Natural image segmentation with non-extensive mixture models

Dusan Stosic, Darko Stosic, Teresa Bernarda Ludermir, Tsang Ing Ren

Abstract: Finite mixture models have been widely used for image segmentation in many computer vision and pattern recognition problems. While images of natural scenes are difficult to model, we can employ emerging concepts from statistical physics to achieve better representations. This paper introduces a new class of finite mixture models for solving such problems. The proposed non-extensive mixture models have real-valued power-law exponents that characterize the degree of correlations. The exponents are used to capture rare or frequent occurring patterns in the image. They can describe complex features found with a hierarchy of sizes in natural images: from small objects with a few dozen pixels to large ones that occupy the entire image. We also present a method to determine the parameters based on maximum likelihood estimation. Our numerical experiments indicate more robust and accurate capabilities of non-extensive mixture models for natural image segmentation than conventional mixture models.