Semantic segmentation of remote sensing image based on Contextual U-Net

Abstract

Semantic segmentation of remote sensing images is a challenging and critical task. The complexity of the remote sensing environment often poses difficulties in accurately capturing object boundaries. To address this challenge, we propose a Contextual U-Net (CU-Net) architecture for semantic segmentation of remote sensing images, which incorporates three collaborative improvements. Firstly, a Boundary Feature Extraction (BFE) module is introduced to fuse semantic feature information from the backbone network with boundary feature information, thereby enhancing the accuracy of edge segmentation in remote sensing images. Secondly, we propose an Adaptive Feature Selection (AFS) module that highlights representative semantic channels for irregular objects, enabling long-distance dependence capture between pixels in the irregular region of the boundary and pixels inside the objects. Thirdly, a Recursive Feature Fusion (RFF) module is introduced to effectively aggregate hierarchical features through adaptive inter-layer feature guidance, facilitating accurate capture of image edges and textures. We collected high-quality remote sensing data through UAVs, comprising 4509 images across 6 different categories. Extensive experiments demonstrate that the proposed CU-Net architecture outperforms state-of-the-art methods.