Segmentation Graph Hierarchies

Abstract

The region's internal properties (color, texture, ...) help to identify them and their external relations (adjacency, inclusion, ...) are used to build groups of regions having a particular consistent meaning in a more abstract context. Low-level cue image segmentation in a bottom-up way, cannot and should not produce a complete final "good" segmentation. We present a hierarchical partitioning of images using a pairwise similarity function on a graph-based representation of an image. The aim of this paper is to build a minimum weight spanning tree (MST) of an image in order to find region borders quickly in a bottom-up 'stimulus-driven' way based on local differences in a specific feature. © Springer-Verlag 2004.