Improved YOLOv7 Model for Insulator Surface Defect Detection

Abstract

Aiming at the problems of insufficient expression ability and imperfect positioning loss function of YOLOv7 model in the single-stage detection network, an improved YOLOv7 model for insulator surface defect detection is proposed. The new method uses the Mish activation function to replace the original activation function, which further improves the expression ability of the network. The Omni-Dimensional dynamic convolution (ODConv) is used to replace some conventional convolutions in the original model, and multiple convolution cores are fused according to the input, considering the attentions of multiple dimensions, and the network structure is optimized. Using SIoU to measure the loss, the angle term is added to the previous loss function to speed up the convergence of the model. In addition, we use a Copy-Paste data augmentation approach for the problem that there are not enough negative samples of insulator defects. Finally, the proposed method improves the detection accuracy of the network for insulator surface defects. Experimental results show that the improved YOLOv7 model has good detection performance, and the mAP on the insulator dataset reaches 88.7%, which achieves the highest score compared with the original YOLOv7 model and a series of classical mainstream algorithms.