Statistical learning and prior image modeling

Abstract

In order to overcome the limitations of piecewise constant phenomenon and computational burden which exist in Markov Random Field (MRF) with pair wise neighborhood and traditional learning style respectively, this paper proposes a clustering learning method from natural image database, no filters included. By this method, we get the distributive law of the blocks abstracted from natural images. Furthermore, we also do the prior image modeling according to the learned law. And the real application in image restoration illustrates its effectiveness by comparison between high order MRF prior model and pair wise MRF prior model.