Abstract: An intermediate-parameter relaxation scheme is described in this paper. For classification problems whose labels are based on the combination of intermediate parameters, this scheme has the advantages that the setting of compatibilities and conditional probabilities can be simplified- and the computational complexity be reduced greatly. With this scheme, satisfactory results have been obtained in improving the topographic primal sketch. The estimation of initial probabilities with histograms of subimages is also described. It is proved that this kind of estimation can speed up the convergence of iteration.