PLE-Net: Automatic power line extraction method using deep learning from aerial images

Abstract

Automatic power line extraction is a crucial task for the safe navigation of inspection robots. Nevertheless, power lines are always against complicated natural backgrounds which bring a certain challenge for accurate power line extraction. Meanwhile, the power lines always occupy a minimal portion image pixels in the aerial images compared with backgrounds which causes serious class imbalance issue. Therefore, the robust and accurate power line segmentation from aerial images is one of the most frequently stated problems faced with these factors. Recently, the deep learning has got wide applications on different segmentation tasks with effective contextual feature generation ability. However, these methods show poor ability on the samples with class imbalance due to insufficient process of local contextual features. To address these issues, combined with the encoder–decoder framework, a novel power line extraction network (PLE-Net) is proposed in this paper to construct an end-to-end attention-based segmentation method for automatic power line extraction from aerial images with a self-attention block and a multi-scale feature enhance block. To capture rich contextual relationships from local feature maps, a feature enhance block is proposed for multi-scale feature expression. And a self-attention block is proposed to embed into the proposed segmentation network to emphasize the regions about power lines. Further, the hybrid loss function with binary cross-entropy (BCE) and Dice is set as the loss function to address the class imbalance issue. Combined with the public datasets of power lines, the proposed segmentation network shows a better segmentation performance on vision images and infrared images through the ablation analysis and comparison experiments.