End-to-End Insulator String Defect Detection in a Complex Background Based on a Deep Learning Model

Abstract

Normal power line insulators ensure the safe transmission of electricity. The defects of the insulator reduce the insulation, which may lead to the failure of power transmission systems. As unmanned aerial vehicles (UAVs) have developed rapidly, it is possible for workers to take and upload aerial images of insulators. Proposing a technology to detect insulator defects with high accuracy in a short time can be of great value. The existing methods suffer from complex backgrounds so that they have to locate and extract the insulators at first. Some of them make detection relative to some specific conditions such as angle, brightness, and object scale. This study aims to make end-to-end detections using aerial images of insulators, giving the locations of insulators and defects at the same time while overcoming the disadvantages mentioned above. A DEtection TRansformer (DETR) having an encoder–decoder architecture adopts convolutional neural network (CNN) as the backbone network, applies a self-attention mechanism for computing, and utilizes object queries instead of a hand-crafted process to give the direct predictions. We modified this for insulator detection in complex aerial images. Based on the dataset we constructed, our model can get 97.97 in mean average precision when setting the threshold of intersection over union at 0.5, which is better than Cascade R-CNN and YOLOv5. The inference speed of our model can reach 25 frames per second, which is qualified for actual use. Experimental results demonstrate that our model meets the robustness and accuracy requirements for insulator defect detection.