BES-Net: Boundary Enhancing Semantic Context Network for High-Resolution Image Semantic Segmentation

This paper focuses on the high-resolution (HR) remote sensing images semantic segmentation task, whose goal is to predict labels in a pixel-wise manner. Due rich complexity and heterogeneity of information HR images, ability extract spatial details (boundary information) context dominates performance segmentation. In this paper, based frequently used fully convolutional network framework, we propose boundary enhancing (BES-Net) explicitly use enhance extraction. BES-Net mainly consists three modules: (1) extraction module for extracting information, (2) multi-scale fusion fusing features containing objects with multiple scales, (3) fused extracted improve intra-class consistency, especially those pixels boundaries. Extensive experimental evaluations comprehensive ablation studies ISPRS Vaihingen Potsdam datasets demonstrate effectiveness BES-Net, yielding an overall improvement 1.28/2.36/0.72 percent mF1/mIoU/OA over FCN_8s when BE MSF modules are combined by BES module. particular, our achieves state-of-the-art 91.4% OA dataset 92.9%/91.5% mF1/OA dataset.