Scholarlink

<p>Manhole covers are a critical component of urban infrastructure, and their damage poses significant threats to road safety and structural integrity. Real-time detection and tagging of manhole covers enable timely maintenance, reducing traffic risks and improving infrastructure reliability. To address this issue, we propose an improved manhole cover hazard detection algorithm, DBG-YOLO, based on the YOLOv8n framework. The proposed DBG-YOLO model integrates a Dilated Reparam Block (DRB) into the C2f module of the backbone network, enhancing both the receptive field and feature representation capabilities. This optimization significantly improves detection accuracy for small objects and complex scenarios. When integrated with comprehensive data augmentation strategies, the framework demonstrates exceptional adaptability to low-light conditions, achieving robust detection even under insufficient illumination. For feature fusion, the neck network incorporates a Bidirectional Feature Pyramid Network (BiFPN) and a Global Attention Mechanism (GAM), forming an advanced multi-scale architecture that effectively addresses partial occlusion challenges. By replacing the conventional loss function with SlideLoss, the model further refines bounding box regression accuracy, ensuring precise localization in demanding environments. Collectively, these innovations, synergized with adaptive data augmentation, provide a holistic solution to detection limitations in low-light and occluded scenarios. Experimental results demonstrate that DBG-YOLO achieves superior performance with 93.6% mAP@50 at 167.5 FPS, outperforming both Faster R-CNN (81.4%/13.4 FPS) and YOLOv5s (91.3%/142.7 FPS) in terms of accuracy and inference speed. This lightweight architecture establishes a new state-of-the-art balance between detection accuracy, computational efficiency, and model complexity. This innovative approach provides robust technical support for real-time detection tasks, enhancing urban infrastructure monitoring, improving road safety, and contributing to the development of smart cities.</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

DBG-YOLO: Efficient detection of hidden dangers of manhole covers based on deep learning YOLO network

  • 期刊论文

  • 发表日期：

    2025年6月21日

查看全文
  • Dingbang Guo,
  • Piaorong Xu,
  • Minxi Cai,
  • Exian Liu,
  • Meng Wang,
  • Zhengping Shan,
  • Feng Jiang

摘要

Manhole covers are a critical component of urban infrastructure, and their damage poses significant threats to road safety and structural integrity. Real-time detection and tagging of manhole covers enable timely maintenance, reducing traffic risks and improving infrastructure reliability. To address this issue, we propose an improved manhole cover hazard detection algorithm, DBG-YOLO, based on the YOLOv8n framework. The proposed DBG-YOLO model integrates a Dilated Reparam Block (DRB) into the C2f module of the backbone network, enhancing both the receptive field and feature representation capabilities. This optimization significantly improves detection accuracy for small objects and complex scenarios. When integrated with comprehensive data augmentation strategies, the framework demonstrates exceptional adaptability to low-light conditions, achieving robust detection even under insufficient illumination. For feature fusion, the neck network incorporates a Bidirectional Feature Pyramid Network (BiFPN) and a Global Attention Mechanism (GAM), forming an advanced multi-scale architecture that effectively addresses partial occlusion challenges. By replacing the conventional loss function with SlideLoss, the model further refines bounding box regression accuracy, ensuring precise localization in demanding environments. Collectively, these innovations, synergized with adaptive data augmentation, provide a holistic solution to detection limitations in low-light and occluded scenarios. Experimental results demonstrate that DBG-YOLO achieves superior performance with 93.6% mAP@50 at 167.5 FPS, outperforming both Faster R-CNN (81.4%/13.4 FPS) and YOLOv5s (91.3%/142.7 FPS) in terms of accuracy and inference speed. This lightweight architecture establishes a new state-of-the-art balance between detection accuracy, computational efficiency, and model complexity. This innovative approach provides robust technical support for real-time detection tasks, enhancing urban infrastructure monitoring, improving road safety, and contributing to the development of smart cities.