This work deals with the statistical restoration of a hidden signal using Pairwise Markov Trees (PMT). The latter PMT, recently introduced in the case of discrete hidden signal, are here applied to unsupervised image segmentation and it is showed that they work better then the classical Hidden Markov Trees (HMT). Further, considering a PMT in a linear Gaussian model with continuous hidden data, which is new, we give the formulas of an original extension of the classical Kalman filter.