摘要
针对遥感图像中建筑物区域尺度跨度大且区域边界模糊导致分割精度低的问题,本文提出了一种基于交叉区域注意力的遥感建筑物分割算法.首先,设计了交叉自注意力模块和分组通道注意力模块用于建立遥感图像区域间和区域内特征的相关性表征,进而引导模型关注待分割目标的区域级细节特征与通道组选择能力;最后,针对分割结果缺乏空间相关性约束问题,提出一种区域一致性监督的损失函数,约束局部区域内像素标签分配的一致性.所提算法在WHU数据集上IoU、Precision、Recall、F1-score分别可达到91.2%、95.28%、95.4%和95.3%;在Massachusetts数据集上IoU、Precision、Recall、F1-score分别可达到74.6%、83.7%、86.9%和85.3%,各项指标均优于主流遥感图像建筑物分割算法.
Aimed at the problem of low segmentation accuracy due to the large-scale span and fuzzy boundary of buildings in remote sensing images,a cross-regional attention algorithm for remote sensing building segmentation is proposed.Firstly,a cross self-attention module and a group channel attention module are designed to establish correlation representations of inter-regional and intra-regional features of remote sensing images,which in turn guide the model to focus on region-level detail features and effective feature channel groups of the target to be segmented.Then,aiming at the problem that segmentation results lack spatial correlation constraints,a loss function of regional consistency supervision is proposed to constrain the consistency of pixel label allocation in local regions.The IoU,Precision,Recall,and F1-score of the proposed algorithm can reach 91.2%,95.28%,95.4%,and 95.3%on the WHU Building dataset,respectively.In the Massachusetts dataset,IoU,Precision,Recall,and F1-score can reach 74.6%,83.7%,86.9%,and 85.3%respectively,all of which are better than the mainstream building segmentation algorithms.
作者
邓博文
徐胜军
孟月波
刘光辉
韩九强
史亚
DENG Bowen;XU Shengjun;MENG Yuebo;LIU Guanghui;HAN Jiuqiang;SHI Ya(School of Information and Control Engineering,Xi′an University of Architecture and Technology,Xi′an 710055,China;School of Electronic Science and Engineering,Xi′an Jiaotong University,Xi′an 710049,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2024年第1期207-215,共9页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61803293,51678470)资助
陕西省自然科学基础研究计划(2020JM472,2020JM473,2019JQ760)资助
陕西省重点研发计划项目(2021SF-429)资助.
关键词
遥感图像
建筑物分割
交叉区域注意力
通道注意力
remote sensing image
building segmentation
cross-regional attention
channel attention