基于改进双U-Net的遥感影像道路提取方法  

Road extraction from remote sensing image based on improved dual U-NET

作　　者：徐丽[1] 王铭磊 屈立成[1] XU Li;WANG Ming-lei;QU Li-cheng(School of Information Engineering,Chang’an University,Xi’an 710064,China)

机构地区：[1]长安大学信息工程学院,西安710064

出　　处：《信息技术》2021年第10期20-25,30,共7页Information Technology

基　　金：陕西省自然科学基础研究计划资助项目(2020JM-258)。

摘　　要：针对从地物信息复杂的遥感影像中提取道路准确率较低的问题,提出了一种改进双U-Net的路网提取方法。在两个U-Net间添加上下文特征提取模块进行连接;通过使用Log-Cosh方法对IoU损失进行平滑处理,并与二元交叉熵损失加权相加作为模型的损失函数;将数据分为无道路、简单道路、复杂道路、阴影遮挡道路、立体交通道路和含干扰项道路等,并选取后三类图像作为测试集。实验表明,改进的U型网络模型与经典图像分割网络相比,路网提取的准确率提升超过0.3%,IoU和F1指数提升均超过了1%,能够有效地在地物信息复杂的遥感影像中提取道路。Regarding the problem of low accuracy in road extraction from remote sensing images under complex ground information,a road network extraction method based on improved dual u-net is proposed.A context feature extraction module is added to connect the two u-nets.The Log-Cosh method is used to smooth the IoU loss,and the weighted sum of the loss and the binary cross entropy loss is used as the loss function of the model.The data is divided into no road,simple road,complex road,shaded Road,three-dimensional traffic road and road with interference,and the last three kinds of images are selected as the test set.The experiment results show that the accuracy of road network extraction is improved by more than 0.3%,and the IoU and F1 indexes are improved by more than 1%compared with the classical image segmentation network,which can effectively extract roads from remote sensing images with complex ground information.

关 键 词：遥感影像 道路提取 双U-Net 上下文特征 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]