Abstract:Transmission tower is an important part of the transmission lines. It is very important to use manual or unmanned aerial vehicle to carry out periodic inspection on the target related to the tower. In order to solve the problems of different tower types, small distance size and image distortion, a tower dependent object detection method based on deep learning framework is proposed. Based on YOLOv5s network, this method uses FReLU activation function instead of SiLU activation function to improve the accuracy of image recognition for tower related objects. SPD-Conv module which is friendly to small target and low resolution is adopted to detect the long distance or small size of target in tower image. At the same time, SPPCSPC space pyramid pool is used to process the target features in batches, which further improves the detection accuracy of tower target. The experimental results show that compared with the original YOLOv5s network, the mAP, Precision and Recall of the improved Yolov5s-FSS network are increased by 2.4%, 0.3% and 1% respectively. The detection performance is significantly improved, which can effectively improve the efficiency of transmission tower inspection.