SFR-Net:A Spatial Feature Enhance Method for Road Extraction

In recent years, road remote sensing segmentation has become a necessary prerequisite for map modeling and urban planning. Convolutional neural networks have been widely used in the field of road remote sensing segmentation, and have achieved good performance. However, satellite road photography has the features of distinct objects and huge disparities, making it difficult to extract picture contextual information. We present a SFR-Net based on the mainstream encoder-decoder architecture, as well as an unique multi-scale information fusion enhancement module. The module is inserted between distinct scale features, modeling the interaction of features at different sizes in order to obtain remote sensing picture context information and effectively increase model segmentation precision. The model's performance is validated on CHN6-CUG in this paper, and the precision is enhanced by 1.12% compared to the current mainstream FCN-UNet. As a consequence, we are certain that SFR-Net will become a trustworthy baseline.