Novel Initial Parameters Computation for EM algorithm-based Univariate Asymmetric Generalized Gaussian Mixture

Abstract

In histogram-based image segmentation, the Asymmetric Generalized Mixture Model (AGGMM) is a powerful tool to fit accurately the real images histograms by handling, among others, any asymmetry of the modes. However, the Expectation Maximization (EM) algorithm, used for the estimation of the mixture model parameters, is known to be very sensitive to starting conditions and can lead to erroneous segmentation results when the initialization is not adequate. In this paper, we propose a new method to initialize the AGGMM. This method is based on geometrical aspects of the histogram. First experimentations implying synthetic images generated by Asymmetric Generalized Mixture Distribution (AGGD) model, reveal a good recovering of the input mixture parameters when applying the proposed method. Second experimentations involving real-world images have shown, how the initial parameters computed by the proposed method permit to achieve better histogram fitting with less EM algorithm running time in comparison to other initialization methods.