A particular Gaussian mixture model for clustering and its application to image retrieval

We introduce a new method for data clustering based on a particular Gaussian mixture model (GMM). Each cluster of data, modeled as a GMM into an input space, is interpreted as a hyperplane in a high dimensional mapping space where the underlying coefficients are found by solving a quadratic programming (QP) problem. The main contributions of this work are (1) an original probabilistic framework for GMM estimation based on QP which only requires finding the mixture parameters, (2) this QP is interpreted as the minimization of the pairwise correlations between cluster hyperplanes in a high dimensional space and (3) it is solved easily using a new decomposition algorithm involving trivial linear programming sub-problems. The validity of the method is demonstrated for clustering 2D toy examples as well as image databases.