BIR-Net: A Lightweight and Efficient Bilateral Interaction Road Extraction Network

Abstract

Road segmentation is a crucial aspect in various fields, including intelligent transport systems and urban planning. This article proposes a solution to the problem of inaccurate road region extraction in small devices with limited resources. The proposed solution is a lightweight and efficient bilateral interaction road extraction network, called BIR-Net. First, the detail branch and semantic branch are constructed to form a bilateral feature extraction network for capturing road detail information and semantic information. Then, the shallow interaction module is designed to address the problem of high intraclass variability and interclass similarity in remote sensing images. By exchanging the information of two branches in real time, the edge features of the road are highlighted. The deep interaction fusion module is proposed to fuse information from the two branches using bilateral guided aggregation. Furthermore, to address the characteristics of slender and curved roads in remote sensing images, we have developed the road perception attention module. This module updates the direction weights in real time to track road information, thereby enhancing the network's ability to perceive all road information. The experimental results indicate that BIR-Net has only 3.66M parameters and 6.49G floating point operations. Moreover, the road segmentation accuracies in CHN6-CUG and DeepGlobe datasets are 59.27% and 58.36%, respectively. The proposed method in this article improves road extraction accuracy while maintaining a lightweight structure.

Keywords

• Bilateral interaction network
• lightweight and efficient
• remote sensing images
• road-aware attention module
• road extraction