Scholarlink

<p>As natural aging and human activities continue to pose significant challenges to the preservation of the structural consistency and color fidelity of the ancient murals in China’s Dunhuang Grottoes, mural restoration plays a vital role in safeguarding this invaluable cultural and artistic treasure. This paper presents a progressive image inpainting method grounded in Generative Adversarial Networks, effectively addressing the limitations of traditional methods in handling local distortion and incomplete details. Featuring a deformable convolution-based structure generator and global stabilization mechanisms, the proposed framework includes a gated gradient penalty for model stability and an adaptive loss function. Furthermore, the Mural Image Dataset was constructed using 3264 mural images, with data augmentation applied to enhance model robustness. The experimental results demonstrate 22.21% SSIM, 2.06% PSNR, and 12.69% FID improvements over existing methods, with further gains of 52.43%, 6.79%, and 13.38%, respectively, versus the baseline. These consistent metric advancements confirm the method’s effectiveness.</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

Progressive generative mural image restoration based on adversarial structure learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月30日

查看全文
  • Jiansheng Guan,
  • Han Li,
  • Xiaolong Cai,
  • Enyi Chen,
  • Jinjiang Lin,
  • Zhicai Ding,
  • Yuting Ni

摘要

As natural aging and human activities continue to pose significant challenges to the preservation of the structural consistency and color fidelity of the ancient murals in China’s Dunhuang Grottoes, mural restoration plays a vital role in safeguarding this invaluable cultural and artistic treasure. This paper presents a progressive image inpainting method grounded in Generative Adversarial Networks, effectively addressing the limitations of traditional methods in handling local distortion and incomplete details. Featuring a deformable convolution-based structure generator and global stabilization mechanisms, the proposed framework includes a gated gradient penalty for model stability and an adaptive loss function. Furthermore, the Mural Image Dataset was constructed using 3264 mural images, with data augmentation applied to enhance model robustness. The experimental results demonstrate 22.21% SSIM, 2.06% PSNR, and 12.69% FID improvements over existing methods, with further gains of 52.43%, 6.79%, and 13.38%, respectively, versus the baseline. These consistent metric advancements confirm the method’s effectiveness.