Defect Insulator Detection Method Based on Deep Learning

In the past, the maintenance of transmission and distribution lines was completed manually, which has low efficiency and poor safety. Taking the typical insulator string falling or burst defect as an example, based on the aerial image of transmission line collected by UAV, this paper proposes an accurate and fast solution of detecting transmission line faults combined with image segmentation and object detection. The method combines semantic segmentation network U-Segnet with the object detection network Yolox based on deep learning. Considering the image size collected by UAV is generally too large and the background is complex, we improve the structure of U-Segnet network and increase its depth, so that we can extract deep level feature information and segment insulators more accurately. At the same time, we add the residual network structure in semantic segmentation network to solve the problem that the network cannot converge due to gradient dispersion. Then the insulators are segmented from the complex background and sent to the object detection network for training. Through experiments, we find that this method can effectively identify normal insulators and defect insulators, and the accuracy can be improved to more than 90%.