摘要
为了解决在遥感图像语义分割任务中存在的目标物体之间巨大尺度差异和丢失空间细节信息导致分割精度下降的问题,提出多尺度互注意力与指导上采样网络.利用多尺度互注意力模块获得不同尺度图像之间的像素关系,平衡不同尺度物体的权重,提高小尺度物体的分割性能.编码指导上采样模块利用编码结构中的信息,指导图像上采样的过程,融合空间细节信息,提升目标物体边界像素的分类效果.在Potsdam数据集和Jiage数据集上的m IoU得分分别为85.52%和86.59%,较次优网络分别提升了1.32%和1.46%.
A network with multi-scale mutual attention and guidance upsampling was proposed in order to solve the segmentation accuracy degradation caused by the huge scale difference between target objects and the loss of spatial details in the semantic segmentation of remote sensing images.The multi-scale mutual attention module was used to obtain the pixel relations between different scale images and balance the weights of different target objects in order to improve the segmentation performance of small-scale objects.The image upsampling process was guided by the information in the coding structure,and spatial details were incorporated to enhance the classification of target object boundary pixels in the coding guidance upsampling module.The mIoU scores of the proposed network on the Potsdam dataset and Jiage dataset were 85.52% and 86.59% respectively,which increased by 1.32% and 1.46%compared with the suboptimal network.
作者
刘春娟
乔泽
闫浩文
吴小所
王嘉伟
辛钰强
LIU Chun-juan;QIAO Ze;YAN Hao-wen;WU Xiao-suo;WANG Jia-wei;XIN Yu-qiang(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;School of Surveying,Mapping and Geographic Information,Lanzhou Jiaotong University,Lanzhou 730070,China;Academician Expert Workstation of Gansu Dayu Jiuzhou Space Information Technology Limited Company,Lanzhou 730070,China)
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2023年第7期1335-1344,共10页
Journal of Zhejiang University：Engineering Science
基金
甘肃省自然科学基金资助项目(21JR7RA289)
甘肃省重点研发资助项目(20YF8GA035)。
关键词
遥感图像
语义分割
多尺度互注意力
小尺度物体
编码指导上采样
remote sensing image
semantic segmentation
multi-scale mutual attention
small scale object
coding guidance upsampling
