Scholarlink

<p>Focusing on the cultural heritage of the Nanjing Ming Dynasty city wall, this study presents CWADE-Net, a deep learning-based city wall anomaly detection framework specially designed to detect surface defects arising from herbaceous/woody and vine-type vegetation invasion, as well as brick spalling. Its novelty lies in the capability to address challenging conditions such as uneven illumination, complex backgrounds, and large defect-scale variations. CWADE-Net jointly integrates illumination enhancement, edge information encoding, and spatial-frequency feature extraction in its backbone to improve feature representation. Its neck employs bidirectional feature fusion to enhance multi-scale semantic interaction. Moreover, we adopt a lightweight detection head that enables real-time model deployment. Experiments on images acquired using a Nikon D300, iPhone 15 Pro Max, and DJI Matrice 4E demonstrate mAP50 scores of 82.4%, 87.9%, and 54.8% for three defect types, outperforming mainstream methods by 5-12%, thus effectively supporting intelligent monitoring, conservation, and World Cultural Heritage nomination efforts.</p><p></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

CWADE-Net: a deep learning framework for vegetation invasion and brick spalling defect detection on Nanjing Ming City Wall

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月28日

查看全文
  • Xianglong Yuan,
  • Nannan Wang,
  • Yuliang Wang,
  • Shenglan Du,
  • Mingming Sui,
  • Yueqian Shen,
  • Shihuan Li,
  • Ziyou Wang,
  • Dong Chen,
  • Jiju Poovvancheri,
  • Liqiang Zhang

摘要

Focusing on the cultural heritage of the Nanjing Ming Dynasty city wall, this study presents CWADE-Net, a deep learning-based city wall anomaly detection framework specially designed to detect surface defects arising from herbaceous/woody and vine-type vegetation invasion, as well as brick spalling. Its novelty lies in the capability to address challenging conditions such as uneven illumination, complex backgrounds, and large defect-scale variations. CWADE-Net jointly integrates illumination enhancement, edge information encoding, and spatial-frequency feature extraction in its backbone to improve feature representation. Its neck employs bidirectional feature fusion to enhance multi-scale semantic interaction. Moreover, we adopt a lightweight detection head that enables real-time model deployment. Experiments on images acquired using a Nikon D300, iPhone 15 Pro Max, and DJI Matrice 4E demonstrate mAP50 scores of 82.4%, 87.9%, and 54.8% for three defect types, outperforming mainstream methods by 5-12%, thus effectively supporting intelligent monitoring, conservation, and World Cultural Heritage nomination efforts.