摘要
利用邻近像素类别上的相关性,在采用EM算法对模型参数求解的过程中,以滤波方法引入像素的空间位置信息,降低了EM对初始值选择的敏感性.该算法在引入了像素的位置信息的同时,保持了EM算法的简单性,并为混合分量个数的选择提供了一种新的实现途径.对实际图像的分割结果证实了算法的有效性.
Unsupervised learning of finite mixture models involves two open problems. The selection of the number of components and the initialization. To circumvent these problems in application to image segmentation, the paper integrates the filter technique into the EM algorithm. Unlike the standard EM algorithm, the proposed algorithm does not require careful initialization. It also does not need a model selection criterion to choose the suitable number of mixture components. Estimation and model selection can be integrated seamlessly in a single algorithm. Furthermore, the proposed algorithm can preserve the good traits of EM while making significant use of the spatial information in a reasonable amount of time. Experiment results on real images show that the proposed algorithm can provide fast segmentation with high perceptual quality.
出处
《计算机学报》
EI
CSCD
北大核心
2006年第6期928-935,共8页
Chinese Journal of Computers
关键词
图像分割
滤波
EM算法
混合模型
模型选择
image segmentation
filtering
EM algorithm
mixture model
model selection
