摘要
针对高分辨率遥感影像道路提取难度大、自动化程度低等问题,设计了一种改进U-Net的道路提取方法。该方法首先构建特征编码器为VGG16的VGGU-Net网络,并采用迁移权值方法初始化特征编码器;其次将提取的特征信息输入到另一U-Net网络,通过双网络联合训练方式提高网络的特征拟合能力;最后结合形态学和滤波算法对提取的道路数据进行后处理。实验结果表明:改进后算法对道路提取效果得到了有效的提升,在测试集上的准确率、召回率和IoU分别达到了93.56%、88.22%和83.17%。
Aiming at the problems that the road extraction of high-resolution remote sensing image is difficulty and low automation,a road extraction method based on U-Net is designed.Firstly,the VGGU-Net network with the feature encoder VGG16 is constructed,and the feature encoder is initialized by the migration weight method.Secondly,the extracted feature information is input to another U-Net network,and the network feature fitting ability is improved through the dual network joint training method.Finally,the extracted road data is post-processed in combination with morphological and filtering algorithms.The experimental results show that the improved algorithm has effectively improved the road extraction effect,and the test accuracy,recall rate and IoU reached 93.56%,88.22%and 83.17%respectively.
作者
金飞
王龙飞
刘智
王番
贾桂芬
JIN Fei;WANG Longfei;LIU Zhi;WANG Fan;JIA Guifen(Information Engineering University,Zhengzhou 450001,China;Dawangdian Middle School,Baoding 071000,China)
出处
《测绘科学技术学报》
北大核心
2019年第4期377-381,387,共6页
Journal of Geomatics Science and Technology
基金
信息工程大学科研项目(2016609602)
关键词
遥感影像
道路提取
迁移权值
双网络联合训练
形态学
remote sensing image
road extraction
VGGU-Net
dual network joint training
morphology