Image segmentation based on multiscale initialized Gaussian mixtures

Abstract

Image segmentation is a key step for image processing and Gaussian Mixture Models(GMMs) are the common models for segmentation. The EM algorithm is usually used to estimete the parameters of GMMs, which is opt to get stuck at local minimum. In this paper we propose a new initialized shceme, multiscale online learning, for EM to aviod local minima and for GMMs to decide the optimal initial number of components. Experimental results have shown that this scheme can effectively improve the precision of segmentation compared to classical EM algorithm. © 2013 Springer-Verlag.