Scholarlink

<p>The digital restoration of tomb murals is critical for cultural heritage preservation. Existing research focuses on repairing internal lacunae and lacks an effective coordinated restoration solution for cross-scale compound deterioration involving coexisting peripheral information loss and internal degradation. To address this, we propose a restoration of tomb Murals based on Wavelet Convolution and Transformer Self-Attention Collaborative Network. The wavelet branch enhances local structure preservation by explicitly decoupling and reconstructing frequency-domain features, while the Transformer branch establishes long‑range semantic dependencies to ensure globally consistent structural extrapolation. An enhanced feature fusion unit collaboratively suppresses structural distortion and detail blur, achieving pixel-level high-fidelity restoration. Additionally, the multi-scale cross-layer feature aggregation module further strengthens the decoder’s reconstruction capability. Experiments on Tang Dynasty tomb mural fragments with missing peripheral information show our method improves structural fidelity (PSNR) by 3.5%, perceptual quality (LPIPS) by 15.1%, visual authenticity (FID) by 30.5%, demonstrating its effectiveness in restoring complex damage.</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

WCT-Net: joint restoration of tomb murals based on wavelet convolution and transformer self-attention collaborative network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月13日

查看全文
  • Juanjuan Li,
  • Meng Wu,
  • Zhiyong Lu,
  • Lu Wang,
  • Huaidong Zhao,
  • Jian Li

摘要

The digital restoration of tomb murals is critical for cultural heritage preservation. Existing research focuses on repairing internal lacunae and lacks an effective coordinated restoration solution for cross-scale compound deterioration involving coexisting peripheral information loss and internal degradation. To address this, we propose a restoration of tomb Murals based on Wavelet Convolution and Transformer Self-Attention Collaborative Network. The wavelet branch enhances local structure preservation by explicitly decoupling and reconstructing frequency-domain features, while the Transformer branch establishes long‑range semantic dependencies to ensure globally consistent structural extrapolation. An enhanced feature fusion unit collaboratively suppresses structural distortion and detail blur, achieving pixel-level high-fidelity restoration. Additionally, the multi-scale cross-layer feature aggregation module further strengthens the decoder’s reconstruction capability. Experiments on Tang Dynasty tomb mural fragments with missing peripheral information show our method improves structural fidelity (PSNR) by 3.5%, perceptual quality (LPIPS) by 15.1%, visual authenticity (FID) by 30.5%, demonstrating its effectiveness in restoring complex damage.