基于Voronoi几何划分和层次化建模的纹理影像分割

Voronoi tessellation and hierarchical model based texture image segmentation

将基于像素MRF分割方法拓展到基于地物目标几何约束的区域MRF分割,提出了一种基于区域和统计的纹理影像分割方法,其基本思想是利用Voronoi划分技术将影像域划分为若干子区域。在此基础上,采用二值高斯马尔科夫随机场(BGMRF,bivariate Gaussian Markov random field)模型,静态随机场模型和Potts模型从邻域、区域及全局层次描述影像的纹理结构,并将该纹理结构模型纳入贝叶斯框架;依据贝叶斯定理构建纹理影像分割模型;利用metropolis-hastings(M-H)算法进行模型参数估计,并依据最大后验概率(MAP,maximum a posterior)准则进行优化,从而完成纹理影像分割。为了验证所提出方法的正确性,分别对合成纹理影像,真实纹理影像及遥感影像进行了分割实验,定性和定量的测试结果验证了提出方法的有效性、可靠性和准确性。

A regional and statistical based algorithm for texture image segmentation was proposed. The Voronoi tessella- tion was used for partitioning the domain of an image into sub-regions corresponding to the components of homogenous regions, to which the texture image needs to be segmented. Bivariate Ganssian Markov random field （BGMRF） model, static random field, and potts model were employed to characterize the interactions between two neighbor pixel pairs in a sub-region, and among sub-regions, respectively. Following Bayesian paradigm, a posterior distribution, which models the texture segmentation for a given texture image, was obtained. A metropolis-hastings algorithm was designed for simulating the posterior distribution. Then, texture segmentation was obtained by maximum a posterior （MAP） scheme. The proposed algorithm was tested with both of synthesized and real texture images. The results are qualitatively and quantitatively evaluated and show that the proposed algorithm works well on both of texture images.