非负矩阵分解在遥感图像变化检测中的应用研究

发布时间：2018-02-24 03:30

本文关键词： 变化检测非负矩阵分解聚类图像融合纹理特征灰度共生矩阵主成分分析　出处：《西南交通大学》2014年硕士论文　论文类型：学位论文

【摘要】：遥感图像的变化检测是通过对同一地区不同时期中两幅或多幅遥感图像进行比较分析,得到图像之间变化信息的一项遥感技术的应用,它目前已广泛地应用于经济和国防建设等诸多领域。非负矩阵分解(Nonnegative Matrix Factorization, NMF)算法是国际上新近提出的一种矩阵分解方法,是一种很重要的矩阵降维技术。NMF的应用领域十分广泛,如图像处理,计算机视觉,文本分析等。本文尝试将NMF运用于遥感图像的变化检测当中。主要包括以下三个内容： (1)NMF对差异图的融合。一般的变化检测常常以单一的差异图作为研究对象,而一幅差异图往往存在局限性,针对这个问题本文提出一种使用NMF融合的变化检测方法,将SAR图像的对数比值图与MRD算子图融合,将光学图像的差值图与t检验图融合。通过仿真实验得到了较好的效果,并与现有的几种变化检测方法对比,验证了该方法的有效性。 (2)基于重点关注区域的变化检测。为了降低噪声干扰引起的虚警率,以及变化信息幅度弱引起的漏检率,本文采用确定重点关注区域的方法来做变化检测。首先运用灰度共生矩阵产生差异图的纹理图像,由于方差纹理能凸显变化区域边界且可分性较强,我们采用方差纹理图来为重点关注区域的确定做铺垫。利用NMF提取纹理图的背景特征,通过计算特征图与纹理图中每个像素邻域块的欧氏距离,从而弱化纹理背景。然后将原差异图与该图像对应像素相乘,得到较理想的变化轮廓显著图。将变化轮廓图通过聚类,并膨胀填充内部区域从而得到重点关注区域。最后,根据重点关注区域修正原差异图,通过对修正后差异图的处理得到了较理想的变化检测结果。 (3)研究了NMF聚类的特性。考虑到差异图存有噪声的问题,运用双边滤波对差异图滤波,通过对像素的二阶邻域扫描得到每个像素的特征向量,然后采用PCA降维,最后利用Semi-NMF聚类得到最终变化检测结果。通过与本文前面两种算法的比较可以看到该算法具有很高的检测精度,并且通过与K-means聚类结果的对比,体现出Semi-NMF聚类的准确性。
[Abstract]:The change detection of remote sensing image is an application of remote sensing technology by comparing and analyzing two or more remote sensing images in different periods in the same area. It has been widely used in many fields, such as economy and national defense construction. Nonnegative Matrix factorization (NMF) algorithm is a recently proposed matrix decomposition method in the world. It is a very important matrix dimension reduction technology. NMF is widely used in many fields, such as image processing, computer vision, etc. Text analysis and so on. This paper tries to apply NMF in remote sensing image change detection. General change detection often takes a single difference map as the object of study, but a difference map often has some limitations. In this paper, a change detection method using NMF fusion is proposed in this paper. The logarithmic ratio graph of SAR image is fused with MRD operator graph, and the difference graph of optical image is fused with t-test graph. A good result is obtained by simulation experiment, and the validity of this method is verified by comparing it with several existing change detection methods. In order to reduce false alarm rate caused by noise interference and miss detection rate caused by weak amplitude of change information, In this paper, the method of determining the focus area is used to detect the change. Firstly, the grayscale co-occurrence matrix is used to produce the texture image of the difference map, because the variance texture can highlight the boundary of the changing region and has strong separability. We use variance texture map to lay the groundwork for determining the region of focus. We use NMF to extract the background features of texture map and calculate the Euclidean distance between each pixel neighborhood block in the feature map and texture map. In order to weaken the texture background, then multiply the original difference map with the corresponding pixels of the image, and obtain a more ideal contour salient map. The change contour map is clustered, and the inner region is filled in so as to get the focus area. Finally, According to the original difference map which focuses on the region correction, an ideal change detection result is obtained by processing the modified difference map. In this paper, the characteristics of NMF clustering are studied. Considering the existence of noise in the differential map, a two-sided filter is used to filter the difference map. The feature vector of each pixel is obtained by the second-order neighborhood scanning of the pixel, and then the dimension reduction of each pixel is achieved by using PCA. Finally, the final change detection results are obtained by using Semi-NMF clustering. By comparing with the two previous algorithms in this paper, we can see that the algorithm has a high detection accuracy, and by comparing with K-means clustering results, the accuracy of Semi-NMF clustering is reflected.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】