Study on insulator defect detection based on improved YOLOv8

Abstract Given the problems of poor detection effect in aerial insulator images resulting in poor detection effect, a multi-defect detection algorithm for aerial insulators based on improved YOLOv8 is proposed. The algorithm first changed the FPN structure to BiFPN by introducing the upsampling and downsampling connections in the upstream and downstream of the feature pyramid, to enhance the interaction and information flow between characteristics at different scales. Secondly, the SPPF is improved by using the LSKA attention mechanism to enhance the ability to extract multi-scale features. Finally, for the detection of tiny fine particles, this paper reheavys the decoupling head and the lightweight decoupling head, extracts the key information above the space and channels, and implements a lower inference latency while maintaining accuracy. The algorithm detects the normal glass insulator, glass sheet dirt, glass sheet defect, polymer sheet dirt, polymer sheet defect, and polymer insulator in the aerial image, and simultaneously detects the flashover of the insulator. The experimental results show that the network used in the insulator defect data set detection accuracy (mAP) reached 87.8%, which increased by 2.6% compared to the benchmark model, and the calculation amount (GFLOPs) decreased by 24%. At the same time, after lightweight model parameters are reduced by 32%, based on the accuracy of the subsequent deployment of lightweight requirements, we can better detect insulators of different scales and different conditions.