Scholarlink

<p>Natural aging and human activities have eroded the Dunhuang murals’ structural integrity and color fidelity; restoration aims to reconstruct their original appearance. Building on a prior GAN-based progressive restoration framework, this paper introduces a synergistic mechanism integrating structural features with spatial-frequency domains to resolve color misalignment caused by decoupling color from structure in traditional methods. This method also serves as a digital guide for manual restoration. The framework includes a structural feature-based color generation module with three components: spatial-frequency dimensionality reduction, an attention-based U-Net backbone, and cross-scale, multi-dimensional feature fusion. A dataset of 6800 mural images and matched line art is released. Experiments show improvements over state-of-the-art methods (1.99%, 16.18%, 42.04% in SSIM, PSNR, LPIPS) and over the baseline (1.38%, 6.51%, 10.98%), confirming efficacy.</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

PGMv2: Structural-spatial-frequency contrastive reduction for mural high-fidelity restoration

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月22日

查看全文
  • Han Li,
  • Donglin Liu,
  • Yuan Lu,
  • Qinfei Chen,
  • Xiaoyan Chen,
  • Jiangyin Huang,
  • Yuan Sun,
  • Xungao Zhong,
  • Duanxi Zheng,
  • Shaobo Kang,
  • Zhicai Ding,
  • Jiansheng Guan

摘要

Natural aging and human activities have eroded the Dunhuang murals’ structural integrity and color fidelity; restoration aims to reconstruct their original appearance. Building on a prior GAN-based progressive restoration framework, this paper introduces a synergistic mechanism integrating structural features with spatial-frequency domains to resolve color misalignment caused by decoupling color from structure in traditional methods. This method also serves as a digital guide for manual restoration. The framework includes a structural feature-based color generation module with three components: spatial-frequency dimensionality reduction, an attention-based U-Net backbone, and cross-scale, multi-dimensional feature fusion. A dataset of 6800 mural images and matched line art is released. Experiments show improvements over state-of-the-art methods (1.99%, 16.18%, 42.04% in SSIM, PSNR, LPIPS) and over the baseline (1.38%, 6.51%, 10.98%), confirming efficacy.