全向矩形校准的高分辨遥感影像细节分割

打开文本图片集

中图分类号：TP79 文献标识码：A

Abstract： To address incomplete feature extraction，blurred boundaries，and omission of smalltargets in high-resolution remote sensing image segmentation，a detail-preserving segmentation method based on omnidirectional rectangular calibration，termed the Omnidirectional Rectangular Calibration Network （ORCNet），is proposed.First，an Omnidirectional Rectangular Calibration State Space Module （ORSM） is introduced to enhance geometric adaptability and target retention via octagonal scanning and geometry-sensitive weight calibration. Next，a Complementary Filtering Hybrid Attention Fusion Module （CFHAF） is developed，integrating channel-，spatial-，and pixel-level atention mechanisms to enableadaptive multiscale feature fusion and improved semantic discrimination. Finally，Dynamic Point Upsampling （DySample）is incorporated to refine boundary detail recovery.The model is trained with a Focal-Dice hybrid loss. On the Massachusetts Buildings dataset， the method achieves an F1 score of 84.64% and an mIoU of 77.07% . On the DeepGlobe Road dataset，an F1 score of 85.32% is obtained，representing a 3.51% improvement over RSMamba. Experimental results indicate that the proposed approach efectively addresses the three primary challenges in remote sensing segmentation，delivering high precision，robust per formance，and strong potential for practical application.

Key words： remote sensing image segmentation；rectangular self-calibration； state space model； multiscale feature fusion; edge detail enhancement； dynamic upsampling

1引言

高分辨率光学遥感卫星作为现代对地观测体系的核心载荷，其亚米级空间分辨率为实现地表精细识别提供了数据基础。（剩余20229字）