计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (9): 219-229.DOI: 10.3778/j.issn.1002-8331.2010-0134

• 图形图像处理 • 上一篇    下一篇

融合BOVW和复杂网络的高光谱遥感图像分类

宁晨,谢红薇,孟丽楠   

  1. 太原理工大学 软件学院,太原 030024
  • 出版日期:2022-05-01 发布日期:2022-05-01

Hyperspectral Remote Sensing Image Classification Based on BOVW and Complex Networks

NING Chen, XIE Hongwei, MENG Linan   

  1. College of Software, Taiyuan University of Technology, Taiyuan 030024, China
  • Online:2022-05-01 Published:2022-05-01

摘要: 为挖掘高光谱遥感图像的深层光谱特征,获取优化特征空间以提高分类准确率,提出了一种基于视觉词典和复杂网络的高光谱遥感图像分类的光谱特征提取方法。通过改进视觉词典方法,使用K-Means方法计算各类样本的聚类中心作为词典,并计算各待测试样本的光谱像素值与词典光谱向量中相同光谱波段的差值,计算出单个待测样本点的词频直方图。同时,为提升所提取特征关于像素点波段之间关联的表达,引入复杂网络理论,将样本点光谱向量矩阵化,利用像素点的矩阵位置和像素值构建复杂网络,再对网络进行阈值动态演化,并提取各个子网络拓扑特征。融合二者所提取的特征进行分类。在Salinas和KSC高光谱遥感数据集上的实验结果表明,该算法都可以取得更优的分类效果。

关键词: 高光谱遥感图像, 视觉词典(BOVW), 复杂网络, 图像分类

Abstract: For the sake of mining the deep features of hyperspectral remote sensing images(HSRSI) and obtaining the optimized feature space to improve the classification accuracy, a spectral feature extraction method combining bag of visual words and complex networks(BOVW-CN) is proposed. By improving BOVW method, K-Means method is used to calculate the clustering centers of all kinds of samples as dictionaries, and the difference between the spectral pixel value of each sample to be tested and the same spectral band in the codebook spectral vectors is calculated, and word frequency histogram of a single sample is calculated. Meanwhile, in order to improve the expression of the extracted features about the correlation between the pixel bands, the complex networks theory is applied. The spectral vector of the sample points is matrixed, and the complex network is constructed by using the matrix position and pixel value of the pixels. Then, the threshold value of the network is dynamically evolved, and the topological features of each sub network are extracted. Finally, the features extracted from them are fused for classification. Experimental results on Salinas and KSC hyperspectral remote sensing data sets show that the algorithm can achieve better classification results.

Key words: hyperspectral remote sensing image, bag of visual words(BOVW), complex networks, image classification