Multi-semantic contrast enhancement for robust insulator defect detection

Abstract

The effectiveness of deep learning-based methods for insulator defect detection has been proven. However, in practical applications of power transmission lines, the complex and variable backgrounds in insulator images, coupled with the difficulty in labeling insulator defects, pose challenges to improving the robustness of such methods. Existing studies often utilize generative adversarial networks or forcefully combine foreground and background to augment training samples, but they overlook the rich semantic information in complex scenes, leading to distorted generated adversarial samples. To address this challenge, an innovative multi-semantic contrast enhancement method that significantly enhances the robustness of defect detection by deeply integrating high-level semantic knowledge and low-level signal priors is proposed. Moreover, through adversarial training using generated samples with diverse semantics and real samples, the robustness of the method is further improved. Experimental results demonstrate that this method surpasses state-of-the-art models, achieving significant performance on three independent cross-scene datasets.