A Box-Supervised Instance Segmentation Method for Insulator Infrared Images Based on Shuffle Polarized Self-Attention

Abstract

Infrared imaging technology, as an effective on-site condition monitoring means of equipment, is widely used in the maintenance task of insulators. Accurate segmentation of insulators in infrared images is of great significance for temperature data acquisition and fault diagnosis. The presence of multiple electrical devices in the infrared image substantially increases the difficulty of insulator instance segmentation. Moreover, segmentation methods based on pixel-level sample annotation consume a lot of time and resources. Therefore, a box-supervised instance segmentation method for insulator infrared images based on shuffle polarized self-attention is proposed. Firstly, an adaptive positive and negative sample matching mechanism is designed in an anchor-free detector, which adaptively adjusts the constraint range of sample matching according to the inherent characteristics of the insulator. Secondly, a shuffle polarized self-attention module is embedded in the mask branch to improve the segmentation accuracy of insulator infrared images by feature shuffling and polarized filtering. Thirdly, the insulator weakly supervised segmentation is realized with box annotations by introducing a pairwise loss term and a projection loss term in the training strategy. The experiments show that the proposed method displays superior segmentation performance compared with other advanced method algorithms, which can effectively improve the intelligence of insulator fault diagnosis in transmission lines.