Color Image Segmentation in Both Feature andImage Spaces

One critical problem in image segmentation is how to explore the information in both feature and image space and incorporate them together. One approach in this direction is reported in this chapter. Watershed algorithm is traditionally applied on image domain but it fails to capture the global color distribution information. A new technique is to apply first the watershed algorithm in feature space to extract clusters with irregular shapes, and then to use feature space analysis in image space to get the final result by minimizing a global energy function based on Markov random field theory. Two efficient energy minimization algorithms: Graph cuts and highest confidence first (HCF) are explored under this framework. Experiments with real color images show that the proposed two-step segmentation framework is efficient and has been successful in various applications. INTRODUCTION Color image segmentation can be modeled as a labeling problem for each pixel in the entire image. In the optimal assignment, pixels having the same label should have the following two properties: one, that they must have small distance in features (such as color, texture, etc.) space, the other is that they must be spatially coherent. One natural way of combining the above properties together is treating spatial position as two IDEA GROUP PUBLISHING This paper appears in the publication, Advances in Image and Video Segmentation edited by Yu-Jin Zhang © 2006, Idea Group Inc. 701 E. Chocolate Avenue, Suite 200, Hershey PA 17033-1240, USA Tel: 717/533-8845; Fax 717/533-8661; URL-http://www.idea-group.com ITB13101