Abstract:

In view of the low detection accuracy of multi-scale defect targets in transmission lines under complex background, an improved YOLOv7(You Only Look Once version7) defect target detection model for transmission lines is proposed. To solve the problem of low defect targets caused by complex background, an improved Swin Transformer module is introduced in the Backbone part to improve the detection accuracy of the model using multi-head attention mechanism to improve the effect of global feature extraction. According to the multi-scale characteristics of the target to be detected, an adaptive feature fusion module is introduced on the basis of the feature pyramid to improve the detection ability of the Neck partial feature fusion network on multiple defect targets of different scales. SIoU(Structured Intersection over Union) loss function is used to improve the accuracy of prediction frame regression and accelerate the model convergence. Experimental results show that compared with YOLOv5, YOLOv7 and Faster R-CNN(Faster Region Proposal Convolutional Neural Network) models, the improved YOLOv7 model has higher detection accuracy, with an average detection accuracy of 96.4% and a detection speed of 29.6 frame∙s^-1, which can provide reference for the detection of multiple types of defect targets of transmission lines.

Key words: YOLOv7, deep learning, transmission line defect detection, small-target detection, multi-scale fusion, Swin Transformer, β-dropout, adaptively spatial feature fusion, loss function