Modélisation stochastique des images texturées

Abstract

This thesis focuses on the characterization of color textures by stochastic models in the wavelet domain. The wavelet decomposition provides a spatial frequency representation that is similar to human perception system. The work in this study concerns firstly the description of the marginal statistics of the subband textures, offering univariate best fitting to the non-Gaussian nature of the subband wavelet. In this context, we introduce the generalized Gamma model to provide more genericity and deal with the heterogeneity in image databases. In a second step, we are interested in the joint characterization by multivariate models describing the dependencies between sub-bands of color components of a texture. We propose a generic multivariate model called generalized multivariate Gamma in case the color textures are represented in the reference space, RGB and a multi-model approach in case the color textures are represented in luminance-chrominance spaces. The performance of the proposed models is experimentally evaluated based on the problem of texture classification. This requires that the modeling process considers a similarity measure on the space of the model. To do this, we propose analytic expressions metrics for the models that we offer, which represents a further contribution of this study.