Image segmentation based on k-means clustering and energy-transfer proximity

Abstract

In image segmentation, measuring the distances is an important problem. The distance should tell whether two image points belong to a single or, respectively, to two different image segments. Although the Euclidean distance is often used, the disadvantage is that it does not take into account anything what happens between the points whose distance is measured. In this paper, we introduce a new quantity called the energy-transfer proximity that reflects the distances between the points on the image manifold and that can be used in the image-segmentation algorithms. In the paper, we focus especially on its use in the algorithm that is based on k-means clustering. The needed theory as well as some experimental results are presented. © 2011 Springer-Verlag.