Deep Learning Based Defect Detection Method for Overhead Transmission Wires

Abstract

Transmission line is the carrier of power transmission, in order to more accurately detect the transmission conductor is susceptible to the influence of the external environment and cause the conductor to break the strand, loose strand to the cable foreign body hanging and other defects. In this paper, a lightweight transmission conductor defect detection algorithm named HorCM_PAM_YOLOv5 based on YOLOv5 is proposed. Firstly, in order to increase the algorithm's ability to spatially interact with different features as well as its detection accuracy, this paper designs a lightweight recursive convolution module HorCM with reference to HorNet to enhance the expression of the algorithm's model. Then in order to cope with the problem of complex background and difficult detection of aerial images, this paper proposes a lightweight parallel attention mechanism module (PAM), so that the defective image channel feature extraction and spatial feature extraction can be independent of each other, which reduces the interference of the background and increases the image's characterisation ability. Afterwards, in order to cope with the problem of large differences in the scale of wire defect images taken by aerial photography, this paper proposes for the MPDIOU loss function, using the distance between the prediction frame and the actual frame of the identified points to minimise the processing, to improve the accuracy of the model and convergence speed. Finally, in order to better reflect the actual engineering application scenarios, our dataset uses UAS machine patrol images from Yunnan Power Supply Bureau Company. Experimental simulations show that with a 3.1% increase in detection speed relative to YOLOv5s and a 5% reduction in model volume, the improved algorithm HorCM_PAM_YOLOv5 still manages to increase its overall performance by 9.8% over YOLOv5s, and its accuracy by 7.2% over YOLOv5s.