Non parametric stochastic expectation maximization for data clustering

We propose a novel approach to perform unsupervised and non parametric clustering of multidimensional data upon a Bayesian framework. The developed iterative approach is derived from the Classification Expectation- Maximization (CEM) algorithm [5], in which the parametric modelling of the mixture density is replaced by a non parametric modelling using local kernels, and posterior probabilities account for the coherence of current clusters through the measure of class-conditional entropies. Applications of this method to synthetic and real data including multispectral imagery are presented. Our algorithm is compared with other recent unsupervised approaches, and we show experimentally that it provides a more reliable estimation of the number of clusters while giving slightly better average rates of correct classification.