Improvement Of Manifold Learning For Face Recognition

Manifold learning is an effective dimensional reduction technique for face feature extraction, which, generally speaking, tends to preserve the local neighborhood structures of given samples. The paper mainly studies three feature extraction technologies: discriminant analysis, manifold learning and sparse representation.Since neighbors of a sample often comprise more inter-class data than intra-class data, which is an undesirable effect for classification. In this paper, we address this problem by proposing a subclass-center based manifold preserving projection (SMPP) approach, which aims at preserving the local neighborhood structure of subclass-centers instead of given samples. We theoretically show from a probability perspective that, neighbors of a subclass-center would comprise of more intra-class data than inter-class data, and thus is more desirable for classification. In order to take advantage of the class separability, we further propose the discriminant SMPP (DSMPP) approach, which incorporates the subclass discriminant analysis (SDA) technique to SMPP. In contrast to related discriminant manifold learning methods, DSMPP is formulated as a dual-objective optimization problem and we use the linear weighted sum method to solute it.Motivated by some related sparse representation algorithm, we present in this paper a novel feature extraction approach named sparsity embedding with manifold information (SEMI), which not only preserves the sparse reconstructive relations, but also maintains the manifold structure of the reconstructed data. Specifically, for a sparse reconstructed sample, we minimize both its difference to the corresponding original sample, and its distance to the original intra-class samples. Provided that this sample lies on different submanifolds from other samples, we additionally maximize, in the objective function, its distance to the original inter-class samples.Experimental results on the public AR, FERET and CAS-PEAL face databases demonstrate that the proposed approaches are more effective than related manifold learning, discriminant manifold learning and sparse representation methods in classification performance.