Foreign object detection algorithm for high voltage transmission lines incorporating atrous convolution

Abstract

There are problems of complex background, serious occlusion, and low accuracy of foreign object detection on high-voltage transmission lines. In this paper, we propose an algorithm incorporating atrous convolution for the detection of four types of foreign objects that are more likely to appear on high-voltage transmission lines: bird nests, balloons, kites, and rubbish. The algorithm is based on the original YOLOv4 model, replacing the original SPP structure with an ASPP structure. By using the atrous convolution instead of the original pooling operation, the network is able to have a larger field of perception during feature extraction, while the resolution of the feature map is not degraded too much to avoid losing too much information. The experimental results show that the accuracy of this method reaches 93.73% on the transmission line dataset used, which is a significant improvement over the original YOLOv4 model and improves the detection capability of foreign object targets that are obscured in complex environments.