Abstract: The expectation maximization algorithm has been classically used to find the maximum likelihood estimates of parameters in mixture probabilistic models. Problems of the EM algorithm are that parameters initialization depends on some prior knowledge, and it is easy to converge to a local maximum in the iteration process. In this paper, a new method of estimating the parameter of GMM based on split EM is proposed, it starts from a single mixture component, sequentially split and estimates the parameter of the mixture components during expectation maximization steps. Extensive experiments show the advantages and efficiency of the proposed method.

Key words: Gaussian Mixture Model（GMM）, Expectation Maximization（EM）, parameters estimation, pattern classification

摘要： 期望最大化（Expectation Maximization，EM）算法是一种求参数极大似然估计的迭代算法，常用来估计混合密度分布模型的参数。EM算法的主要问题是参数初始化依赖于先验知识且在迭代过程中容易收敛到局部极大值。提出一种新的基于分裂EM算法的GMM参数估计算法，该方法从一个确定的单高斯分布开始，在EM优化过程中逐渐分裂并估计混合分布的参数，解决了参数迭代收敛到局部极值问题。大量的实验表明，与现有的其他参数估计算法相比，算法具有较好的运算效率和估算准确性。

关键词: 高斯混合模型, 期望最大化, 参数估计, 模式分类