DTUR-Net: a dual-task UR neural network for road segmentation and centerline extraction from high-resolution remote sensing images

Simultaneous road segmentation and centerline extraction from high-resolution remote sensing images is a challenging task in remote sensing image analysis. There is a lack of information exchange between the two major tasks of road segmentation and centerline extraction, resulting in the extracted road surface and centerline exhibiting issues of fragmentation and incompleteness to a certain extent. To address this, we propose a network architecture that combines road segmentation and centerline extraction, called the dual-task UR neural network. It employs feature reuse operations to adopt road segmentation results and remote sensing images as prior knowledge for centerline extraction, facilitating information transfer and parameter sharing between the two tasks. The road segmentation results constrain the generation of centerlines, effectively alleviating potential issues in the centerline. At the same time, a multiscale stripe convolution module is utilized to integrate road semantic information at different scales, capturing context features that conform to road structures in the proposed network. In addition, a feature enhancement extraction module is introduced to strengthen the degree of feature extraction and extract deeper level feature information. Furthermore, binary cross-entropy and focal loss composite functions are employed to alleviate problems arising from sample imbalance. Finally, experiments are conducted on the Massachusetts dataset and CHENG_ROAD dataset, comparing the proposed method with other mainstream road extraction models. Experimental results demonstrate that the proposed method outperforms the compared methods in terms of extraction accuracy and visual performance.