Experimental issues of functional merging on probability density estimation

This paper introduces a new technique for model adaptation of normal mixtures by merging their normal components. The merging technique is based on the angle (Arc-Cosine distance) between normal components in the mixture. Starting from an over-dimensioned mixture, we work out the underlying number of modes in a multimodal distribution in terms of a probabilistic measure of the best model. We illustrate the performance of functional merging on the automatic estimation of the number of lines in a degraded ancient manuscript (British library Beowulf poem) and the location of cells in microscope images.