Pararmeter estimation in gibbs-markov image models

A. El-Baz, A.A. Farag
2003 Sixth International Conference of Information Fusion, 2003. Proceedings of the  
This paper introduces two novel approaches to estimate the clique potentials in discrete and multilevel realizations of Gibbs Markov random field (GMRF) models. The first approach employs a genetic algorithm (GA) in order to arrive at the closest synthesized image that resembles the original "observed" image. The Second approach is used to estimate the parameters of Gaussian Markov random field. Given an image formed of a number of classes, an initial class density is assumed and the parameters
more » ... and the parameters of the densities are estimated using the EM approach. Convergence to the true distribution is tested using the Levy distance. The segmentation of classes is performed iteratively using the ICM algorithm and a genetic algorithm (GA) search approach that provides the maximum a posteriori probability of pixel classification. During the iterations, the GA approach is used to select the clique potentials of the Gibbs-Markov models used for the observed image. The algorithm stops when a fitness function, equivalent to the maximum a posteriori probability, does not change. The approach has been tested on synthetic data and is shown to provide satisfactory results.
doi:10.1109/icif.2003.177340 fatcat:parsa3uisre55lveku5mjhnvpe