Research On Image Segmentation Based On Dirichlet Process Mixture Model

With the development of digitization and information,for a huge amount of picture information,Image processing is playing an increasingly important role in our life.Image segmentation is the most important and basic technical means in image processing.Its function is to divide different regions in the image,so as to get the target that you are interested in.With the continuous innovation of science and technology,image segmentation is widely used in various fields,such as face recognition,traffic monitoring,military engineering and so on.There are many existing image segmentation methods.In recent years,due to the flexibility of mixture model in image segmentation,more and more researchers have paid attention to it,especially the finite mixture model.However,due to the need to manually determine the number of components in the finite model,it is difficult to take advantage of complex data in many application scenarios.Therefore,we propose a model method based on Dirichlet process to address the defects of the finite mixed model in this paper and applied to do color natural image and medical image segmentation.This paper will introduce the existing segmentation methods and point out the existing problems,then propose two new methods to solve these problems,and verify the effectiveness of our model method with experiments.The main research work of this paper is summarized as follows:First of all,many finite mixture models can get good segmentation effect for Gaussian data,but they do not perform well when applied to non-Gaussian data.Therefore,a Beta-liouville mixture model for color images which is based on Dirichlet process is proposed in this paper.The model use the Beta-Liouville distribution as the base distribution,and its advantage is that it can still obtain good statistical characteristics for non-Gaussian data.In addition,our model is based on the Dirichlet process mixture model(also can call as the infinite Beta-liouville mixture model),compared with the finite mixture model,our Dirichlet process mixture model does not need to manually select the number of clustering,and the model will adjust itself according to the complexity of data.In the meantime,the generalized mean(GM)is added to the spatial relationship in the model,so that the robustness and noise resistance of our model are improved,thereby enhancing the segmentation effect o f the model.For parameter estimation of the model,we used the collapsed variational Bayes.Experimental results show that our method is better than some existing methods based on Dirichlet finite mixture model in natural image segmentation.Secondly,for medical images,considering the complexity and statistical characteristics of the data,we put forward another inverted Beta-Liouville mixture model which based on the Dirichlet process(also can call as infinite beta-liouville mixture model).The basic distribution adopted by the model is the inverted Beta-Liouville distribution,which can be regarded as a generalized inverted Dirichlet distribution.Moreover,it can show multiple symmetric and asymmetric patterns,that is,the distribution can be tilted to the left,right or symmetric.Thus,its generalization ability is significantly stro nger than that of the Beta-Liouville distribution,so it is greatly applicable.This model is also an infinite mixture model and can adjust its own complexity according to the data.In addition,in order to improve the robustness of the model,the generalized mean is also used to impose spatial constraints in this model.Finally,EM algorithm is prone to overfitting and falling into local minima,and variational inference is adopted in the learning process of model parameters.The two methods proposed in this paper are respectively verified by different experiments.The first method use the Berkeley color image dataset which available on the Internet,while the second approach use data sets that are open source online simulate and real brain MRI images.Through quantitative analysis of various experiments,the superiority of the two model algorithms proposed in this paper over the finite mixed model is proved.