Experiments in Simulating Natural Textures
From here on, we start experimental investigations of the proposed Gibbs image models with multiple pairwise pixel interaction and parameter learning schemes. This chapter presents results of simulation of various natural image textures taken from the MIT Media Laboratory “VisTex” image database (Pickard et al., 1995) or obtained by digitizing some examples from the album of Brodatz (1966). First, we consider the self-similarity between the different patches of visually homogeneous textures that can be revealed. It follows that the self-similarity is quantitatively described by the model-based interaction maps and resulting interaction structures and Gibbs potentials. Then multiple simulation results obtained by the CSA are presented to show that homogeneous, or stochastic textures may be separated from textures which are weakly homogeneous and inhomogeneous in terms of our models by comparing the training and simulated samples.
KeywordsTraining Sample Interaction Structure Stochastic Approximation Simulated Sample Gibbs Potential
Unable to display preview. Download preview PDF.