Power line detection in millimetre‐wave radar images applying convolutional neural networks
Wei Xiong, Jingsheng Luo, Chaopeng Yu
Abstract
Abstract High‐voltage power lines and their supporting towers create hazardous low‐altitude flight conditions for helicopters. Herein, a novel power line detection and classification is proposed for millimetre‐wave (MMW) radars applying Hough transform (HT) and convolutional neural network (CNN). HT is implemented to detect candidate lines in MMW radar images. Unfortunately, in the presence of background clutter and manmade structures, some false power lines are also brought in, thus it is vital to differentiate the actual power lines. Moreover, it is important to classify the power lines with and without Bragg pattern, and offer different threat levels. The authors firstly introduce the incidence angle and azimuth angle vectors of candidate lines as a supplement to power vectors, and propose a new CNN architecture for classification that uses powerful dense blocks and defined index pooling. With the trick of index pooling, the incidence angle and azimuth angle vectors can be merged with the corresponding power vectors together to mine the features, including those of Bragg pattern. Taking advantage of scale transformation, the proposed network adapts to different input scales. Extensive experiments demonstrate that the proposed method achieves state‐of‐the‐art results and is not sensitive to the input scales and noise levels.