Scholarlink

<p>Silk manuscripts contain important historical information, but their characters are often degraded by fading, blurring, broken strokes, and local structural loss. This paper proposes a receptive-field-enhanced Transformer restoration model for blurred character images from Mawangdui silk manuscripts. Built on a U-Net-style framework, the model introduces a perception former block (PFB) to capture multi-scale stroke context and long-range dependencies, and a skip feature fusion (SFF) module to improve cross-level feature transmission and preserve stroke details. A task-specific dataset with 4202 training images and 427 test images is constructed, and experiments are conducted under random, 10%–35%, and 35%–60% mask settings. Compared with representative restoration methods and a diffusion-based text-image inpainting baseline, the proposed method improves structural restoration and achieves higher PSNR and SSIM under most mask settings. Model complexity, inference time, NIQE, and CLIP-IQA are also reported.</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

Blurred character image restoration in Mawangdui silk manuscripts using receptive field transformer networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Yukang Hua,
  • Ke Ren,
  • Hangbo Zhang,
  • Qi Jia,
  • Bowen Liu,
  • Yanbo Zhang

摘要

Silk manuscripts contain important historical information, but their characters are often degraded by fading, blurring, broken strokes, and local structural loss. This paper proposes a receptive-field-enhanced Transformer restoration model for blurred character images from Mawangdui silk manuscripts. Built on a U-Net-style framework, the model introduces a perception former block (PFB) to capture multi-scale stroke context and long-range dependencies, and a skip feature fusion (SFF) module to improve cross-level feature transmission and preserve stroke details. A task-specific dataset with 4202 training images and 427 test images is constructed, and experiments are conducted under random, 10%–35%, and 35%–60% mask settings. Compared with representative restoration methods and a diffusion-based text-image inpainting baseline, the proposed method improves structural restoration and achieves higher PSNR and SSIM under most mask settings. Model complexity, inference time, NIQE, and CLIP-IQA are also reported.