Scholarlink

<p>Under severe noise interference and structural erosion in oracle bone rubbing images, discriminative glyph structures are often incomplete and unstable, making existing object detection methods highly vulnerable to complex backgrounds and structural degradation. To address this challenge, this paper proposes SGD-Det, a structure-guided oracle bone inscription detection framework that explicitly incorporates glyph structural priors into the detection process. The proposed method introduces a glyph-structure-guided attention mechanism that progressively models glyph boundary perception, multi-scale structural relationships, and structure–semantic collaboration, enabling the detector to focus on structurally informative regions even under severe degradation. In addition, a high-resolution feature representation is introduced to preserve fine-grained stroke details and improve detection stability for densely distributed and small-scale glyph targets. By combining structure-guided modeling with training strategies tailored to rubbing characteristics, the framework significantly enhances robustness to noise interference and structural damage. Extensive experiments on an oracle bone inscription dataset show that the proposed method achieves 84.8% mAP@50 and 44.5% mAP@50:95, while maintaining a lightweight model size of 3.3M parameters. Compared with strong recent baselines, SGD-Det delivers more stable detection under severe structural degradation and complex rubbing noise, demonstrating clear advantages in preserving fine-grained stroke details and improving localization robustness for small and densely distributed oracle characters. These results indicate that explicitly modeling glyph structural priors provides a more reliable and practically effective solution for oracle bone detection in degraded historical documents.</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

Sgd-det:structure-guided oracle character detection in degraded rubbing images

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Zixuan Wang,
  • Xuxing Qi,
  • Cheng Xu,
  • Zhen Liu,
  • Jiazheng Yuan,
  • Bofeng Mo,
  • Lu Chen,
  • Hongzhe Liu

摘要

Under severe noise interference and structural erosion in oracle bone rubbing images, discriminative glyph structures are often incomplete and unstable, making existing object detection methods highly vulnerable to complex backgrounds and structural degradation. To address this challenge, this paper proposes SGD-Det, a structure-guided oracle bone inscription detection framework that explicitly incorporates glyph structural priors into the detection process. The proposed method introduces a glyph-structure-guided attention mechanism that progressively models glyph boundary perception, multi-scale structural relationships, and structure–semantic collaboration, enabling the detector to focus on structurally informative regions even under severe degradation. In addition, a high-resolution feature representation is introduced to preserve fine-grained stroke details and improve detection stability for densely distributed and small-scale glyph targets. By combining structure-guided modeling with training strategies tailored to rubbing characteristics, the framework significantly enhances robustness to noise interference and structural damage. Extensive experiments on an oracle bone inscription dataset show that the proposed method achieves 84.8% mAP@50 and 44.5% mAP@50:95, while maintaining a lightweight model size of 3.3M parameters. Compared with strong recent baselines, SGD-Det delivers more stable detection under severe structural degradation and complex rubbing noise, demonstrating clear advantages in preserving fine-grained stroke details and improving localization robustness for small and densely distributed oracle characters. These results indicate that explicitly modeling glyph structural priors provides a more reliable and practically effective solution for oracle bone detection in degraded historical documents.