Deep nested U-structure network with frequency attention for building semantic segmentation

Abstract The automated segmentation of buildings from remotely sensed imagery has undergone extensive research and application across various industrial domains. Despite this, several challenges persist, including incomplete internal extraction, low accuracy in edge segmentation, and difficulties in predicting irregular targets. We have introduced a novel approach to address these issues: an end-to-end residual U-structure embedded within a U-Net, enhanced by a frequency attention module and a hybrid loss function. The novel residual U-structure is introduced to replace the encode-decode blocks of traditional U-Nets, and the hybrid loss function is utilized to guide segmentation for more complete and accurate segmentation masks. A frequency attention module is also implemented to emphasize essential features and minimize irrelevant ones. A comparison of the proposed framework with other baseline schemes was conducted on four benchmark data sets, and the experimental results demonstrate that our framework performs better segmentation than other baseline state-of-the-art schemes.