结合特征场和标号场MRF的区域化纹理图像分割  

Texture Segmentation Using Gaussian Markov Random Field and Region Tessellation

作　　者：杨圣 YANG Sheng(College of Surveying, Mapping and Geographical Science, Liaoning Technical University,Fuxin 103000,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,辽宁阜新123000

出　　处：《测绘与空间地理信息》2018年第5期66-72,共7页Geomatics & Spatial Information Technology

摘　　要：为了实现对纹理图像的分割,需利用建模像素间相互作用关系,因此本文利用在标号场和特征场中分别建模邻域多边形和邻域像素之间的作用关系,并提出一种基于马尔科夫随机场(Markov Random Field,MRF)的区域化纹理图像分割方法。即利用Voronoi划分技术,将图像划分为若干个多边形;在标号场上利用Gibbs分布建模相邻多边形标号间的相互作用,在特征场上利用高斯分布建模多边形内邻域像素间光谱测度的相关性;结合贝叶斯定理建立分割模型;通过最大期望值(Expectation Maximization,EM)算法来估计模型参数,进而获得最优分割结果。本文分别对合成纹理图像、自然纹理图像和遥感图像进行分割实验,并对分割结果进行定性和定量评价。通过计算混淆矩阵得出Kappa值为0.97,满足了优秀分类器的标准。本文提出的算法具有很强的抗噪和描述复杂光谱测度的能力,可行性好,准确性高。In order to solve this problem,a region-based algorithm is proposed. The proposed algorithm can describe the spatial relation of neighbor pixels and can also guarantee that the labels of neighbor pixels are the same simultaneously. First,the Voronoi technology is applied to divide the image domain into a set of non-overlapping polygons. Second,MRF model and Gibbs distribution are used to describe the relationship between neighbor polygons. In this way,the label field is constructed. Third,the Gaussian distribution is used to describe the texture structure in a polygon. Fourth,the proposed algorithm is considered that pixels in the same polygon belong to the same label. So the MRF model is used to describe the relationship between neighbor pixels with the same pixel label.Then the feature field is established. Finally,the segmentation model is obtained by using the Bayesian paradigm to combine the label field and the feature field. The parameters of the segmentation model are estimated by using the EM（ Expectation Maximization） algorithm. In this paper,the synthetic texture images,natural texture images and remote sensing images are segmented. And the segmentation results are evaluated qualitatively and quantitatively. By calculating the confusion matrix,the Kappa value is 0.97,which meets the standard of the excellent classifier. Solve this problem,which is ensure that the labels of the neighbor pixels are same when segmenting the image. And it has the strong ability to anti noise and describe the complex spectrum. The proposed algorithm also has the good feasibility,high accuracy.

关 键 词：纹理图像分 高斯马尔可夫随机场 VORONOI划分 EM/MPM 

分 类 号：P237[天文地球—摄影测量与遥感]