Boosting power line inspection in bad weather: Removing weather noise with channel-spatial attention-based UNet

Power line inspection based on UAVs can effectively improve the inspection efficiency. With the development of object detection algorithms, automatic detection and recognition for power line components based on UAVs can further improve inspection speed. However, the performance most object detection methods are easily affected by noise, which is commonly encountered in the natural world such as rain, snow, haze, etc. In this paper, we aim at improving the power line detection performance of UAV inspections in bad weather. Specifically, we first construct a power line components detection (PLCD) dataset, which includes 1943 power line images captured by UAVs and corresponding annotated bounding boxes. Then we generate three sub-datasets named PLCD-R, PLCD-S, PLCD-H to simulate captured images under rain, snow, and haze conditions, respectively. A new image restoration model termed CSUNet is proposed for better remove the noise and restore images. Extensive experimental results demonstrate our CSUNet can well remove the noise captured in bad weather and improve the downstream object detection performance. The PLCD dataset and the codes will be publicly available to facilitate future research.