Variational Bayesian Image Modelling.pdf

Variational Bayesian Image Modelling Li Cheng LICHENG @CS .UALBERTA .CA Department of Computing Science, University of Alberta, Canada Feng Jiao FJIAO @UWATERLOO .CA School of Computer Science, University of Waterloo, Canada Dale Schuurmans DALE @CS .UALBERTA .CA Shaojun Wang SWANG@CS .UALBERTA .CA Department of Computing Science, University of Alberta, Canada Abstract exhibit strong spatial correlations (Besag, 1986). We present a variational Bayesian framework Hidden Markov Random Fields (HMRF) are a particu- for performing inference, density estimation and larly natural model to apply in domains with numerous, model selection in a special class of graphical correlated hidden variables of this form, and have been models—Hidden Markov Random Fields (HM- extensively applied in areas such as computational vision RFs). HMRFs are particularly well suited to im- (Forbes Peyrard, 2003; Heitz Bouthemy, 1993). How- age modelling and in this paper, we apply them ever, because of the large number of hidden variables and to the problem of image segmentation. Unfor- their complex graphical structure, density estimation in tunately, HMRFs are notoriously hard to train these models is computationally hard. The problem is made and use because the exact inference problems even harder by the f