Deterministic annealing em and its application in natural image segmentation

Abstract

In this paper, we present a color image segmentation algorithm based on a finite mixture model and examine its application to natural scene segmentation. Gaussian mixture model (GMM) is first adopted to represent the statistical distribution of multi-colored objects. Then a deterministic annealing Expectation Maximization (DAEM) formula is used to estimate the parameters of the GMM. The experimental results show that the proposed DAEM can avoid the initialization problem unlike the standard EM algorithm during the maximum likelihood (ML) parameter estimation and natural scenes containing texts are segmented more efficiently than the existing EM technique. © Springer-Verlag 2004.