A Bayesian approach to 2D non minimum phase AR identification

The authors deal with estimation of autoregressive (AR) noncausal models of bidimensional signals. The problem of factorizing an image into an excitation with a given marginal p.d.f. and a IIR filter is formulated in a Bayesian conceptual framework. The proposed solution is an iterative procedure for the minimization of the a posteriori risk associated to a given cost function. The procedure implies the inversion of a Toeplitz-block-Toeplitz covariance matrix and the iterated solution of a set of normal equations associated with a nonlinear estimation stage.<>