结合密集注意力和并行上采样的遥感图像道路分割  

Remote Sensing Image Road Segmentation Combining Intensive Attention and Parallel Upsampling

作　　者：张颖 李小霞[1,2] 李永龙 吕念祖 王皓冉 顾书豪 王学渊 ZHANG Ying;LI Xiao-xia;LI Yong-long;LV Nian-zu;WANG Hao-ran;GU Shu-hao;WANG Xue-yuan(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China;Robot Technology Used for Special Environment Key Laboratory of Sichuan Province,Mianyang 621010,China;Tsinghua Sichuan Energy Internet Research Institute,Chengdu 610042,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621010 [2]特殊环境机器人技术四川省重点实验室,四川绵阳621010 [3]清华四川能源互联网研究院,成都610042

出　　处：《小型微型计算机系统》2021年第11期2356-2361,共6页Journal of Chinese Computer Systems

基　　金：四川省科技计划项目(2020YFSY0062,2019YJ0449)资助;国家自然科学基金项目(61771411)资助.

摘　　要：针对高分辨率遥感图像中道路背景信息复杂且细节信息易丢失导致分割精度低的问题,本文提出一种结合密集注意力和并行上采样的遥感图像道路分割网络.在U型网络编码器-解码器的中间部分设计了密集空洞空间金字塔注意力模块,其中空间注意力分支可扩大有效感受野并获取密集的多尺度空间信息,通道注意力分支有利于提取全局信息并增强通道间的相关性,结合空间和通道注意力建立全局上下文依赖关系,减少背景信息的干扰;在解码器部分提出多路并行上采样模块,将不同尺度的特征图进行通道衰减后上采样恢复到原始图像大小,并行的特征图增强了分割模型结合多层次特征的能力,且更有利于道路细节信息的保持.实验结果表明,本文方法在DeepGlobe数据集上测试的召回率、准确率、精准率和F1-score分别达到0.805、0.994、0.821和0.803,各项指标以及分割效果均优于目前主流的遥感图像道路分割算法.Aiming at the problem of low segmentation accuracy due to the complicated road background information in high-resolution remote sensing images and easy loss of detailed information,this paper proposes a remote sensing image road segmentation network that combines intensive attention and parallel upsampling.A dense hollow spatial pyramid attention module is designed in the middle part of the U-shaped network encoder-decoder,in which the spatial attention branch can expand the effective receptive field and obtain dense multi-scale spatial information,and the channel attention branch is conducive to extracting global information and enhance the correlation between channels,space and channel attention is combined to establish a global context-dependence relationship and reduce the interference of background information.In the decoder part,a multi-channel parallel up-sampling module is proposed,and feature maps of different scales are attenuated,up-sampled,and restored to the original image size.The parallel feature map enhances the ability of the segmentation model to combine multi-level features,and is more conducive to the preservation of road detail information.The experimental results show that the recall rate,accuracy rate,precision rate,and F1-score of the method tested on the DeepGlobe dataset are 0.805,0.994,0.821,and 0.803,respectively,and all indicators and segmentation effects are better than the current mainstream remote sensing image road segmentation algorithm.

关 键 词：遥感图像 道路分割 密集注意力 并行上采样 通道衰减 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]