Research On Image Segmentation Based On Mixture Model

In recent years,with the advent of the information age and the digital age,image segmentation technology has been widely used in fields such as medicine and military engineering.Image segmentation technology can help to analyze various quantitative and qualitative image data which are obtained,so that these data can be better applied to engineering applications.Image segmentation is a key step to analyze and understand the image.It is the most important and basic technique in digital image processing.It is a kind of basic computer vision technology,which can accurately segment the image and have an important value for both engineering applications and quantitative analysis.For example,new medical imaging technologies such as computed tomography(CT),magnetic resonance imaging(MRI)and ultrasonography have been widely used in medical diagnosis,preoperative planning,treatment,postoperative monitoring and other aspects.At present,many methods such as edge detection,region-based segmentation,threshold-based segmentation have been proposed.Among them,the segmentation algorithm based on statistical models is a more flexible and effective segmentation method in the field of image segmentation.Among them,the Gaussian mixture model is the most common.However,the clustering performance of the finite Gaussian mixture model is greatly affected by the non-typical samples or outliers and could not be suitable for the segmentation of high-noise images such as medical image segmentation.Dirichlet mixture model(DMM)can be applied into clustering of both symmetrical and asymmetric modal data better compared with GMM,so it much likely to classify data samples accurately and have better performance on anti-noise than GMM and finite Student's mixture model(SMM).However,if we only consider this model,the ability of anti-noise only obtained from the statistical characteristics of the data itself since we does not really consider the secondary influence of external interference from the original data,which can not obtain a good segmentation result.Therefore,in order to solve this problem,we should consider Dirichlet Mixed Modelbased on the spatial constraints(SC-DMM).In addition,for some single-channel images,if the original data is used for clustering directly,the Dirichlet distribution data format can not be obtained,and therefore,the original data needs to be normalized.In this case,this dissertation will consider another model called Inverted Dirichlet Mixture Model(IDMM-SC),which can allow single-channel data to be directly categorized without the need for normalization and retain the full range of statistical characteristics.Finally,considering the complexity and multi-statistic properties of data,we propose a more generalized mixture model called Beta-Liuville mixture model based on spatial constraints(BLMM-SC).In addition,some of the previous models only consider adding space constraints to the posterior,resulting in the result that the model is not robust enough against noise.Therefore,in this model,we use the generalized mean(GM)to impose space constraints on both prior and posterior.In the related experiment,the mixture model is robust against noise and good segmentation results are obtained.