Estimating the Number of Components of Mixture Models for Medical Image
Chang Jin-yi · 2010
Estimating the number of mixture models is the key part of clustering analysis and density estimation for medical image.In order to overcome the over-fitting problem of the method of information criteria,we proposed a new estimation method which is based on a feature function of Gaussian mixture models(GMMs).First,the feature function of medial image was defined on the GMMs.Second,constructed a new criterion with the feature function to estimate the number of components of the mixture models.At last,we proposed an algorithm to compute the new criterion.Our new criterion uses a parameter to adjust the value of log-feature function and to keep the balance effect of the penalized function.Experiments on the simulate data and real CT image show our criterion can determine a more reasonable number of components K than others information model selection criteria and avoid the over-fitting problem of the medical image.