Scholarlink

<p>Bamboo and wooden slips were the primary writing media from the Warring States period to the Han Dynasty and are an important source for studying ancient Chinese political, legal, and intellectual history. Super-resolution (SR) reconstruction of slip images is an essential preprocessing step for character restoration and content recognition. To address this task, we propose a dual-stream disentangled SR model, termed DSF-SRFormer, which separately models character structures and background textures. At the character level, a Directional Stroke Attention Block (DSAB) captures stroke-aligned features using multi-directional convolutional kernels corresponding to dominant writing directions. For full-slip reconstruction, a Long Strip Attention Block (LSAB) models long-range dependencies to enhance contextual consistency in vertically arranged texts. An edge-aware loss combined with perceptual loss optimizes stroke sharpness and visual quality. Experimental results show that DSF-SRFormer improves performance on both character-level and full-slip SR tasks in terms of PSNR, SSIM, and LPIPS, producing faithful stroke structures.</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

Dual-stream disentangled super-resolution for bamboo slip text images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月9日

查看全文
  • Wenhao Wang,
  • Tiejun Wang,
  • Xiaoyan Hu,
  • Chengjie Xu,
  • Hualong Du,
  • Lin Tao,
  • Chengjie Zhou

摘要

Bamboo and wooden slips were the primary writing media from the Warring States period to the Han Dynasty and are an important source for studying ancient Chinese political, legal, and intellectual history. Super-resolution (SR) reconstruction of slip images is an essential preprocessing step for character restoration and content recognition. To address this task, we propose a dual-stream disentangled SR model, termed DSF-SRFormer, which separately models character structures and background textures. At the character level, a Directional Stroke Attention Block (DSAB) captures stroke-aligned features using multi-directional convolutional kernels corresponding to dominant writing directions. For full-slip reconstruction, a Long Strip Attention Block (LSAB) models long-range dependencies to enhance contextual consistency in vertically arranged texts. An edge-aware loss combined with perceptual loss optimizes stroke sharpness and visual quality. Experimental results show that DSF-SRFormer improves performance on both character-level and full-slip SR tasks in terms of PSNR, SSIM, and LPIPS, producing faithful stroke structures.