DRA-Net: A Dual-Branch Residual Attention Network for Pixelwise Power Line Detection

Abstract

The accurate extraction of transmission lines from UAV aerial photography is a necessary but challenging task for intelligent power inspection due to some factors, such as complex backgrounds, varying illumination, and class imbalance issue. Although there have been some research achievements in transmission line extraction, especially U-Net and its variant networks, there are still some shortcomings, such as insufficiency processing of local features, limited noise denoising ability, poor edge feature representation ability, etc. In this paper, we propose a dual-branch residual attention network, called DRA-Net, which introduces a dual-branch encoder composed of residual convolutional neural network (RCNN) branch and recurrent residual convolutional neural network (RRCNN) branch to extract richer semantic information and a context fusion block (CFB) to fuse temporal and spatial information effectively. In addition, an U-shape noise de-noising block (UND) is proposed to reduce interference from complex backgrounds and an edge enhancement block (EEB) is also proposed to strength the capacity of segmentation network to extract useful edge feature information from noise. Experiments demonstrate that proposed DRA-Net has obtained an excellent segmentation manifestation on remote sensing images, with a 93.26% Dice coefficient and 93.19% mIOU value on public power line dataset (PLD), and a 96.40% Dice coefficient and 96.04% mIOU value on self-built overhead power line (OPL) dataset.