SAR image classification with normalized gamma process mixtures

We propose a novel image prior for the non-parametric Bayesian mixture model based unsupervised classification of SAR images. We modified the Normalized Gamma Process prior that constitutes a more general form of the Dirichlet Process prior in order to enclose the contribution of the adjacent pixels into the classification scheme. This yields an image classification prior embedded in a mixture model that allows infinite number of clusters and enables reaching to smoothed classification maps. Based on the classification results obtained on synthetic and real TerraSAR-X images, it is shown that the proposed model is capable of accurately classifying the pixels. It applies a simple iterative update scheme at a single run without performing a hierarchical clustering strategy as used in the previously proposed methods. It is also demonstrated that the model order estimation accuracy of the proposed method outperforms the conventional finite mixture models.