Unsupervised restoration in Gaussian Pairwise Mixture Model

The idea behind the Pairwise Mixture Model (PMM) we propose in this work is to classify simultaneously two sets of observations by introducing a joint prior between the two corresponding classifications and some inter-dependence between the two observations. We address the Bayesian restoration of PMM using either MPM or MAP criteria, and an EM-based parameters estimation algorithm by extending the work done for classical Mixture Model (MM). Systematic experiments conducted on simulated data shows the effectiveness of the model when compared to the MM, both in supervised and unsupervised contexts.