Scholarlink

<p>The Dunhuang manuscripts, a pivotal legacy of the Silk Road, are preserved mainly as fragmented pieces, presenting substantial challenges for reassembly. Currently, fragment matching relies heavily on expert judgment, with handwriting style consistency as a key visual clue. However, this approach suffers from low efficiency and entails considerable temporal demands. To address this issue, we propose a hybrid network model that integrates a MobileNet-based convolutional branch for local feature extraction with a Transformer branch for capturing global contextual patterns, thereby simulating the expert visual comparison process. Building upon the handwriting style matching strategy, we reformulate the matching task into a consistency classification problem, effectively achieving reduced complexity and improved interpretability. We constructed the DSE-Rejoin dataset with 63,608 high-resolution images of Dunhuang manuscript fragments featuring diverse handwriting styles and damage levels. Experiments on this dataset achieved 95.17% accuracy, 95.06% precision, 95.17% recall, and 95.06% F1 score.</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

Dunhuang manuscript fragment reassembly based on patch-level handwriting style recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月10日

查看全文
  • Yutong Zheng,
  • Langtai Cheng,
  • Tieshan Zhang,
  • Jiaqi Dai,
  • Zeli Tong,
  • Mingkun Chen,
  • Yanping Xiang,
  • Xuan Liu,
  • Xuelong Li

摘要

The Dunhuang manuscripts, a pivotal legacy of the Silk Road, are preserved mainly as fragmented pieces, presenting substantial challenges for reassembly. Currently, fragment matching relies heavily on expert judgment, with handwriting style consistency as a key visual clue. However, this approach suffers from low efficiency and entails considerable temporal demands. To address this issue, we propose a hybrid network model that integrates a MobileNet-based convolutional branch for local feature extraction with a Transformer branch for capturing global contextual patterns, thereby simulating the expert visual comparison process. Building upon the handwriting style matching strategy, we reformulate the matching task into a consistency classification problem, effectively achieving reduced complexity and improved interpretability. We constructed the DSE-Rejoin dataset with 63,608 high-resolution images of Dunhuang manuscript fragments featuring diverse handwriting styles and damage levels. Experiments on this dataset achieved 95.17% accuracy, 95.06% precision, 95.17% recall, and 95.06% F1 score.