MFE-Net: Multi-Scale Attention Feature Enhancement Fusion Network for Building Extraction From Remote Sensing Images

In the field of building extraction from remote sensing images, existing techniques often cause incomplete segmentation due to the diverse and complex shapes of remote sensing buildings, and present the problem of sticking when segmenting neighboring targets. To address the above problems, we propose a Multi-Scale Attention Feature Enhancement Fusion Network (MFE-Net), which is composed of a ConvNeXt backbone network and three specially designed modules; among them, the Multi-scale Global Context Feature Fusion (MGCF) module is used to supplement the global contextual information and enhance the understanding of global features. The Channel Global Feature Enhancement (CGFE) module is used to reduce potential channel bias during feature fusion. The Global Grouped Coordinate Attention (GCA) module enhances the localization and emphasis of small target features. Experiments conducted on the publicly available Massachusetts buildings dataset and the Inria aerial imagery dataset show that the building IoU achieved by MFE-Net are 75.22 % and 82.57 %, respectively. The experimental results show that MFE-Net exhibits superior segmentation performance compared to the current state-of-theart segmentation networks.