Abstract: [Objectives] The traditional power grid inspection method has problems such as high labor intensity and low efficiency, so UAV(Unmanned Aerial Vehicle) inspection has become one of the development directions of intelligent power grid inspection. [Methods] This paper proposes a UAV inspection algorithm based on lightweight deep learning network YOLOv5-Mv3 for detecting grid insulators and their foreign objects, by using Shandong Golden Power Grid as the research object. Firstly, the dataset is constructed by taking pictures of the grid inspection by UAV, and the dataset is trained. Secondly, for the grid insulators and their foreign objects, Mobilenetv3 is used to replace CSPDarknet53 as the feature extraction network, and YOLOv5-Mv3 is improved by lightweighting to reduce the model parameters and the amount of computation, so that it can ensure the accuracy and meet the real-time detection requirements. [Results] Finally, experiments show that the mAP value of this paper's algorithm reaches 84.7%, and the actual detection frame rate can be up to 56.6, and the improved YOLOv5-Mv3 has higher detection accuracy and faster detection speed compared to the Faster RCNN, SSD, and YOLOv4 models. [Conclusions] The algorithm improves the efficiency of the UAV inspecting the power grid, realizes the requirement of lightweight and high efficiency, and is more in line with the power grid intelligent inspection requirements.

Key words: unmanned aerial vehicle, YOLO algorithm, insulator, target detection, transmission line, power inspection, machine vision, image processing