SKFD-Isomap for face recognition

Abstract

Recently, neuroscientists emphasized the manifold ways of perception and proposed Isomap for manifold learning. Favorable results have been achieved using Isomap for data description and visualization. However, since the unsupervised Isomap is developed based on minimizing the reconstruction error with multidimensional scaling (MDS) without using the class specific information, it may not be optimal from the perspective of pattern classification. Therefore, an improved version of Isomap, namely SKFD-Isomap, is proposed using class information to construct the neighborhood, and kernel Fisher discriminant (KFD) to achieve the nonlinear embedding. A nearest neighbor classifier is then applied in the subspace for classification. Experimental results show the effectiveness of the proposed approach. © 2005 by International Federation for Information Processing.