TAR-Net: A Triple Attention Residual Network for Power Line Extraction from Infrared Aerial Images

Abstract

AbstractTransmission line segmentation is of great significance for the intelligent power inspection, which could well serve the path planning and navigation of different inspection platforms. Combined with the strong context extraction ability, deep learning has provided an effective means for pixel-level image segmentation. However, the power lines are always against with complex natural environment, such as different lighting conditions, different visibility, different natural environment, etc, which will bring a great effect for accurate power line extraction from aerial images. Meanwhile, compared with the background environment, the image pixels of the power line account for a small proportion, which will lead to the problem of unbalanced pixel proportion, and it also will affect the whole segmentation performance. Faced with the above issues, with the encoder-decoder framework, a triple attention residual network, namely TAR-Net, is proposed in this paper for accurate power line extraction from aerial images. To realize effective feature extraction, a residual U-Net network is built to acquire strong contexts. Faced with the class imbalance issue, a triple attention block is proposed to make the segmentation network better focus on the power lines. Further, a dense convolution block is proposed for feature enhancement of local feature maps. Combined with public data sets on infrared aerial images, experiment results show that the proposed segmentation network could acquire a better segmentation performance on power lines against complex environments compared with other advanced segmentation models.KeywordsPower linesDeep learningResidual networkTriple attention