Multivariate image segmentation using Laplacian Eigenmaps

We are exploring the novel technique of Laplacian Eigenmaps (LE) [1] as a means of improving the clustering-based segmentation of multivariate images. A computationally efficient scheme, taking advantage of the ability of LE-algorithm to learn the actual manifold of the multivariate data, is introduced. After embedding the local image characteristics in a high-dimensional feature space, the skeleton of the intrinsically low dimensional manifold is reconstructed. A low-dimensional map, in which the variations in the local image characteristics are presented in the context of global image variation, is then computed. The non-linear projections on this map serve as inputs to the fuzzy c-means algorithm boosting its clustering performance significantly. The final segmentation is produced by a simple labelling scheme that works pixelwise. The experimental results using RGB-images were very promising and showed that robustness to noise and generic character are the main advantages of our method.