电子学报 ›› 2018, Vol. 46 ›› Issue (12): 3021-3028.DOI: 10.3969/j.issn.0372-2112.2018.12.028

• 学术论文 • 上一篇    下一篇

可变类谱聚类遥感影像分割

李玉, 袁永华, 赵雪梅   

  1. 辽宁工程技术大学测绘与地理科学学院遥感科学与应用研究院, 辽宁阜新 123000
  • 收稿日期:2016-05-06 修回日期:2018-06-30 出版日期:2018-12-25
    • 通讯作者:
    • 李玉
    • 作者简介:
    • 袁永华 女,1991年5月出生于河南省商丘市,2014年于河南城建学院获得学士学位,现为辽宁工程技术大学在读硕士生,主要研究方向为图像分割.E-mail:790599549@qq.com;赵雪梅 女,1989年9月出生于辽宁省阜新市,2012年于辽宁工程技术大学获得学士学位,2017年于辽宁工程技术大学获得博士学位,现于中国科学院遥感与数字地球研究所做博士后研究工作,主要研究方向为遥感图像分割及基于深度学习的Landsat图像分类.E-mail:374010101@qq.com
    • 基金资助:
    • 国家自然科学基金 (No.41271435,No.41301479); 辽宁省自然科学基金 (No.2015020090)

Spectral Clustering of Variable Class for Remote Sensing Image Segmentation

LI Yu, YUAN Yong-hua, ZHAO Xue-mei   

  1. Institute for Remote Sensing Science and Application, School of Geomatics, Liaoning Technical University, Fuxin, Liaoning 123000, China
  • Received:2016-05-06 Revised:2018-06-30 Online:2018-12-25 Published:2018-12-25
    • Corresponding author:
    • LI Yu

摘要: 为实现遥感影像分割中类别数的准确、自动判别,提出了一种可变类谱聚类算法.根据影像的相似图构建权值矩阵和标准Laplacians矩阵,计算Laplacians矩阵较小特征值对应的特征向量生成特征向量矩阵,并视其与像素对应的向量行为像素特征点集;研究Laplacians矩阵处于不同(近似)块对角结构时类属同一目标类像素特征点的聚集性,定义聚类度指标,计算不同分割类别数对应聚类度;选择聚类度将发生最后一次较大跳变时的分割类别数作为算法估计类别数,并采用FCM(Fuzzy C-Means)算法划分该类别数对应像素特征点集实现影像分割.分别采用提出算法和基于特征间隙的算法分割合成及真实遥感影像.实验结果表明提出算法可准确地判别影像类别数.

关键词: 遥感影像, 可变类分割, 相似图, 谱聚类

Abstract: This paper presents a spectral clustering algorithm based image segmentation to correctly and automatically determine the number of classes. Firstly, the weighted matrix and the normalized Laplacians matrix are established with the similarity graph corresponding to a given image. Then, the eigenvectors corresponding to the smaller eigenvalues of the normalized Laplacians matrix are calculated to generate eigenvectors matrix and the pixel feature points set is constructed by means of treating each line of the eigenvectors as a different data point. Secondly, when the Laplacians matrix is in different approximate block diagonal structure, the proposed algorithm exploits the clustering property of the pixel feature points belonging to the same class and calculates the corresponding clustering degree of the different number of segmentation classes by defining the index of clustering degree. Finally, when the clustering degree is the last one to have a greater degree of jumping, the number of the segmentation classes is selected as the number of classes in this paper. The FCM algorithm is used to partition the pixel feature points set corresponding to the number of classes selected to realize the image segmentation. Synthesized and real remote sensing images are used for testing the proposed algorithm. The results show that the proposed algorithm can identify the number of classes in an image correctly.

Key words: remote sensing image, Segmentation with unknown number classes, Spectral clustering, Similarity graph

中图分类号: