Msmdff-net: multi-scale fusion coder and multi-direction combined decoder network for road extraction from satellite imagery

Using deep learning to extract roads from satellite images is one of the most popular methods. However, the existing encoder-decoder-based deep networks usually produce fragmented roads, due to the complex spatial and color characteristics of the road. In this paper, motivated by the road multi-scale information, we proposed a multi-scale and multi-direction feature fusion network (MSMDFF-Net) to reduce the fragmentation of road extraction results. The proposed method mainly consists of three processes: 1) In the initial stage, the image details from different directions were transmitted; 2) At different encoding stages, the multi-scale information of the image was fused; 3) In the decoding process, the matching modules of road characteristics were used to up-sample the feature map. Extensive experiments on the popular datasets (LSVD and Deep-Globe datasets) demonstrate that the MSMDFF-Net has higher accuracy and generalization performance with less fragmentary road results.