作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (8): 196-205. doi: 10.19678/j.issn.1000-3428.0062451

• 图形图像处理 • 上一篇    下一篇

融合多尺度特征的遥感影像道路提取方法

赫晓慧1,3, 宋定君2, 李盼乐2, 田智慧1,3, 周广胜3   

  1. 1. 郑州大学 地球科学与技术学院, 郑州 450052;
    2. 郑州大学 信息工程学院, 郑州 450001;
    3. 中国气象科学研究院郑州大学生态气象联合实验室 郑州 450052
  • 收稿日期:2021-08-23 修回日期:2021-10-09 发布日期:2021-10-11
  • 作者简介:赫晓慧(1978-),女,教授、博士生导师,主研方向为人工智能、遥感影像处理、数据挖掘;宋定君,硕士研究生;李盼乐,博士;田智慧、周广胜,教授、博士。
  • 基金资助:
    河南省重大科技专项(201400210900)。

Remote Sensing Image Road Extraction Method Combined with Multi-Scale Features

HE Xiaohui1,3, SONG Dingjun2, LI Panle2, TIAN Zhihui1,3, ZHOU Guangsheng3   

  1. 1. School of Earth Sciences and Technology, Zhengzhou University, Zhengzhou 450052, China;
    2. School of Information Engineering, Zhengzhou University, Zhengzhou 450001, China;
    3. Laboratory of Ecological Meteorology Jointly Hosted by Zhengzhou University and Chinese Academy of Meteorological Sciences, Zhengzhou 450052, China
  • Received:2021-08-23 Revised:2021-10-09 Published:2021-10-11

摘要: 针对遥感影像中由于道路信息错综复杂,导致道路提取不完整、精确度低等问题,提出一种新型遥感影像道路提取方法。融合多尺度特征改善道路提取的整体效果,基于深度残差网络设计混合空洞卷积,并通过定义卷积核各值的间距增大特征提取感受野,从而丰富多尺度特征融合模块中的浅层道路语义信息。在编码端提取特征后,利用权重分布的方法匹配感受野尺度,使用不同层级间的特征对全局先验信息进行表征,提高多尺度融合特征获取浅层道路语义信息的准确性,并将改进的多孔空间金字塔池化模块融入到深度残差网络中,挖掘并深度聚合道路浅层次和深层次的语义信息。在两种高分辨率遥感数据集Cheng_Data_Roads和Zimbabwe_Data_Roads上的实验结果表明,所提方法的F1值和MIoU值分别为91.16%和83.63%,准确率、召回率等评价指标均明显优于U-net、ResUnet、D-Linknet等语义分割方法。

关键词: 深度学习, 空洞卷积, 深度残差网络, 多尺度特征融合, 遥感影像道路

Abstract: This paper proposes a new road extraction method from remote sensing images that targets the accuracy problems arising from intricate road information.The overall effect of road extraction is improved by fusing multi-scale features.The hybrid hole convolution is designed based on a deep residual network, and the feature extraction receptive field is increased by defining the spacing of each value in the convolution kernel, thereby enriching the shallow road semantic information in the multi-scale feature fusion module.After feature extraction of the encoder, the weight distribution method is used to match the scale of the receptive field, and the features at different levels areused to characterize global prior information, thus improving the accuracy of multi-scale fusion features to obtain semantic information from shallow roads.The improved porous spatial pyramid pooling module is integrated into the deep residual network to mine and deeply aggregate deep semantic information from roads.Experimental results on two high-resolution remote sensing datasets of Cheng_Data_Roads and Zimbabwe_Data_Roads, show that the proposed method achieves improvements of 91.16% and 83.63% for the F1 and MIoU values, respectively.Furthermore, the accuracy, recall rate, and other evaluation metricsare significantly higher than those of U-net, ResUnet, D-Linknet, and other semantic segmentation methods.

Key words: deep learning, atrous convolution, deep residual network, multi-scale feature fusion, remote sensing image road

中图分类号: