As the feature dimension increases，the original PMK suffers from distortion factors that increase linearly with the feature dimension.This paper proposes a new method by consistently dividing the feature space into two subspaces while generating several levels.In each subspace of the level，the original pyramid matching is used.Then a weighted sum of every subspace at each level is made.To optimize the added kernel matrix，we get a p.s.d.kernel which can be used in kernel based learning methods (such as SVM).Experiments on dataset Caltech-101 and ETH-80 show that：compared with other related algorithms which need hundreds of times of original computational time，It takes only about 46 times of original computational time to obtain the same accuracy by using the method of DP-PMK.