基于多尺度融合注意力改进UNet的遥感图像水体分割  

Water body segmentation in remote sensing images based on multi-scale fusion attention module improved UNet

作　　者：石甜甜 郭中华 闫翔[1,2] 魏士钦 SHI Tian-tian;GUO Zhong-hua;YAN Xiang;WEI Shi-qin(School of Physics and Electronic-Electrical Engineering,Ningxia University,Yinchuan 750021,China;Ningxia Key Lab on Information Sensing&Intelligent Desert,Ningxia University,Yinchuan 750021,China)

机构地区：[1]宁夏大学物理与电子电气工程学院,宁夏银川750021 [2]宁夏大学沙漠信息智能感知重点实验室,宁夏银川750021

出　　处：《液晶与显示》2023年第3期397-408,共12页Chinese Journal of Liquid Crystals and Displays

基　　金：宁夏自然科学基金(No.2020AAC03026);宁夏大学研究生创新研究项目(No.GIP2021006,No.CXXM202221)。

摘　　要：针对遥感图像水体分割任务,提出了一种多尺度融合注意力模块改进的UNet网络——A-MSFAM-UNet,该方法在GF-2遥感图像水体分割任务中实现了端到端高分辨率遥感图像水体分割。首先,针对以往注意力模块全局池化操作带来的局部信息不敏感问题,设计了一种多尺度融合注意力模块(MSFAM),该模块使用点卷积融合通道全局信息、深度可分离卷积弥补全局池化造成的信息丢失。MSFAM用于UNet跳跃连接后的特征融合部分重新分配特征点权重以提高特征融合效率,增强网络获取不同尺度信息的能力。其次,空洞卷积用于VGG16主干网络扩展感受野,在不损失分辨率的情况下聚合全局信息。结果表明,A-MSFAM-UNet优于其他通道注意力(SENet、ECANet)改进的UNet,在GF-2水体分割数据集上平均交并比(MIoU)、平均像素精度(MPA)和准确率(Acc)分别达到了96.02%、97.98%和99.26%。A multi-scale fusion attention module improved UNet network is proposed for the water body segmentation task of remote sensing images,A-MSFAM-UNet,which achieves end-to-end high-resolution remote sensing images in the GF-2 remote sensing image water body segmentation task.Firstly,aiming at insensitivity of local information caused by global pooling operation of previous attention module,a multiscale fusion attention module(MSFAM)is designed,which uses point convolution to fuse channel global information and depthwise separable convolution.The loss of information caused by global pooling is made up.MSFAM is adopted to redistribute the weights of feature points after UNet skip connection to improve the efficiency of feature fusion and enhance network ability to obtain information at different scales.Secondly,the atrous convolution is applied to VGG16 backbone network to expand receptive field and aggregate global information without loss of resolution.The results show that A-MSFAM-UNet outperforms other channel attention(SENet,ECANet)improved UNet,and achieves mean intersection over union(MIoU)、mean pixel accruary(MPA)and accuracy(Acc)of 96.02%,97.98%and 99.26%on the GF-2 water body segmentation dataset.

关 键 词：遥感图像 注意力模块 深度可分离卷积 特征融合 空洞卷积 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置] TP391[自动化与计算机技术—控制科学与工程]