Spectral clustering algorithms have wide applications in pattern recognition and image segmentation.They can cluster samples in any form of the feature space and have global optimal solutions.In this paper

a new graph-based spectral cluster algorithm called Dcut is applied to image segmentation.Dcut completely satisfies the general criterion of the cluster algorithms:maximizing the within-cluster similarities while minimizing between-cluster associations.Compared with Ncut

Dcut has better grouping performance in image segmentation.In order to overcome Dcut's shortcoming i.e.slow speed for image segmentation

two fast Dcut algorithms

i.e.subspace-based Dcut (SDcut) and block-based SDcut (BSDcut)

are proposed.SDcut and BSDcut have Dcut's grouping performance whihe at the same time reducing the computational complexity.Experiments based on texture images and real images demonstrate the advantages of the proposed algorithms.