GLSANet: Global-Local Self-Attention Network for Remote Sensing Image Semantic Segmentation

Learning long-range contextual dependence is important for remote sensing (RS) image segmentation in complex patterns. Meanwhile, exploring local context information is conducive to the discrimination of fine details. Only underlining either global semantic correlations or local context details is insufficient to achieve accurate segmentation. In this letter, we propose an architecture with the global-local self-attention (GLSA) mechanism, called GLSANet, which can simultaneously consider both global and local contexts for segmentations. Particularly, the GLSA mechanism consists of the global atrous self-attention (GASA) and local window self-attention (LWSA) mechanisms. GASA can learn long-range semantic relations in a gapped manner, while LWSA can locally capture contextual details. As a bridge between the two self-attention (SA) branches, a context fusion module (CFM) is further designed to adaptively integrate global and local contexts. The experiments with public datasets show that the proposed GLSANet significantly refines semantic segmentation and outperforms other competing methods.