Detecting Broken Glass Insulators for Automated UAV Power Line Inspection Based on an Improved YOLOv8 Model

Intelligent drone inspection of power lines using computer vision techniques is a very promising research topic. In particular, the detection of broken glass insulators due to their essential role in the proper functioning of electrical transmission lines. However, detection of such defects is challenging due to their small size coupled with complex aerial image backgrounds and limited dataset availability. In this regard, this paper uses advanced object detection algorithms to achieve real-time monitoring of broken glass insulators using drones by offering main contributions on two aspects. First, a large-scale aerial image dataset from Vietnam was meticulously constructed, including 1,010 labeled original images. It will serve as a valuable resource for those engaged in the automation of power line inspections. Second, a novel improved Yolov8 detection model is introduced, incorporating the Gather and Distribute mechanism for the fusion of different feature maps within the Neck component of Yolov8. Experimental results demonstrate that the model surpasses equivalent-scale models, such as Yolov8-m, Yolov7, RTDETR-l, Gold-Yolo and Yolov6-v3.0-m, exhibiting notable improvements. The model shows improvements of up to 2,1% in mean Average Precision mAP:50 while maintaining resource efficiency, with a smaller model size of at least 17.31% compared to other models, and it exhibits real-time performance capabilities. Overall, these promising contributions could advance the development of intelligent, reliable and automated anomaly detection systems suitable for drone inspections of power lines in the context of industry 4.0. Access to both the source code and the Vietnamese dataset is available through this link: https://github.com/phd-benel/Yolov8_gold .