An Improved Insulator and Defect Detection Method Based on ARA-YOLOv5s

The accurate identification of insulators and defects is very important for the safe and stable operation of power systems. The detection accuracy is poor due to diverse scales of insulators and complex backgrounds. Particularly, the relatively small defects will increase the difficulty. To solve these problems, this paper proposes the ARA-YOLOv5s (the YOLOv5s based on asymmetric-dilated-multi-scale-convolution, receptive-field-coordinate-attention-convolution and adaptive-spatial-feature-fusion) model for detecting insulators and defects. Firstly, the ADMC module is proposed to improve the feature extraction by using asymmetric atrous convolution with different dilates and an adaptive weights fusion mechanism. Secondly, aiming at the complex background interference, the RFCAConv (receptive field coordinate attention convolution) module is introduced to emphasize the important features by the fusion of receptive field attention convolution and the CA (coordinate attention) attention mechanism. Thirdly, the head makes use of the ASFF (adaptive spatial feature fusion), which can adaptively select and fuse feature information from different scales as required, thus improving the detection performance. the $\mathrm{mAP} \text{@} 0.5$, [email protected]:0.95 increased by 4.7%, 2.5%, respectively. More importantly, the improvement in detecting defects is more significant, which indicates the superiority of recognizing smallscale targets.