测绘学报 ›› 2019, Vol. 48 ›› Issue (5): 597-608.doi: 10.11947/j.AGCS.2019.20180062

• 摄影测量学与遥感 • 上一篇    下一篇

多尺度全卷积神经网络建筑物提取

崔卫红, 熊宝玉, 张丽瑶   

  1. 武汉大学遥感信息工程学院, 湖北 武汉 430079
  • 收稿日期:2018-02-10 修回日期:2018-07-23 出版日期:2019-05-20 发布日期:2019-06-05
  • 通讯作者: 熊宝玉 E-mail:xiongbbyy@foxmail.com
  • 作者简介:崔卫红(1971-),女,博士,副教授,研究方向为遥感影像信息提取与变化检测。E-mail:whcui@whu.edu.cn
  • 基金资助:
    国家自然科学基金(41101410)

Multi-scale fully convolutional neural network for building extraction

CUI Weihong, XIONG Baoyu, ZHANG Liyao   

  1. School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China
  • Received:2018-02-10 Revised:2018-07-23 Online:2019-05-20 Published:2019-06-05
  • Supported by:
    The National Natural Science Foundation of China (No.41101410)

摘要: 针对VGG16网络在高空间分辨率遥感影像中进行大型建筑物提取时存在空洞的现象,提出一种基于多尺度影像的建筑物提取方法。将原始影像进行不同尺度的下采样,提取不同尺度下的建筑物特征,并将这些多尺度特征相加合并,同时为了减少网络参数数量,用全卷积上采样过程代替原始VGG16网络中的全连接层进行建筑物提取。以0.5 m分辨率的上海市嘉定区影像和1 m分辨率的Massachusetts地区影像进行试验,精度分别达97.09%和96.66%,表明本文方法的有效性。

关键词: 大型建筑物, 多尺度, 全卷积上采样

Abstract: Some holes occurred when extracting large buildings in high spatial resolution remote sensing images with VGG16. A method of building extraction based on multi-scale features is proposed to solve this problem. Firstly, the original images were downsampled at different scales. Then, it could be extracted that the features of buildings at different scales and fused them. To reduce the number of network parameters, the fully convolutional upsampling was used to replace the fully connected layer in the original VGG16 model. The study images were from the 0.5 m resolution in Jading of Shanghai and 1 m resolution Massachusetts building dataset. The accuracy of buildings extraction were 97.09% and 96.66% respectively. The result showed the effectiveness of the proposed method.

Key words: large buildings, multi-scale, fully convolutional upsampling

中图分类号: