A deeply supervised vertex network for road network graph extraction in high-resolution images

Extracting road network maps for high-resolution remote sensing images is a critical and challenging remote sensing topic, with significant importance for traffic navigation, disaster management, autonomous driving, and urban planning. Although deep learning has demonstrated its powerful feature extraction capabilities in image processing tasks, including road extraction, existing road extraction algorithms still have some limitations. In particular, semantic segmentation-based methods often lack supervision of the connection relationships between roads and topological correctness constraints, resulting in fragmented and poorly connected road network graphs. Moreover, graph-based methods lack effective supervision strategies for vertex misdetection issues. To deal with these issues, we propose a deeply supervised vertex network (DSVNet) for road network graph extraction. First, to effectively supervise road vertices, we design a road vertex supervision module that yields improved vertex prediction accuracy. Second, to merge the benefits of segmentation-based methods and graph-based methods, we establish a parallel semantic segmentation branch based on the vertex querying task, thereby achieving enhanced road extraction accuracy. Furthermore, we introduce deformable attention into the model to boost its performance and computational efficiency. We validate the effectiveness of DSVNet on two large-scale public datasets. A large number of experimental results show that our research approach achieves the best road detection performance. • A new graph-based end-to-end road extraction algorithm is proposed. • Introducing deformable attention into road extraction algorithms. • Designed a structure for collaboration between segmentation and vertex query tasks. • A road vertex supervision module is designed for supervised vertex prediction. • The algorithm is validated on two large publicly available remote sensing datasets.