Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (23): 240-247.DOI: 10.3778/j.issn.1002-8331.2007-0342

• Graphics and Image Processing • Previous Articles     Next Articles

Road Extraction Network of Remote Sensing Image Based on SPUD-ResNet

LI Daidong, HE Xiaohui, LI Panle, TIAN Zhihui, ZHOU Guangsheng   

  1. 1.School of Information Engineering, Zhengzhou University, Zhengzhou 450001, China
    2.School of Earth Science and Technology, Zhengzhou University, Zhengzhou 450052, China
    3.Chinese Academy of Meteorological Sciences, Beijing 100081, China
  • Online:2021-12-01 Published:2021-12-02



  1. 1.郑州大学 信息工程学院,郑州 450001
    2.郑州大学 地球科学与技术学院,郑州 450052


A road extraction method based on SPUD-ResNet is designed and implemented to solve the problem of extracting long and narrow road structures due to the rich detail and strip distribution of road targets in remote sensing images. Firstly, the residual network encoder is constructed by using void convolution, and then connected to the corresponding decoder by skip connection, which effectively accelerates the convergence of the network and preserves the detailed road information. Secondly, in order to capture the long-distance dependence of the narrow road structure more effectively, the strip pooling module and the mixed pooling module are built respectively to enhance the ability of the network to obtain the strip road characteristics. Finally, a mixed loss function is designed by using the geometric information of the road structure and the structural similarity index to refine the road boundary and remove the fuzzy prediction from the road extracted results. Experimental results on the Massachusetts Roads dataset show that the proposed method achieves 83.4%, 84.5% and 83.9% in recall, accuracy and F1-score indicators, respectively, which improves the effect of road extraction.

Key words: remote sensing image, road extraction, strip pool, deep learning, ResNet


针对遥感影像中道路目标细节丰富且呈带状分布的特点,造成狭长的道路结构提取困难的问题,设计并实现了一种基于SPUD-ResNet的道路提取方法。该方法利用空洞卷积构建残差网络编码器,并通过跳跃连接与对应解码器相连,有效加速网络的收敛并保留道路的细节信息;为了更有效地捕获狭长道路结构的长距离依赖关系,分别构建条形池化模块和混合池化模块,增强网络对条形道路结构特征的获取能力;利用道路结构的几何信息和结构相似性指数设计混合损失函数,精细化道路边界,去除道路提取结果中的模糊预测。在Massachusetts Roads数据集上的实验结果表明,所提方法在召回率、精确度和F1-score指标分别达到了83.4%、84.5%和83.9%,提升了道路提取的效果。

关键词: 遥感影像, 道路提取, 条形池化, 深度学习, ResNet