Research on transmission line ice-cover segmentation based on improved U-Net and GAN

Severe ice cover can cause line dancing, insulator flashing, tower tilting, and even collapse of the tower. which is threatening the safety of transmission line system and has great impact on the industry production and people life. In this paper, we propose a deep learning-based ice-cover segmentation method (S_UNet), which is promising for the intelligent operation of the power grid. It uses the encoder-decoder structure from U-Net to improve the backbone network performance, at the same time it is a lightweight network structure with attention mechanism, migration learning, and auxiliary loss, etc. Through experimental analysis, S_UNet's false and partial detection phenomenon still need to be improved. So an optimized network SGAN_UNet composed by Generative Adversarial Networks (GAN) and S_Unet is proposed. It uses the ground truth labeled images to compare with predicted images to improve segmentation performance. The experimental results showed that the SGAN_UNet network reduces false and partial detection problems in the ice-cover segmentation, which has 1.10% improvement in mIoU comparing with S_UNet.