experimental results on three categories of damaged mural datasets (cracks, peeling, fading) demonstrate superior performance over baseline methods, with significant improvements in PSNR, SSIM, and LPIPS, enhancing both restoration accuracy and visual quality.