Anomaly detection of power line insulator from aerial imagery with attribute self-supervised learning

Unmanned aerial vehicles (UAVs) can conveniently capture the insulator images in aerial scenes. With the aerial insulator images, we can effectively conduct insulator status inspections. However, the insulator anomaly detection is challenging with low accuracy due to the complex background in insulator images as well as the insulator variety. To cope with this problem, we propose a novel insulator anomaly detection method. Specifically, we first use the Generative Adversarial Network (GAN) to coarsely detect insulator defects based on the reconstruction error of insulator area. Then, we incorporate the foreground attribute learning and structure attribute learning to dynamically improve our model's sensitivity for detecting the insulator defects. The foreground attribute learning aims to highlight the foreground regions of aerial insulator images, which makes the image features more robust to the background interference. Also, by using the structure attribute learning method, our model can learn normal structure pattern of insulators more effectively, increasing the ability to distinguish the abnormal sample. With these strategies, the proposed model reduces the influence of background interference and becomes more discriminative to the insulator defects. Extensive experiments on one real-world UAV images dataset have demonstrated the effectiveness of the proposed method for insulator anomaly detection.