Manifold sparsity preserving projection for face and palmprint recognition

Abstract

Sparsity Preserving Projection (SPP) has been recently successfully applied on pattern recognition applications and is the basis for a series of follow up extensions. However, being unsupervised for dimensionality reduction, SPP does not employ the discriminative information of class labels when projecting data into a smaller subspace. This paper proposes a manifold sparsity learning method called Manifold Sparsity Preserving Projection (MSPP) for the face and palmprint recognition. Our method employs the manifold structure for better preserving the sparsity of data in the embedding space. Differing from recent localized sparsity learning methods such as Local Sparse Representation Projections (LSRP) and Local Sparse Preserving Projections (LSPP), which enforce a one-to-one matching between a sample and its sparsely reconstructed model, our method employs manifold data structure to ensure that a sample and all its classmate’s sparsely reconstructed models remain as close as possible in the new space. We show that when manifold and sparsity information are simultaneously accounted for, their discriminative power is significantly leveraged. An immediate bonus of our approach is that there is no tuning parameter whatsoever for performance variation. We analytically demonstrate the aforementioned features of our approach and then, using a series of experiments on ORL, Yale, IIT Delhi near infrared facial, and NIR set of PolyU Multispectral palmprint databases its pragmatic consequences will be pictured.