Improved insulator defect detection network considering target characteristics and sample location classification

Abstract Insulators are crucial components of the power system. An enhanced insulator detection network, based on YOLOv8, addresses unequal training samples, inadequate target localization and classification accuracy in the existing insulator unmanned aerial vehicle inspection algorithm. Firstly, the ADown down-sampling component and DynamicConv are incorporated into the backbone network to enhance feature representation. Secondly, Focal-IoU and Adaptive Training Sample Selection are used during training to adjust the weight of each sample based on their quantity and difficulty level, enhancing focus on rare and challenging targets. Finally, to address difficult target localization and classification, design a task-aligned detection head called ‘Align Head’ to strengthen the link between localization and classification branches. Experiments show that the proposed method increases mAP@0.5 (mean average precision at a threshold of 0.50) by 7.5% over the baseline, with an FPS of 81.57, demonstrating superior performance.