The subject of this thesis relates to variational bayesian and mean field approximation approaches for sources separation. My work is centered on the application of these methods in hyperspectral imagery with an aim of spectral reduction, segmentation and classification. My principal contributions are :
the modeling of the problem of hyperspectrales images analysis in term of sources separation and the choice of bayesian approach for the resolution of the problem: In this approach the paramount point is prior modeling the images. The model suggested is a compound markov random field for the sources with hidden labels associated for sources.
The resolution of the problem by variational approaches: Indeed the bayesien estimate makes it possible to take into account uncertainties and all prior knowns on the model of observations. But in general, and in particular with modeling by hidden field the sources which we propose the exploration of the joint posterior law or the effective calculation of the posterior estimators require approximations. For that I propose two approaches:
1) the use of the variational approach in an EM algorithm for the estimate of the mixture matrice and the hyperparameters. Indeed the step(E) of the algorithm requires an integration on the hidden variables modelled by Potts Markov model which cannot be made in an analytical way. Then I used the variationel approach or the mean field approximation for the estimate of this law.
2) a variational approach which proposes a separable law on all the parameters and the hidden variables of the proposed sources separation model. This method makes it possible to have a criterion of convergence and a criterion of selection of the model which corresponds in our case to the numbers of sources. This type of approach proves less expensive in computing times than the usual sampling procedures, such as methods MC