HRGA-Net: Hierarchical Rotation Gaussian Attention Network for Accurate Insulator Detection from UAV Images

Insulator detection plays a critical role in the power system operation. However, it is challenging to perform insulator detection because of the following issues: (1) The insulator is always masked by complex environments, e.g., fog and forest; (2) The insulator varies greatly in the background. To solve such problems, a Hierarchical Rotation Gaussian Attention Network (HRGA-Net) is proposed for insulator detection, where the Rotation Gaussian Convolution Attention (RGCA) module is devised to distinguish the insulator from the complicated environments by exciting the effective channel dependencies and learning the multi-view spatial distribution information around the Gaussian neighborhood; then, the Multipolar Rotation Gaussian Attention (MRGA) module is devised to detect the multi-scale insulators from the aerial photographs by capturing different-scale and multi-view details from hierarchical convolutional layers. Experiments conducted on the benchmarks show that the proposed HRGA-Net can achieve excellent performance with the mAP@.5 of 99.33%, 99.28%, 96.53%, and 97.27% on SFID, UPID, CPLID, and IDID datasets, respectively, further assisting in the power system operation.