Image Clustering And Retrieval Based On MRF Models

Image clustering and retrieval play important roles in the field ofcomputer vision especially in senior semantic understanding, and the problemsof construction and inference have been viewed as two key research directions.MRF models are statistical models which can describe the interactioninformation between elements. Under those models, the internal structure ofimages can be formed simply and reasonably in the way of probability.Compared with traditional discriminated models, the fuzziness and ambiguitycan be reduced in image understanding. The inference tasks based on MRFmodels can be equivalent to the problems of minimization energy. ICM andgraph-cut are two kinds of effective discrete optimization methods which areuseful to solve the problems of image clustering and retrieval.Focusing on image clustering and retrieval problems, the major works ofthis paper are as follows:(1) This paper summarizes the research of image processing under theMRF models, and then introduces the significance values of using MRFmodels for image clustering and retrieval.(2) We outline the basic theory of MRF, including definitions, energyfunctions, and inference algorithm, meanwhile we mainly focus on the discreteoptimization algorithms of ICM and graph-cut.(3) An image clustering method based on MRF model has been proposed,by the way of translating the problem of clustering to energy minimization.The unary and dualistic energies of are estimated by Gaussian mixture modeland Potts model respectively. We also present a new neighborhood constructmethod based on K-nearest neighborhood systems and get the calculationsunder the framework of Bayesian by using ICM algorithm.(4) A new region-based retrieval method is proposed under graph-cutmethod, and the whole framework has been presented. We mainly focus onanalyzing the effectiveness of graph-cut algorithm which plays an importantrole in image retrieval.