Statistical region measures for separation of figure from ground

With a deformable template and knowledge of image statistics, two-dimensional region cues can be used to separate an object from its background. While such cues are useful, sometimes they are limited to cases where pixel populations are known. By analysing its statistical properties, we show that a new region term, related to the Kullback-Leibler divergence, is independent of density type. While it requires prior knowledge of the image, we show it is tolerant to inaccuracies in this information. Hence, limitations in existing methods are discussed and an improved approach demonstrated