Scholarlink

<p>The study of ancient Chinese bamboo slips from the Warring States period is a crucial task in digital humanities, yet it is hampered by severe background noise and long-tailed character distributions. To address these issues, we propose a framework centered on a Self-supervised Bilateral Branch Network (SBBN). The framework makes three key contributions: (1) We establish zgzj1001, a benchmark dataset for this task. (2) We design a Spatial Self-Attention module to specifically enhance character features while actively suppressing background interference. (3) SBBN effectively overcomes the class imbalance problem by integrating a self-supervised contrastive learning branch to learn robust representations of rare tail classes. Experiments on zgzj1001 show that our method achieves a 96.31% balanced accuracy, offering a competitive and practical solution compared to baseline models. This work provides a reliable tool for paleographic research and a functional paradigm for cultural heritage digitization.</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

Recognition of Warring States bamboo slip characters via a self-supervised bilateral branch network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月20日

查看全文
  • Wenbo Jiang,
  • Wenda Xie,
  • Yidi Yang,
  • Lirui Liu,
  • Rennian Wang

摘要

The study of ancient Chinese bamboo slips from the Warring States period is a crucial task in digital humanities, yet it is hampered by severe background noise and long-tailed character distributions. To address these issues, we propose a framework centered on a Self-supervised Bilateral Branch Network (SBBN). The framework makes three key contributions: (1) We establish zgzj1001, a benchmark dataset for this task. (2) We design a Spatial Self-Attention module to specifically enhance character features while actively suppressing background interference. (3) SBBN effectively overcomes the class imbalance problem by integrating a self-supervised contrastive learning branch to learn robust representations of rare tail classes. Experiments on zgzj1001 show that our method achieves a 96.31% balanced accuracy, offering a competitive and practical solution compared to baseline models. This work provides a reliable tool for paleographic research and a functional paradigm for cultural heritage digitization.