Insulator Defect Detection of Lightweight Rotating YOLOv5 Based on Adaptive Feature Fusion

With the construction of smart grid, aerial insulator defect detection based on computer vision has become an important task to ensure grid safety. When the target detection model is too large, it is not conducive to the edge deployment of aerial inspection UAV; Moreover, different aerial photography angles and distances will cause the insulator string in the image to have any direction and less defect information. In order to solve these problems, this paper proposes a rotating GBS-AFP-YOLOv5 model with the combination of lightweight and adaptive features. Firstly, an improved YOLOv5 based on lightweight GBS is proposed by Ghost convolution, which can effectively extract features while reducing the complexity of the model. Then, an adaptive information interaction feature pyramid (AFP) is proposed by combining CARAFE upsampling operator and ECA attention, which effectively fuses the feature information of shallow and deep defects and improves the model performance. Then, a more accurate insulator string detection method is realized by using rotating frame combined with ring label smoothing technology. Finally, the normalized wasserstein distance (NWD) is introduced to improve the loss function, which further enhances the detection ability of the model for small targets with defects. Based on the insulator data set, the test results show that the model has a good defect detection performance, which is improved from mAP0.5 to 0.923 on the basis of only 4.32M parameters.