Abstract: Power line segmentation is of great significance to the automatic obstacle avoidance and the low-altitude flight safety of unmanned aerial vehicles. Whereas the traditional line-based and segment-based algorithms can only be applied in some simple scenes, they are prone to give false positive and false negative results in complex scenes. In recent years, the rapid development of deep learning has greatly promoted the research of power line segmentation. However, there are still three problems we observed in related research: 1) Less consideration of practical application; 2) inadequate use of power line characteristics; 3) ignoring the problem of lacking large-scale power line data. Starting from the application requirements of power line segmentation, this paper improved the traditional F1-Score evaluation matrix, and proposed a more suitable one, termed WG-F1-Score, for power line segmentation. Furthermore, in view of the characteristics of power line, this paper proposed a lightweight real-time semantic segmentation network, dubbed as SaSnet, which included two versions with different depths, lite and general. To solve the problem of lacking large-scale data for power line segmentation, this paper proposed a self-supervised learning algorithm, inpainting based self-supervised learning (IBS). Based on the IBS algorithm, trained with a very small amount of labeled data, SaSnet achieved the state-of-the-art performance in both segmentation performance and inference speed on the public dataset. The experiment results on the Jetson AGX Xavier show that SaSnet has preliminary capabilities for practical applications.