Insulators Recognition Based on Enhancement of High-Level Feature with Improved Attention Method

The integrity of the insulator has a great impact on the reliability and safety of the overhead transmission line. In order to inspect the defects, image segmentation is an important step. The performance of segmentation network will have large influence on the insulator defect inspection strategies. In this case, inspired by the classical Fully Convolutional Networks (FCNs) and Attention mechanism, this paper explores the direction of improving network’s performance by enhancing the encoding quality of high-level features to contrive an optimal way of insulator image segmentation. We introduce a cascaded high-level feature enhancement network, which uses cascaded contracting path of U-Net as encoder and cascaded High-level Feature Enhancement Module (HFE) as decoder. The HFE is committed to selectively aggregating the optimal feature information of high- and low-level features, so as to enhance the ability of decoder to reconstruct high-resolution details. In addition, the HFE can provide independent attention to each value of input features, so as to comprehensively optimize the target related feature. We validated the effectiveness of our method in the insulator segmentation task. The experimental results show that our method achieved considerable improvement compared with the state-of-the-art FCNs including U-Net and attention-based U-Net.