An important problem in multiresolution analysis of signals and images consists in restoring continuous hidden random variables x = {x s } s∈S from observed ones y = {y s } s∈S. This is done classically in the context of Hidden Markov Trees (HMT). HMT have been generalized recently to Pairwise Markov Trees (PMT). In this paper we propose a smoothing restoration algorithm for Gaussian PMT.