Scholarlink

<p>Ancient textiles embody rich historical information but often suffer from pattern loss and structural degradation due to long-term aging and wear. Although digital technologies provide non-contact solutions for their preservation, existing restoration methods still struggle to maintain structural continuity and reconstruct fine-grained textures. To address these issues, this study proposes a structure-guided image restoration framework for ancient textiles based on a CNN-Transformer architecture. The proposed method decouples structural and texture modeling into cooperative subnetworks, effectively constraining the global pattern layout and enhancing the local texture consistency during the image completion process. Experimental results on both simulated data and real-world ancient textile images demonstrate that the proposed method outperforms several existing approaches. For a missing rate of 20%, the proposed method achieves a PSNR of 35.65 dB and an SSIM of 0.975, while maintaining a stable performance under higher missing-rate scenarios.</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

Digital restoration for textile cultural relics via a structure-guided neural network based on CNN-Transformer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月20日

查看全文
  • Wenting Yang,
  • Binjie Xin,
  • Feifei He,
  • Jiyuan Liu,
  • Md All Amin Newton

摘要

Ancient textiles embody rich historical information but often suffer from pattern loss and structural degradation due to long-term aging and wear. Although digital technologies provide non-contact solutions for their preservation, existing restoration methods still struggle to maintain structural continuity and reconstruct fine-grained textures. To address these issues, this study proposes a structure-guided image restoration framework for ancient textiles based on a CNN-Transformer architecture. The proposed method decouples structural and texture modeling into cooperative subnetworks, effectively constraining the global pattern layout and enhancing the local texture consistency during the image completion process. Experimental results on both simulated data and real-world ancient textile images demonstrate that the proposed method outperforms several existing approaches. For a missing rate of 20%, the proposed method achieves a PSNR of 35.65 dB and an SSIM of 0.975, while maintaining a stable performance under higher missing-rate scenarios.