An exact inference approach of posterior probability is proposed for multiscale image segmentation to overcome limitation of local optimum as well as unbounded convergence rate of traditional wavelet-domain hidden Markov tree (HMT) segmentation methods. Hidden Markov forest (HMF) model is constructed by characterizing inter-scale statistical dependence between fine scale pixel and high scale wavelet coefficients. Bottom-up likelihood estimation and up-bottom posterior inference on each sub-tree of the HMF model are performed, in which consistency of unary and pairwise distribution is guaranteed. Binary segmentation at multiscales are achieved by threshold the posterior probability. Experimental results of typical dynamic background segmentation as well as comparison with segmentation by weighted aggregation (SWA) algorithm demonstrate the effectiveness of our method.