电子学报 ›› 2020, Vol. 48 ›› Issue (1): 66-74.DOI: 10.3969/j.issn.0372-2112.2020.01.008

• 学术论文 • 上一篇    下一篇

基于深度卷积神经网络的图上半监督极化SAR图像分类算法

魏志强1, 毕海霞2, 刘霞3   

  1. 1. 西安电子工程研究所, 陕西西安 710100;
    2. 西安交通大学电子与信息工程学院, 陕西西安 710049;
    3. 西安理工大学应用数学系, 陕西西安 713300
  • 收稿日期:2018-11-14 修回日期:2019-04-16 出版日期:2020-01-25
    • 通讯作者:
    • 毕海霞
    • 作者简介:
    • 魏志强 男,1974年12月出生,安徽利辛人.毕业于复旦大学获博士学位,现为西安电子工程研究所研究员,主要研究方向为雷达系统工程、太赫兹技术、SAR图像处理.E-mail:zqwei@fudan.edu.cn;刘霞 女,1984年1月出生,陕西榆林人.2015年毕业于西安交通大学获博士学位,现为西安理工大学讲师.主要研究方向为机器学习、学习理论、非线性泛函分析.E-mail:liuxia1232007@163.com
    • 基金资助:
    • 国家自然科学基金 (No.61806162)

A Graph-Based Semi-Supervised PolSAR Image Classification Method Using Deep Convolutional Neural Networks

WEI Zhi-qiang1, BI Hai-xia2, LIU Xia3   

  1. 1. Xi'an Electronic Engineering Research Institute, Xi'an, Shaanxi 710100, China;
    2. School of Electronic and Information Engineering, Xi'an Jiaotong University, Xi'an, Shaanxi 710049, China;
    3. Department of Applied Mathematics, Xi'an University of Technology, Xi'an, Shaanxi 713300, China
  • Received:2018-11-14 Revised:2019-04-16 Online:2020-01-25 Published:2020-01-25
    • Corresponding author:
    • BI Hai-xia
    • Supported by:
    • National Natural Science Foundation of China (No.61806162)

摘要: 为实现在只有少量标记数据情况下的高质量的图像分类,本文提出了一种基于深度卷积神经网络的图上半监督极化SAR图像分类算法.该算法将极化SAR图像建模为无向图,并基于该无向图,定义了包含半监督项,卷积神经网络项和类标光滑项的能量函数.算法所采用的卷积神经网络提取抽象的数据驱动的极化特征.半监督项约束了有标记像素的类标在分类过程中保持不变.类标光滑项约束了像素间类标的光滑性.基于对PauliRGB图像进行超像素分割而产生的初始化类标图,交替迭代优化所定义的能量函数直至其收敛.在两幅真实极化SAR图像上的实验结果表明,该算法达到了优异的分类效果,其性能优于当前已有算法.

关键词: 极化SAR图像分类, 半监督, 卷积神经网络, 图模型

Abstract: To realize high quality image classification with few labeled data, a graph-based semi-supervised PolSAR image classification method using deep neural networks is proposed in this paper. The PolSAR image is modeled as a graph,based on which we design an energy function which incorporates a semi-supervision term, a convolutional neural network (CNN) term and a pairwise smoothness term. CNN is responsible for extracting discriminative polarimetric features. The semi-supervision term enforces that class information of labeled pixels keep fixed during the classification. The pairwise smoothness term enforces class label smoothness. Started from an initialized class label map generated using superpixel segmentation of polarimetric PauliRGB image, the proposed method iteratively optimizes the energy function until the model converges. Experimental results conducted on two benchmark PolSAR images show that our approach effectively improves the classification accuracy with limited numbers of labeled pixels.

Key words: polarimetric synthetic aperture radar (PolSAR) image classification, semi-supervised method, convolutional neural network (CNN), graph model

中图分类号: