Scholarlink

<p>The Dunhuang murals, an integral part of Chinese cultural heritage, face severe degradation due to natural factors, posing challenges for inpainting. Traditional methods often fail to address these complexities, struggling with global semantic understanding, fine-detail reconstruction, and computational efficiency. To overcome these limitations, we propose MWT-Net (Mamba Prior and Wavelet-Sparse Transformer Network), which combines Mamba priors with a wavelet-based sparse Transformer in a dual-branch architecture. The prior-guided branch captures long-range dependencies and global context via state-space modeling, while the main branch uses multi-scale sparse modeling for enhanced texture reconstruction. A GDT-FF module improves attention expressiveness, preserves fine details, and reduces redundancy. Experiments show that MWT-Net outperforms existing approaches on multiple benchmarks, achieving a 2.67 dB gain in PSNR, a 9.60% increase in SSIM, and reductions of 0.1113 and 0.0250 in LPIPS and FID. These results highlight the potential of MWT-Net to produce coherent and realistic mural inpainting, contributing to cultural heritage preservation. Our code and dataset are available at <a href="https://github.com/github662/dunhuang.">https://github.com/github662/dunhuang.</a></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

Enhancing mural inpainting: a dual-branch approach integrating mamba priors and wavelet-sparse transformers

  • 期刊论文

  • 发表日期：

    2026年3月16日

查看全文
  • Zhigang Xu,
  • Jie Li

摘要

The Dunhuang murals, an integral part of Chinese cultural heritage, face severe degradation due to natural factors, posing challenges for inpainting. Traditional methods often fail to address these complexities, struggling with global semantic understanding, fine-detail reconstruction, and computational efficiency. To overcome these limitations, we propose MWT-Net (Mamba Prior and Wavelet-Sparse Transformer Network), which combines Mamba priors with a wavelet-based sparse Transformer in a dual-branch architecture. The prior-guided branch captures long-range dependencies and global context via state-space modeling, while the main branch uses multi-scale sparse modeling for enhanced texture reconstruction. A GDT-FF module improves attention expressiveness, preserves fine details, and reduces redundancy. Experiments show that MWT-Net outperforms existing approaches on multiple benchmarks, achieving a 2.67 dB gain in PSNR, a 9.60% increase in SSIM, and reductions of 0.1113 and 0.0250 in LPIPS and FID. These results highlight the potential of MWT-Net to produce coherent and realistic mural inpainting, contributing to cultural heritage preservation. Our code and dataset are available at https://github.com/github662/dunhuang.