结合双重注意力机制的遥感图像道路分割
Road segmentation of remote sensing images combined with dual attention mechanism
龚轩 1郭中华 2丁荣荣 2顾旭璐 2闫梓旭3
作者信息
- 1. 宁夏大学电子与电气工程学院,宁夏银川750021;宁夏大气探测技术保障中心,宁夏银川750002
- 2. 宁夏大学电子与电气工程学院,宁夏银川750021
- 3. 宁夏计量质量检验检测研究院,宁夏银川750021
- 折叠
摘要
为解决光学遥感图像道路分割所存在的漏判、误判等问题,提出了一种改进U型网络结构的语义分割模型,融入双重通道注意力机制和改进空间金字塔池化结构的残差特征提取U型网络(RSD-UNet).首先,编码模块采用具有残差结构的ResNet—34,避免神经网络出现梯度消失;其次,融入串行改进的SPPCSPC池化模块,提高网络的感受野、解决道路特征多尺度问题;最后,在上采样操作后融入多频谱通道和空间的维度的双重注意力机制(DAM).实验结果表明:在CHN6—CUG数据集上,对比基准网络UNet,指标IoU和F1分数提高了4.4％和3.07％.因此,RSD-UNet能够较好实现对于光学遥感图像道路分割.
Abstract
A new semantic segmentation model,ResNet-SPPCSPC-dual channel attention-UNet (RSD-UNet ),based on an improved U network structure,is proposed to solve the problem of false negative and misjudgement in road segmentation of optical remote sensing images.Firstly,the coding module adopts ResNet-34 with residual structure to avoid gradient disappearance of neural network.Secondly,serial modified SPPCSPC pooling module is integrated to improve the receptive field of the network and solve the multi-scale problem of road characteristics.Finally,the dual attention mechanism (DAM)of multi-spectral channels and spatial dimensions is integrated after the up-sampling operation.The experimental results show that on the CHN6-CUG dataset,compared with the benchmark network UNet,the index IoU and F1 scores increase by 4.4％ and 3.07％.Therefore,RSD-UNet can better achieve road segmentation for optical remote sensing images.
关键词
双重注意力机制/遥感图像/道路分割Key words
dual attention mechanism/remote sensing image/road segmentation引用本文复制引用
基金项目
宁夏自然科学基金资助项目(2020AAC03026)
宁夏大学研究生创新研究项目(CXXM202221)
出版年
2024
NETL
NSTL
万方数据