Scholarlink

<p>To address the issues of low contrast and blurred edges in Dunhuang murals, which often lead to artifacts and edge-detail distortions in restored areas, this study proposes a mural restoration algorithm via the fusion of edge-guided and multi-scale spatial features. First, an encoder extracts low-level features, and an Edge-Gaussian Fusion Block enhances edge details using a rotation-invariant Scharr filter and Gaussian modeling to refine low-confidence features. In the decoding phase, a hybrid pyramid fusion mamba block applies dense spatial pyramid pooling to aggregate multi-scale semantic information, while a Pyramid Fusion Mamba Module reduces redundant semantics for improved feature expressiveness. Finally, a Spatially Enhanced Mamba Module captures long-range dependencies and performs pixel-level restoration. Experiments on the Dunhuang mural dataset show significant improvements: PSNR increases by 0.04–0.67%, SSIM improves by 0.71–0.84%, L1 error reduces by 10.47–20.95%, and LPIPS decreases by 1.18–14.45%.</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

Mural restoration via the fusion of edge-guided and multi-scale spatial features

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月4日

查看全文
  • Zhongmin Liu,
  • Yang Liu,
  • Wenjin Hu

摘要

To address the issues of low contrast and blurred edges in Dunhuang murals, which often lead to artifacts and edge-detail distortions in restored areas, this study proposes a mural restoration algorithm via the fusion of edge-guided and multi-scale spatial features. First, an encoder extracts low-level features, and an Edge-Gaussian Fusion Block enhances edge details using a rotation-invariant Scharr filter and Gaussian modeling to refine low-confidence features. In the decoding phase, a hybrid pyramid fusion mamba block applies dense spatial pyramid pooling to aggregate multi-scale semantic information, while a Pyramid Fusion Mamba Module reduces redundant semantics for improved feature expressiveness. Finally, a Spatially Enhanced Mamba Module captures long-range dependencies and performs pixel-level restoration. Experiments on the Dunhuang mural dataset show significant improvements: PSNR increases by 0.04–0.67%, SSIM improves by 0.71–0.84%, L1 error reduces by 10.47–20.95%, and LPIPS decreases by 1.18–14.45%.