Joint segmentation of images with non Gaussian mixture models

The idea behind the Pairwise Mixture Model is to classify simultaneously two sets of observations by introducing a joint prior between the two corresponding classifications and some statistical relations between the two observations. We address both the Gaussian case and non-Gaussian parametric case built with copula-based parametric models and non-Gaussian margins. We also provide EM and ICE algorithms for automatic parameters estimation in order to make classification algorithms unsupervised. The model is illustrated through the segmentation of vectorial images (color and IRM). Results are compared to the segmentations obtained using independent mixture models on individual bands.