Unsupervised Bayesian classifier applied to the segmentation of retina image

In this paper, we use a stochastic model based on the finite normal mixture distribution identification for retina image segmentation. Local unsupervised methods blind and contextual, using the Expectation-Maximisation (EM) family algorithms for parameter estimation are tested. To get rid of the spatial dependence effect of pixels, a decorrelation processing is used before parameter estimation. The segmentation is then performed by Bayesian decision rule. Segmentation results are presented to prove the effectiveness of different approaches.