A real-time insulator condition detection model for UAV inspection based on FG-YOLO

Abstract In the field of target detection, especially for UAV inspection insulator state detection, accurate identification of small targets and complex background environment is always a major challenge. Traditional methods solve this problem by strengthening feature acquisition, but ignore the increase of computational complexity and resource consumption, resulting in insufficient hardware resources of UAV. We innovatively enhance YOLOv8-N model comprehensively and propose a Feature Guided-You Only Look Once (FG-YOLO), which effectively solves the problems of background noise interference and insufficient global information capture by designing context anchor concat and C2 Locality-Aware Attention (C2fLA) modules. Secondly, we introduce lightweight neck networks such as SCDown, DySample and lightweight guided convolutional detection detector head to maintain the original performance of the model, while reducing the model parameters and computational complexity and enhancing the robustness of features. In addition, we design global channel directed attention mechanism module to improve the sensing ability of network to size targets through multi-scale feature fusion. Experiments show that the improved FG-YOLO has excellent performance and potential in real-time detection and feature capture in three public datasets, with mAP50 reaching 88.2, 99.6 and 99.8 respectively. Moreover, FG-YOLO has inference speed of 25 frames s −1 in edge device MAIX-II Axera-Pi, meeting the real-time detection requirements of insulator defects.