Scholarlink

<p>Digital preservation of historic murals is essential for protecting cultural heritage. Despite centuries of damage, advances in inpainting offer new restoration possibilities. However, existing methods often distort features like color and texture, and suffer from significant pixel-level blurring. We propose a Coordinated Attention Aggregation Transformation (CAAT) GAN architecture with U-Net discriminators to address these limitations. The CAAT generator extracts contextual information from distant regions via a Coordinated Aggregation Transformation Block, expanding the receptive field and improving content inference in missing areas to restore original color and texture. The U-Net discriminator further refines results by providing both global and local confidence scores. We also introduce DunHuang-Mural, a dataset of 7983 high-resolution historical murals. Trained on 6386 images and evaluated on 1597, our CAUGAN achieves significant gains in visual fidelity and structural consistency over existing methods, demonstrating its utility for archeological mural restoration.</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

Supporting historic mural image inpainting by using coordinate attention aggregated transformations with U-Net-based discriminator

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月28日

查看全文
  • Junjie Zhang,
  • Shuang Bai,
  • Xianyi Zeng,
  • Kaixuan Liu,
  • Hua Yuan

摘要

Digital preservation of historic murals is essential for protecting cultural heritage. Despite centuries of damage, advances in inpainting offer new restoration possibilities. However, existing methods often distort features like color and texture, and suffer from significant pixel-level blurring. We propose a Coordinated Attention Aggregation Transformation (CAAT) GAN architecture with U-Net discriminators to address these limitations. The CAAT generator extracts contextual information from distant regions via a Coordinated Aggregation Transformation Block, expanding the receptive field and improving content inference in missing areas to restore original color and texture. The U-Net discriminator further refines results by providing both global and local confidence scores. We also introduce DunHuang-Mural, a dataset of 7983 high-resolution historical murals. Trained on 6386 images and evaluated on 1597, our CAUGAN achieves significant gains in visual fidelity and structural consistency over existing methods, demonstrating its utility for archeological mural restoration.