Fast power line detection based on semantic flow

Abstract

In this paper, a Semantic-flow-based fully convolutional network model (SFCN) is proposed to solve the problems of low recall rate in the extraction of thin and long transmission lines in UAV images and are easily affected by complex background and illumination. The backbone of the model network adopts a smaller number of channels to reduce the number of parameters and speed up the learning speed. The inverted residual module is used to enhance the feature learning ability of the network under low number of channels and to prevent model degradation. The semantic flow module replaces the skip connection to complete the accurate fusion of high-dimensional features and low-dimensional features, and finally outputs the pixel-by-pixel recognition results. The method in this paper can realize quickly power line detection. Compared with the regular semantic segmentation models ENet, UNet, NestedUnet, DeepLabv3_plus, GCN, SegFormer, FCHarDNet, BiSeNetv2, and DDRNet, the method in this paper performs the best, with an F1 value of 83.693% and a recall rate of 80.64%.