Abstract: Gaussian mixture model is one of the simple,effective and widely used tools in image segmentation.How-ever,the fitting result is not accurate enough when the number of mixture components in the traditional Gaussian mixture model is determined.In addition,because the spatial relationship between pixels is not considered,the segmentation results are easily affected by noise,and the segmentation accuracy is not high.To remedy the defects of the traditional Gaussian model,this paper proposed a multi-class Gaussianmixture model and a neighborhood information based Gaussianmixture model for image segmentation.The multi-class Gaussian mixture model decomposes the traditional mixture model.The traditional mixture model is composed of M different weighted distributions,and multi-class Gaussianmixture model decomposes each of the M components into R different distributions,that is,the multi-class Gaussian mixture model is composed of M different weighted distributions,and each of the M distributions is obtained by mixing R different distributions,thus improving the fitting accuracy of the model.The neighborhood information based Gaussianmixture model adds spatial information to the prior probability and posterior probability in the model,thus enhancing the information association and antinoise capability among pixels.The segmentation results were evaluated by the indexes of structural similarity,misclassification rate and peak signal-to-noise ratio.The experimental results show that compared with the existing segmentation method of mixture model,the segmentation accuracy of the proposed method in this paper is greatly improved,and the noise is effectively restrained.

Key words: Gaussian mixture model, Image segmentation, Multi-class, Neighborhood information