Segmentation strategies for the alpha-tree data structure

Abstract

The alpha-tree is a versatile algorithm for color image segmentation employing attribute constraints to control the partitioning of the alpha-connected image space. Attribute constraints are enforced using a maximization strategy that returns the set of the largest connected components complying with these constraints assuming no prior violations from nested sub-components. This article presents two new strategies extending the way this set is defined. These are the non-target clustering and attribute maximization strategies, that give access to segments that could not be defined previously. Collectively they enable the handling of texture-rich regions that cannot be clustered into meaningful segments, and compute the unsupervised segmentation of images by seeking for extreme attribute values.