Benchmarking segmentation results using a Markov model and a Bayes information criterion

Abstract

Features are derived from wavelet transforms of images containing a mixture of textures. In each case, the texture mixture is segmented, based on a 10-dimensional feature vector associated with every pixel. We show that the quality of the resulting segmentations can be characterized using the Potts or Ising spatial homogeneity parameter. This measure is defined from the segmentation labels. In order to have a better measure which takes into account both the segmentation labels and the input data, we determine the likelihood of the observed data given the model, which in turn is directly related to the Bayes information criterion, BIC. Finally we discuss how BIC is used as an approximation in model assessment using a Bayes factor.