Transmission-Lines Foreign Objects Detection Using MSYOLO

Foreign objects detection on transmission lines has the problems that the accuracy is low due to the images captured by UAVs often have small targets, and most models can hardly meet the requirements of real-time detection. Thispaper proposes a novel multiscale-adaptive YOLO-based network model with SIoU loss (MSYOLO) for precise detection of foreign objects on transmission lines. First, an enhanced scale-aware pyramid network which combined dual focus feature aggregation network and a unified feature enhancement module was proposed to improve the detection rate of small targets. Second, SIoU and Fast NMS was introduced to further correct and accelerate the bounding box regression. Experimental results show that our proposed method reaches 86.65 AP50, 46.22%AR and 149 FPS, which is superior to the other anchor-based methods of YOLO series.