A multilevel spectral hypergraph partitioning approach for color image segmentation

In many image processing applications, and in the human visual system, relationships among objects of interest are more complex than pairwise. Simply approximating complex relationships as pairwise ones can lead to loss of information. A natural way to describe complex relationships, without loss of information, is to use hypergraphs. In this paper, we use a Color Image Neighborhood Hypergraph representation (CINH), which extracts all features and their consistencies in the image data and whose mode of use is close to the perceptual grouping. We formulate a color image segmentation problem as a CINH partitioning problem. A new multilevel spectral hypergraph partitioning approach is presented. Our experiments on the Berkeley images database showed encouraging results compared with the graph partitioning strategy based on Normalized Cut (NCut) criteria.