Scholarlink

<p>Oracle Bone Inscriptions (OBIs) are invaluable primary sources for studying the Shang Dynasty. However, effective retrieval of OBIs character images from literature remains challenging due to the massive volume of data, complex character variations, and limited textual annotations. This paper proposes a novel image-based retrieval method for OBIs literature by integrating deep learning with vector database techniques. Specifically, ResNet-50 enhanced with Atrous Spatial Pyramid Pooling (ASPP) is employed to extract robust multi-scale features from complex rubbing images. The L2 normalized feature vectors are then stored in a FAISS vector database for efficient inner-product similarity matching. For evaluation, we constructed the OBIs-DIset, a dedicated dataset containing about 17,000 images extracted from real OBIs literature. Experimental results show that the proposed method achieves a mAP of 85.94% and a retrieval time of 56.25 ms, significantly outperforming baseline approaches.</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

A deep learning method for image-based retrieval of oracle bone inscriptions literature

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月28日

查看全文
  • Zhaoan Dong,
  • Boyong Wang,
  • Jing Xiong,
  • Bang Li,
  • Xiaofan Wang

摘要

Oracle Bone Inscriptions (OBIs) are invaluable primary sources for studying the Shang Dynasty. However, effective retrieval of OBIs character images from literature remains challenging due to the massive volume of data, complex character variations, and limited textual annotations. This paper proposes a novel image-based retrieval method for OBIs literature by integrating deep learning with vector database techniques. Specifically, ResNet-50 enhanced with Atrous Spatial Pyramid Pooling (ASPP) is employed to extract robust multi-scale features from complex rubbing images. The L2 normalized feature vectors are then stored in a FAISS vector database for efficient inner-product similarity matching. For evaluation, we constructed the OBIs-DIset, a dedicated dataset containing about 17,000 images extracted from real OBIs literature. Experimental results show that the proposed method achieves a mAP of 85.94% and a retrieval time of 56.25 ms, significantly outperforming baseline approaches.