测绘学报 ›› 2022, Vol. 51 ›› Issue (1): 135-144.doi: 10.11947/j.AGCS.2021.20200508

• 影像处理与重建 • 上一篇    下一篇

高分辨率遥感影像建筑物提取多路径RSU网络法

张玉鑫, 颜青松, 邓非   

  1. 武汉大学测绘学院, 湖北 武汉 430079
  • 收稿日期:2020-10-14 修回日期:2021-07-21 发布日期:2022-02-15
  • 通讯作者: 邓非 E-mail:fdeng@sgg.whu.edu.cn
  • 作者简介:张玉鑫(1997-),女,硕士,主要研究方向为语义分割和建筑物提取。E-mail:zhangyuxin_whu@whu.edu.cn
  • 基金资助:
    四川省科技计划(2019YFG0460)

Multi-path RSU network method for high-resolution remote sensing image building extraction

ZHANG Yuxin, YAN Qingsong, DENG Fei   

  1. School of Geodesy and Geomatics, Wuhan University, Wuhan 430079, China
  • Received:2020-10-14 Revised:2021-07-21 Published:2022-02-15
  • Supported by:
    Sichuan Science and Technology Program(No. 2019YFG0460)

摘要: 针对卷积神经网络在提取建筑物的过程中,存在建筑物边界不准确和建筑物内部空洞等问题,提出以RSU模块(residual U-block)为核心的MPRSU-Net (multi-path residual U-block network)。该模块利用编码器-解码器结构和残差连接,实现了局部特征和多尺度特征的融合。由于一个RSU模块提取的信息有限,MPRSU-Net进一步通过多路径结构并行了不同尺度的RSU模块,并在这些模块之间进行信息交换,提高了特征聚集效率。在分辨率为0.3 m的WHU和Inria建筑物数据集上进行试验,精度分别达95.65%和88.63%,IoU分别达91.17%和79.31%,验证了本文方法的有效性。此外,本文方法相较于U2Net,计算量明显降低,模型参数量减少68.63%,表明本文方法具有一定的应用价值。

关键词: 高分辨率遥感影像, 建筑物提取, 多尺度, 卷积神经网络, 多路径

Abstract: Inaccurate boundaries and holes are two major problems when extracting buildings from high-resolution remote sensing images by a convolution network. To solve these problems, we proposed the MPRSU-Net (multi-path residual U-block network), which is based on the RSU (residual U-block). The RSU is able to fuse local features and multi-scale features, with the help of the encoder-decoder structure and the residual connection. However, a single RSU is not enough to gather enough information, MPRSU-Net parallels RSU blocks of different scales by the multi-path structure and exchanges information among these blocks to further enhance the feature aggregation efficiency. Experimental results showed that the MPRSU-Net achieved 95.65%,88.63% precision, and 91.17%,79.31% IoU on 0.3 m resolution WHU and Inria building datasets, which showed the effectiveness of the proposed method. In addition, compared with the U2 Net, MPRSU-Net is much lighter in computation and reduces the amount of model parameters by 68.63%, demonstrating that the method has some application value.

Key words: high-resolution remote sensing image, building extraction, multi-scale, convolutional neural networks, multi-path

中图分类号: