Multilateral Semantic With Dual Relation Network for Remote Sensing Images Segmentation

Abstract

Semantic segmentation of remote sensing images is an extensively employed and demanding task. Although deep convolutional neural networks have significantly increased the accuracy of semantic segmentation, the problems of losing detailed features in segmentation and ignoring rich contextual information of images still exist. To solve these challenges, we propose a multilateral semantic with dual relation network (MSDRNet) for remote sensing images segmentation. The proposed MSDRNet consists of two parallel modules, the detail semantic module (DSM) and the global semantic module (GSM), for extracting image detail and global features, respectively. Subsequently, improved spatial relation block (SRB) and channel relation block (CRB) are introduced in two separate parallel modules to further enhance the contextual connection of the images. Finally, a feature refinement module (FRM) is added to balance the multilateral features between the features extracted from the two branches. We display the robustness and effectiveness of the proposed MSDRNet on the publicly available ISPRS Potsdam and Vaihingen datasets. We further experimented with the Gaofen Image Dataset (GID), which contains information on larger scale features, to demonstrate the validity of our model. The results of extensive experiments conducted on the above three datasets show that the proposed approach outperforms several state-of-the-art semantic segmentation methods.