结合EM/MPM算法和Voronoi划分的图像分割方法

Combining the EM/MPM and Voronoi Tessellation for Image Segmentation

为了在模型参数先验分布知识未知情况下实现基于区域和统计的图像分割,并同时获取更加精确的模型参数,提出了一种结合Voronoi划分技术、最大期望值(Expectation Maximization,EM)和最大边缘概率(Maximizationof the Posterior Marginal,MPM)算法的图像分割方法。该方法利用Voronoi划分技术将图像域划分为若干子区域,待分割图像中的同质区域可以由一组子区域拟合而成,并假定各同质区域内像素强度服从同一独立的正态分布,从而建立图像模型,然后结合EM/MPM算法进行图像分割和模型参数估计,其中,MPM算法用于实现面向同质区域的图像分割,EM算法用于估计图像模型参数。为了验证提出的图像分割方法,分别对合成图像和真实图像进行了分割实验,并和传统的基于像素的MRF分割结果进行对比,测试结果的定性和定量分析表明了该方法的有效性和准确性。

In order to realize image segmentation without prior knowledge on model parameters, this paper presents a new region and statistics based approach, which combines Voronoi tessellation technique and expectation-maximization/maximi- zation of the posterior marginal （EM/MPM） algorithm. By Voronoi tessellation, the domain of a give image is partitioned into ~oronoi polygons. The homogeneous regions in which the pixel intensities are assumed to follow independent and iden- tical Gaussian distribution are constructed by a subset of polygons. After an image model is built, EM/MPM scheme is uti- lized to govern image segmentation, where the MPM criterion is for segmentation and the EM algorithm is for parameter esti- mation. To verify the validness of the proposed algorithm, testing is carried out with synthetic and real images, respective- ly, and the results show that the proposed algorithm works well on images segmentation. This paper also compares segmen- tation results of real and synthetic images using the proposed approach and pixel based MRF approach. As a result, the Voronoi polygons fitting homogeneous regions is more accurate and proper to trace the texture structures.