Robust mixture decomposition via maximization of trimmed log-likelihood with application to image database organization

We present a robust mixture decomposition technique that also finds the number of components required to model the mixture automatically. The algorithm maximizes the trimmed log-likelihood to achieve robustness, and finds the 'optimal' number of components by maximizing the entropy of the component cardinalities. We present results on synthetic and real data to illustrate the performance of the algorithm under noisy conditions. Preliminary results indicate that the proposed algorithm can be used to organize image databases for efficient retrieval.