A comparison of EM and GMVQ in estimating Gauss mixtures: application to probabilistic image retrieval

Expectation-maximization (EM) is the dominant algorithm for estimating the parameters of a Gauss mixture (GM). Recently, Gauss mixture vector quantization (GMVQ) based on the Lloyd algorithm has been applied successfully as an alternative for both compression and classification. We investigate the performance of the two algorithms for GMs in image retrieval. The asymptotic likelihood approximation is used as a similarity criterion to compare GMs directly. The two algorithms result in very close retrieval performance. We demonstrate that the closeness comes from the close mutual approximation of the GM estimated parameter values and that the two algorithms have similar convergence speed. Our analysis shows that GMVQ has roughly half the computational complexity of EM.