Image segmentation using finite mixtures and spatial information

摘要

The finite mixture is a flexible and powerful probabilistic modelling tool. It can be used to provide a model-based clustering in the field of pattern recognition. However, the application of finite mixtures to image segmentation faces some difficulties. First, the estimation of the number of components is still an open question. Second, mixture-based data clustering does not consider spatial information, which is important for smooth regions to be obtained in the segmentation results. In this paper, the spatial information is used as a prior knowledge of the number of components. The spatial information does not tell the value of the number of components; instead, it provides some indirect information about this value. An Expectation Maximization based algorithm is developed to estimate mixture density using the indirect information. The experimental results with simulated data of 2D Gaussian mixture show that the proposed algorithm is capable of estimating the number of components accurately without using any model selection criteria. The experimental results of image segmentation show that the proposed algorithm has better performance in generating smooth regions in the segmentation results compared to the common algorithms that use model selection criteria to estimate the number of components.