Linguistic color image segmentation using a hierarchical Bayesian approach

In this work, we combine Bayesian techniques with a color categorization model, which leads to a method for the linguistic segmentation of color images. The categorization model considers the 11 universal color categories proposed by Berlin and Kay (Basic Color Terms: Their Universality and Evolution. Berkeley: Uni- versity of California; 1969). The likelihood for each cate- gory is represented by a linear combination of quadratic splines, and as a result, each voxel in the color space L*u*v* is described as a vector of probabilities, whose components express the degree to which the voxel belongs to a given color category. This gives rise to a probabilis- tic dictionary which is used for the segmentation, in which prior spatial granularity constraints are incorpo- rated via an entropy-controlled quadratic Markov mea- sure field (ECQMMF) model, as proposed by Rivera et al. (IEEE Trans Image Process 2007;16:3047-3057). We give a generalization of ECQMMF that allows one to consider the perceptual interactions between the basic colors that were experimentally established by Boynton and Olson (Color Res Appl 1987;12:94-105). 2009 Wiley Periodicals, Inc. Col Res Appl, 34, 299 - 309, 2009; Published