The novel technique of Laplacian eigenmaps (LE) is studied as a means of improving the clustering-based segmentation of color images. Taking advantage of the ability of the LE algorithm to learn the actual manifold of the multivariate data, a computationally efficient scheme is introduced. After embedding the local image characteristics, extracted from overlapping regions, in a high-dimensional feature space, the skeleton of the intrinsically low-dimensional manifold is constructed using spectral graph theory. Using the LE-based dimensionality reduction technique, a low-dimensional map is computed in which the variations of the local image characteristics are presented in the context of global image variation. The nonlinear projections on this map serve as inputs to the Fuzzy C-Means (FCM) algorithm, boosting its clustering performance significantly. The final segmentation is produced by a simple labeling scheme. The application of the presented approach to color images is very encouraging and illustrates the effectiveness of the performance over alternative methods.