Insulator defect detection based on improved YOLOv5 algorithm

Due to the issues with a high number of parameters and the sluggish detection speed of the present insulator defect identification algorithm, an enhanced YOLOv5 target detection approach is proposed. To begin, replacing the YOLOv5 backbone network with the lightweight network MobileNetV3 minimizes the amount of processing and the number of parameters used during feature extraction, resulting in faster detection. Afterward, clustering with K-means++ is used to produce anchor frames that are better suited for insulators on transmission lines with thin structures. The network model can then be used to discover small targets for insulator flaws by adding a coordinate attention mechanism at the feature fusion stage. Last but not least, the edge regression is optimized for accurate target localization, which raises the recognition rate and improves the loss function. The revised YOLOv5 network model has a file size of just 7.2 MB, which is only 51% smaller than that of the original YOLOv5 network.This results in a reduction in the number of parameters, an increase in detection speed and accuracy, and a 98% average detection accuracy for mAP. The inference time is also reduced by 21 ms.