摘要
针对在遥感影像的建筑物提取过程中,建筑物密集且离散分布带来提取效果一般的问题,采用一种特征信息增强的U-net网络。模型使用MobileNet主干网络做编码器,用于影像的建筑物特征提取,考虑到下采样时低维信息逐渐丢失,以致边缘提取效果不佳,网络结合形态学的膨胀和闭运算优化提取结果的精度。实验结果表明,在多场景高分辨率的武汉大学遥感影像建筑物数据集上,结合形态学后处理的M-Unet(MobileNet U-net)提取结果不仅在视觉效果上表现优异,而且在精确度、召回率、F1-score、平均交并比MIou(Mean Intersection over union)4个指标上分别达到96.2%、76.6%、84.6%和74.5%,均优于相同主干网络下的Pspnet和Segnet。
In the process of building extraction from remote sensing image,the problem of edge extraction precision decreased caused by intensive and discrete distribution of buildings,U-net with enhanced feature information is proposed.Model uses backbone of MobileNet as encoder,which is used to extract building feature of image.Considering that the low-dimensional information lost gradually in subsampling,which results in the extraction of building edge is not integrated,M-Unet(MobileNet U-net)combined morphological expansion and closed operation to optimize the accuracy of extraction.The experiment was carried out on the multi-scene and high-resolution remote sensing image building data set of Wuhan University,the results show that the method proposed not only performed well on visual effect,but also achieves 96.2%,76.6%,84.6%and 74.5%respectively in the four indexes of accuracy,recall rate,F1-score and MIou(Mean Intersection over union),which superior Pspnet and Segnet under the same backbone.
作者
于坤
王贺封
焦月正
李武乾
YU Kun;WANG Hefeng;JIAO Yuezheng;LI Wuqian(School of Mining and Geomatics Engineering,Hebei University of Engineering,Handan 056038,China;Handan Key Laboratory of Natural Resources Spatial Information,Handan 056038,China)
出处
《测绘与空间地理信息》
2021年第10期50-54,共5页
Geomatics & Spatial Information Technology
基金
教育部人文社科青年基金项目(19YJCZH155)
河北省高等学校科学技术研究重点项目(ZD2020312)资助。
