Insulator Defect Detection Method Based on The Improved MG-YOLOv5s Model

As an important component in the power system, whether the insulator can work properly directly affects the stable operation of the power system. Currently, the detection of insulator defects is mostly based on image processing methods, and the commonly used object detection methods have poor recognition effects for small defects. Therefore, this paper proposes an insulator defect detection method based on the improved MobileNetv3 and Ghost Module Integrated You Only Look Once version 5 small (MG-YOLOv5s) model for small defects. Compared with You Only Once version 5 small (YOLOv5s), this model replaces the backbone feature extraction network with MobileNetv3 and uses the Ghost Module to lighten the Neck and Head parts of the model, reducing the model size and increasing the detection speed; it adopts Scaled Intersection over Union (SIoU) as the regression loss and uses Focal Loss to optimize the confidence loss, solving the problems such as the mismatch between the predicted box and the real box direction and the balance of positive and negative samples and difficult-to-classify samples; it replaces the attention Squeeze-and-Excitation Networks (SENet) and the activation function Hard-Swish with Efficient Channel Attention Network (ECANet) and Leaky Rectified Linear Unit (ReLU) respectively, retaining the detailed information of the target, effectively improving the feature expression ability of the network and enhancing the recognition confidence of the model for insulator defects in complex environments. The experimental results show that although the average detection accuracy Multi-class Average Precision (mAP) of the improved model is slightly reduced, the model size is reduced to 26.01% of the original, only about 7.07MB, and the detection speed is increased by about 13.79%, reaching 33f·s<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−1</sup>. The recognition accuracy AP of the small pollution flashover defect is increased by about 10.41%, further verifying the feasibility and effectiveness of this model in insulator defect detection.