28 Regularized Enighbourhood Preserving Embedding in Face Recognition

Neighbourhood Preserving Embedding (NPE) is a manifold based dimensionality reduction technique in face recognition, which is able to discover the underlying informative facial structures. In this paper, we propose an enhanced version of NPE, namely Regularized NPE (RNPE) by regulating the NPE space for gaining better recognition performance. In RNPE, the eigenspace of the within-class scatter matrix is decomposed into three subspaces and the eigenfeatures of the subspaces will be regularized according to an eigenspectrum weighting model, respectively. Then, NPE is performed on the regularized features for dimensionality reduction. The proposed method obtains better recognition rate and is experimentally verified in PIE and FRGC databases.