Scholarlink

<p>Unmanned aerial vehicle (UAV)-based power line inspections often face significant challenges in detecting small defects due to their limited size and the complexity of surrounding backgrounds. This paper proposes a new approach that integrates a graph neural network (GNN)-based feature extractor with a cross-scale dual attention detector to address these issues, called Graph-Enhanced YOLO (GEYOLO). The proposed method comprises a backbone network for feature extraction, a neck network for multi-scale feature fusion, and a prediction head for final detection. By leveraging graph neural network and a dual-attention mechanism, the model enhances the detection of small objects in complex environments, resulting in improved accuracy and robustness. GEYOLO achieves an mAP of 77.0%, outperforming the second-best model (CEH-YOLO, 72.4%) by 4.6% while maintaining real-time inference at 22.3 ms per image (≈ 45 FPS), close to YOLOv8 (19.8 ms). This balance of precision and efficiency demonstrates GEYOLO’s superiority for small-defect UAV inspection. This approach significantly enhances the efficiency and reliability of UAV-based inspections, offering a practical and valuable solution for automated infrastructure maintenance.</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

Improving small object detection in power line inspection with graph-enhanced YOLO

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Jiangtao Shi,
  • Wenqi Fan,
  • Zhi Yang,
  • Junhui Li,
  • Mengxuan Li

摘要

Unmanned aerial vehicle (UAV)-based power line inspections often face significant challenges in detecting small defects due to their limited size and the complexity of surrounding backgrounds. This paper proposes a new approach that integrates a graph neural network (GNN)-based feature extractor with a cross-scale dual attention detector to address these issues, called Graph-Enhanced YOLO (GEYOLO). The proposed method comprises a backbone network for feature extraction, a neck network for multi-scale feature fusion, and a prediction head for final detection. By leveraging graph neural network and a dual-attention mechanism, the model enhances the detection of small objects in complex environments, resulting in improved accuracy and robustness. GEYOLO achieves an mAP of 77.0%, outperforming the second-best model (CEH-YOLO, 72.4%) by 4.6% while maintaining real-time inference at 22.3 ms per image (≈ 45 FPS), close to YOLOv8 (19.8 ms). This balance of precision and efficiency demonstrates GEYOLO’s superiority for small-defect UAV inspection. This approach significantly enhances the efficiency and reliability of UAV-based inspections, offering a practical and valuable solution for automated infrastructure maintenance.