Corner Guided Instance Segmentation Network for Power Lines and Transmission Towers Detection.

Gaoyi Zhu,Wenxuan Zhang,Mei Wang,Jie Wang,Xia Fang

DOI: https://doi.org/10.1016/j.eswa.2023.121087

IF: 8.5

2023-01-01

Expert Systems with Applications

Abstract:Power lines (PLs) and transmission towers (TTs) detection from aerial images is the prerequisite step for unmanned aerial vehicles (UAVs) navigation, path planning, and obstacle avoidance. Unlike the previous works which only focus on PLs detection or TTs detection, this work aims to detect PLs and TTs together based on the instance segmentation algorithm. However, PLs and TTs in aerial images present numerous challenges to object detection and segmentation researchers—the slender characteristic of PLs, the foreground–background class imbalance, and the foreground–foreground class imbalance. To cope with the above challenges, this work proposes a corner-guided instance segmentation network (CGISnet) for detecting PLs and TTs together from aerial images. Specifically, an attention-based multiscale feature extraction module facilitates the extraction of the basic spatial feature information in the case of serious background noise interference. To accumulate more feature information about the objects, especially the slender objects, a corner-guided region of interest (ROI) align module detects the object as a pair of corners to obtain the corresponding ROI information for further segmentation. In addition, a weighted cross entropy loss function is utilized to resolve the foreground–foreground class imbalance issue. Extensive experiments were conducted on self-built and two public datasets to demonstrate the effectiveness and robustness of the proposed CGISnet. The results demonstrate that our approach not only detects PLs and TTs with high accuracy, but also achieves competitive performance in detecting objects with different shapes and sizes.