Edge-adaptive clustering for unsupervised image segmentation

When used for image segmentation, most standard clustering algorithms can shift image boundaries due to intensity fluctuations within an image. In this paper, a novel approach to clustering is proposed for performing unsupervised image segmentation based upon a generalization of the standard K-means clustering algorithm. By incorporating a new term into the objective function of the K-means algorithm, boundaries between regions in the resulting segmentation are forced to occur at the same locations as edges in the observed image. A straightforward iterative algorithm is derived for minimizing this edge-adaptive K-means objective function. The result is an efficient segmentation algorithm that reconstructs boundaries in the image more accurately than standard methods.