Semantic segmentation networks for powerlines incorporating self-supervision and self-attention

The obstacle avoidance problem is critical when Unmanned Aerial Vehicles(UAVs) encounter transmission lines during aerial flight. One key in semantic segmentation of transmission lines is to solve the problem of long-distance image segmentation discontinuity problem. To this end, we propose a semantic segmentation algorithm with the addition of a self-attention module to improve the U-Net for the semantic segmentation of transmission lines. The self-attention module extracts the global features on different scales of U-Net, which enhances the ability to capture the global features of transmission lines. To further optimize the training process, maximum pooled label downsampling is proposed to enhance the learning ability of unbalanced classes of transmission lines. Moreover, a self-supervised pre-training of convolutional neural network image mask modeling is proposed to improve the quality of pre-training weights. In addition, for validation on a large-scale transmission line dataset, the instance segmentation labels of the TTPLA transmission line and tower dataset are processed to produce a novel semantic segmentation dataset. Experiments show that the improved network has higher segmentation accuracy than the original U-Net through the self-attention mechanism for capturing global features, the optimized depth-supervised process, and the self-supervised pre-training. On the TTPLA transmission line semantic segmentation dataset, the metric IoU increases by 2.32% compared to the original U-Net, reaching 71.45%. Our results prove that the proposed algorithm enhances the association among the semantic features of long-distance transmission lines in images and improves the integrity of transmission line semantic segmentation. The approach can enhance the obstacle avoidance capability of UAVs.