Scholarlink

<p>The Liye Qin bamboo slips are vital for studying Qin society and local administration, but prolonged burial and poor preservation have left many characters blurred or damaged. Because annotation depends heavily on expert paleographic knowledge, labeled data remain scarce. To address these challenges, we propose a task-oriented semi-supervised framework for Liye Qin bamboo-slip character restoration. We construct a dedicated dataset containing 875 labeled image pairs and 10,091 unlabeled degraded images. Within an EMA-based teacher–student scheme, the framework combines weak–strong degradation consistency and soft confidence weighting (SCW) to exploit unlabeled samples and suppress unreliable pseudo-supervision. Experimental results show that the framework outperforms representative baselines, improving PSNR by 0.877 dB and SSIM by 0.019, while reducing LPIPS and FID by 26.67% and 25.89%, respectively. These results demonstrate the effectiveness of the proposed method, providing a practical reference for the digital restoration and preservation of excavated manuscripts.</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

Semi-supervised restoration of damaged characters in Liye Qin bamboo slips

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Shihui Zhou,
  • Ying Zeng,
  • Dongbo Ou,
  • Kai Shi,
  • Haoyin Liu,
  • Jie Tian,
  • Jintian Lu

摘要

The Liye Qin bamboo slips are vital for studying Qin society and local administration, but prolonged burial and poor preservation have left many characters blurred or damaged. Because annotation depends heavily on expert paleographic knowledge, labeled data remain scarce. To address these challenges, we propose a task-oriented semi-supervised framework for Liye Qin bamboo-slip character restoration. We construct a dedicated dataset containing 875 labeled image pairs and 10,091 unlabeled degraded images. Within an EMA-based teacher–student scheme, the framework combines weak–strong degradation consistency and soft confidence weighting (SCW) to exploit unlabeled samples and suppress unreliable pseudo-supervision. Experimental results show that the framework outperforms representative baselines, improving PSNR by 0.877 dB and SSIM by 0.019, while reducing LPIPS and FID by 26.67% and 25.89%, respectively. These results demonstrate the effectiveness of the proposed method, providing a practical reference for the digital restoration and preservation of excavated manuscripts.