Merged Intersection Over Union: A Metric and a Loss for Aerial Images in Transmission Lines
Hui Zhang, Jianming Du, Chenjun Xie, Feng Zhao · 2023
With the development of information technology and new energy, efficient power inspection is crucial for the safe and stable operation of the power grid. In recent years, State Grid of China has vigorously promoted intelligent transmission line inspection, gradually replacing the initial manual inspection of transmission lines with Unmanned Aerial Vehicle (UAV) inspection. The UAV inspection mainly relies on UAV high-altitude collection of transmission line images, and uses object detection (OD) and recognition algorithms to automatically determine various device defects in the transmission line. Due to the current mainstream OD methods using IoU to measure the degree of overlap between the predicted box and the real box, IoU is not only unable to measure scenes where the object boxes do not coincide, but also extremely sensitive to small-scale object position changes. If we continue to use IoU to measure high-altitude transmission line defect objects with large scale span, the OD algorithm model cannot achieve better performance. In order to alleviate this problem, we propose a new evaluation metric method, which is the Merged Intersection Over Union (MIoU). MIoU is a plug and play component that can be embedded into label assignment and loss function of any anchor-based detector to replace commonly used IoU metric. We conducted experiments on the MIoU performance in the transmission line component fault dataset, and a large number of experiments showed that using our proposed metric model improved 6.8% AP compared to the benchmark model, and significantly improved the recall rate of the model on certain key defect classes.