系统工程与电子技术 ›› 2021, Vol. 43 ›› Issue (8): 2108-2115.doi: 10.12305/j.issn.1001-506X.2021.08.11

• 传感器与信号处理 • 上一篇    下一篇

基于改进的深度学习网络的SAR图像瞬时海岸线自动提取算法

王彬1,2,*, 王国宇1   

  1. 1. 中国海洋大学信息科学与工程学院, 山东 青岛 266100
    2. 青岛科技大学信息科学技术学院, 山东 青岛 266061
  • 收稿日期:2020-12-16 出版日期:2021-07-23 发布日期:2021-08-05
  • 通讯作者: 王彬
  • 作者简介:王彬(1981—), 女, 讲师, 博士, 主要研究方向为SAR图像处理与分析|王国宇(1962—), 男, 教授, 博士, 主要研究方向为图像处理与分析、模式识别
  • 基金资助:
    国家自然科学基金资助课题(61702295)

Instantaneous coastline automatic extraction algorithm for SAR images based on improved deep learning network

Bin WANG1,2,*, Guoyu WANG1   

  1. 1. College of Information Science & Engineering, Ocean University of China, Qingdao 266100, China
    2. School of Information Science & Technology, Qingdao University of Science & Technology, Qingdao 266061, China
  • Received:2020-12-16 Online:2021-07-23 Published:2021-08-05
  • Contact: Bin WANG

摘要:

针对目前合成孔径雷达(synthetic aperture radar, SAR)在对大尺度瞬时海岸线提取方面的图像解译过程中, 仍然存在精度低与自动化水平差的问题, 提出一种基于深度学习网络的瞬时海岸线自动提取算法。首先, 将SAR图像进行Lee滤波增强来抑制相干斑。其次, 通过升级残差网络为主干网络,分4级提取海水目标的特征。然后, 将4级特征经过全局卷积网络、密集连接网络和解码器网络配合,充分提取目标的本质特征, 并通过上采样产生海水分割结果。最后, 利用Sobel算子分离出海岸线并和原SAR图像融合以便清晰查看结果。通过与全卷积网络与细化网络的海岸线提取实验结果进行对比, 证明所提算法对海岸线的提取更加准确, 能够减少虚警和漏警, 具有更好的性能。

关键词: 海岸线提取, 合成孔径雷达图像阴影水体提取, 深度学习, 编码解码网络

Abstract:

Aiming at the problems of low accuracy and automation in the process of image interpretation of large-scale instantaneous coastline extraction by synthetic aperture radar (SAR), a instantaneous coastline automatic extraction algorithm based on deep learning network is proposed. Firstly, the enhancing Lee filter is used to suppressing speckle noise for SAR image. Secondly, the features of seawater targets are extracted in four levels by upgrading the residueal network as the backbone network. Then, the four levels of features pass through global convolutional network, dense connection network and decoder network to extract the essential features. The sea water segmentation results are obtained by the upsampling process. Finally, the results can be clearly viewed by separating the coastline using Sobel operator and fusing it with the original SAR image. Compared with the coastline extraction experimental results of full convolutional network and refinement network, the experimental results show that the proposed algorithm has less false alarm and miss alarm results, and the accuracy extraction of the obtained coastline results, which has better performance.

Key words: coastline extraction, synthetic aperture radar (SAR) image shadow water extraction, deep learning, coding and decoding network

中图分类号: