A Lightweight YOLOv5 Transmission Line Defect Detection Method Based on Coordinate Attention

Abstract

At present, in the power industry, there has always been a demand for intelligent computing and real-time feedback on the edge side using embedded devices. Due to the number of parameters, calculations, and memory usage of the deep learning model, its deployment on edge devices is severely affected. Based on this, this paper proposes a lightweight object detection network based on coordinate attention. The network is based on YOLOv5, decouples the large convolution kernels in the network in channel and space, reduces the parameters of the convolution kernel and the calculation amount of convolution operations, and realizes the lightweight processing of the network. In addition, a lightweight coordinate attention module is introduced into the network, and the model can obtain a larger area of information by embedding position information into the attention map without introducing large overheads, so that the model can increase a small amount of calculation while being significant improve the mAP of the model. The lightweight YOLOv5 model based on coordinate attention makes it possible to deploy on embedded devices with limited resources and achieve better detection results. Lightweight YOLOv5l, YOLOv5m, YOLOv5s, and YOLOv5n reduce FLOPs by about 60.94%, 55.69%, 46.25%, and 46.51%, respectively.