Bayesian texture model selection by harmonic mean

The paper presents a model selection method for texture images, more specifically, it finds the most adequate model for the pixels´ interaction. This approach relies on a Bayesian framework, that probabilities all the quantities and determines the joint a posteriori law for the models and the parameters. In order to compute the a posteriori model probabilities, the model parameters are marginalized by means of sampling, performed independently for each model in a within-model sampling strategy using a Metropolis-Hastings (M-H) algorithm. The resulting chains are used to compute the evidence of each model by an harmonic averaging of the likelihoods computed for the aforementioned sampled values. The work presented in the following consists in a complex and comprehensive formalism based on state of the art methods for parameter estimation, model selection techniques and sampling algorithms, the novelty being the design of such an approach for a texture model selection problem. An image processing application of this kind raises serious difficulties regarding the large amount of data, the data correlation and the highly non-linear dependencies of the data with respect to the parameters. Despite these challenges, our method successfully solves the problem of texture model selection and parameter estimation.