RE-Net: Road Extraction from Remote Sensing Images with Deep Learning and Geometric Priors

Road extraction is one of significant research hot-spots in the area of remote sensing images processing and applications. Recently, deep learning based methods have become the mainstream research method due to their high degree of automation and superior performance. However, the unique characteristics of roads in airborne or satellite remote sensing images still limit the accuracy and connectivity of road extraction, such as complex and diverse backgrounds, small proportions, large spans, multiple confusing categories, occlusion, and limit resolution. To address some of these challenges, we propose a new Road Exaction Network (RE-Net) with deep learning and geometric priors for road extraction from remote sensing images for better accuracy and connectivity of road extraction. The proposed RE-Net is composed by Detail branch module, Fusion module and MHNet which is composed by MSCAN backbone and Ham module. The main contributions of this paper are as follows. A new Road Extraction network with Detail supervision, Fusion module and MSCANN as backbone net, using dual branch semantic segmentation model based on the geometric prior knowledge of roads. The segmentation branch uses MSCAN as the backbone network, which can better adapt to the geometric shape of the road and extract multi-scale strip-like features. In addition, a lightweight Ham module to enhance the modeling ability of the model for long-term distance dependencies, and uses a refined feature fusion module to fully fuse the shallow spatial features and the semantic features. The detail branch combines road edge priors, which can supervise the model to pay more attention to road edge information. The experimental results on the DeepGlobe dataset and CHN6-CUG dataset demonstrate that the proposed RE-Net can effectively improve the accuracy and connectivity of road extraction from remote sensing images.