窗口注意力融合多分支特征的高分遥感图像语义分割  

Semantic segmentation of high-resolution remote sensing images based on window-attention-fusion multi-branch features

作　　者：徐世博 闫钧华[1,2] 张寅 王达伟[1,2] 张新志 XU Shibo;YAN Junhua;ZHANG Yin;WANG Dawei;ZHANG Xinzhi(Key Laboratory of Space Photoelectric Detection and Perception,Ministry of Industry and Information Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 21l106,China;College of Astronautics,Nanjing University of Aeronautics and Astronautics,Nanjing 21l106,China)

机构地区：[1]南京航空航天大学空间光电探测与感知工业和信息化部重点实验室,南京211106 [2]南京航空航天大学航天学院,南京211106

出　　处：《测绘科学》2024年第12期104-113,共10页Science of Surveying and Mapping

基　　金：中央高校基本科研业务费项目(NJ2024027,NJ2023029)。

摘　　要：针对通用语义分割模型分割高分辨率图像时,难以平衡精度和显存消耗问题,该文提出窗口化融合多分支特征语义分割模型。在特征编码阶段采用三分支网络架构,分别提取高分辨图像不同尺度不同侧重点的特征信息;针对不同分支特征的特点,基于自注意力机制设计窗口相关融合模块来融合局部窗口特征,基于互注意力机制设计分支交叉融合模块来融合不同分支特征,并将融合处理后的特征解码,完成高分辨率图像分割任务。实验结果表明,所提算法在DeepGlobe和Inria Aerial数据集上的mIoU分别达到了73.87和74.93,同时显存消耗较少,优于已有主流算法。Addressing the challenge of balancing accuracy and memory consumption in general semantic segmentation models when segmenting high-resolution images,a window-fusion multi-branch feature semantic segmentation model was proposed in this paper.During the feature encoding phase,a three-branch network architecture was used to extract features from high-resolution images at different scales and with different focuses.A window relation fusion module was designed based on the self-attention to integrate local window features,and a cross branch fusion module was designed based on the cross-attention to merge features from different branches.The integrated features were then decoded to complete the segmentation task of high-resolution images.Experimental results showed that the proposed algorithm achieved mIoU scores of 73.87 and 74.93 on the DeepGlobe and Inria Aerial datasets,respectively,with less memory consumption,outperforming existing mainstream algorithms.

关 键 词：语义分割 高分辨率遥感影像 多分支网络架构 注意力机制 

分 类 号：P237[天文地球—摄影测量与遥感]