Abstract:

To address the problem of multi-scale detection of multi-class defect targets in transmission lines in complex environments, a YOLOv5 transmission line multi-class defect target detection algorithm was proposed based on Transformer and perceptual field modules, abbreviated as TR-YOLOv5. First, a YOLOv5 network was built to address the problem of low saliency of defect targets caused by complex backgrounds, which hindered accurate detection. The Transformer module was introduced in the Backbone part. By utilizing a multi-head attention structure to capture the correlations and global information between the pixels of feature maps, the feature expression capability of the defect targets was enhanced, thereby improving the detection accuracy of the model. Secondly, since the target being detected is impacted by multiple scales, a perceptual field module was introduced in the Neck part to extract features of different scales of the target. Null convolution was also employed to increase the perceptual field, while more detailed features were reserved for the subsequent PANet structure. Furthermore, the Neck feature fusion capability was bolstered to enhance the detection accuracy of the model for multi-scale defective targets. In addition, to enhance the precision of predicted border regression, the CIOU function was introduced to further boost the detection accuracy of the algorithm. Finally, the proposed algorithm was validated using the data of a power inspection department for the past three years. The experimental results demonstrated that the proposed algorithm could surpass seven comparative algorithms in terms of detection accuracy and real-time performance, with an average detection accuracy of 95.6% and the inspection image detection speed for 1280×720 resolution reaching 125 frames/second.

Key words: YOLOv5, transmission line defect detection, dilation convolution, Transformer, receptive field block, loss function