Multi-scale detail enhancement network for remote sensing road extraction

Extracting high-precision road networks from remote sensing images is of great significance for optimizing urban spatial layout, improving urban traffic efficiency, and improving residents' travel conditions. However, factors such as occlusions, shadows, and complex backgrounds bring some challenges to road extraction, and existing methods often produce incomplete and incoherent road extraction results, with accuracy and generalization ability to be improved. Based on the in-depth understanding of road shape characteristics, a Multi-scale Detail Enhancement Network (MSDE-Net) is proposed to improve the continuity and accuracy of remote sensing road extraction. MSDE-Net uses the Multi-Scale Convolution (MSConv) module to construct the encoder decoder structure, which can simultaneously focus on the local and global multi-scale information of the image. Secondly, to improve the accuracy of segmentation results, Cross Block is introduced to integrate information across spatial scales and to enhance the network's ability to understand the context; Finally, to improve the continuity of the road segmentation, an edge enhancement (DE block) module is embedded to strengthen the network's focus on thin and weak feature structures. The experimental results show that the mIoU of MSDE-Net on CHN6-CUG, DeepGlobe, and Massachusetts datasets reaches 80.08%, 81.26%, and 78.91%, respectively, which provides better performance compared with other mainstream image segmentation networks and recently proposed road extraction networks.