Foreign Object Detection on Insulators Based on Improved YOLO v3

As an important component of transmission lines, insulator plays an essential role in the stable operation of the power grid. However, the outdoor environment in which the insulators are located can easily lead to the hanging of foreign objects. This paper proposes a novel method for foreign object detection on insulators based on the improved YOLO v3: Dense-YOLO v3. A dense network is designed to replace one of the convolutional layers of the original network in order to realize the multi-layer feature reuse and fusion of the insulator, which improves the detection accuracy. In addition, we amplify the training set to improve the training effect of the network and propose a wrong detection cost function to measure the risk of false detection. The experiment shows that the proposed algorithm has a detection precision rate reaching up to 94.54%. Meanwhile, the Dense-YOLO v3 outperforms YOLO v3 and Faster R-CNN, both in terms of detection accuracy and wrong detection cost. The result shows that the presented approach can be applied to the UAV inspection of transmission lines.