Foreign Object Detection Method in Aerial Images of Power Transmission Lines Based on Scale Adaptive YOLOv5

Foreign object detection is an important part of power inspection work. It aims to investigate objects that have not yet invaded transmission lines but are likely to threaten the operation of transmission lines. However, the complex environment of high-voltage transmission lines and the variable scale of foreign objects makes it difficult. Therefore, a Scale Adaptive You Only Look Oncev5 (SA-YOLOv5) based foreign object detection method is proposed. Unmanned aerial vehicles collect aerial images containing two types of foreign objects: plastic film and house to build a dataset. The StemBlock module is introduced to downsample shallow features, reducing calculation parameters while maintaining strong feature expression capabilities. The scale-adaptive residuals module is designed to fit foreign object features with large scale differences. Finally, migrating pre-trained model weights strategy is adapted to accelerate model fitting. Experimental results prove that SA-YOLOv5 locates and identifies foreign objects more accurately, and meets the accuracy requirements for inspection of foreign objects in transmission lines to a certain extent.