On the suitability of different probability distributions for the task of image segmentation

The task of image segmentation can be approached by learning a probabilistic mixture model of the observed image data. If this is performed in an unsupervised fashion the problem may be divided into two tasks. Namely, learning the latent structure, and modelling the appearance of each latent component. In this work the task of finding the latent structure is set aside in favour of an in-depth analysis of the influence of the choice of the model distribution and the choice of the feature space. Hence, the latent structure is assumed to be known or predicted by an oracle. Further, an overview of possible distributions is given which may be used to model the data instead of the ubiquitous normal distribution. Lastly, a new class of multivariate distributions utilising flexible marginal distributions and a Gaussian copula to model the correlation structure of the marginals is presented and compared to other distributions in terms of their capability to model challenging natural images.