Local sparsity preserving projection and its application to biometric recognition

Dimensionality reduction techniques based on sparse representation have drawn great attentions recently and they are successfully applied to biometric recognition. In this paper, a new unsupervised dimensionality reduction method called Local Sparsity Preserving Projection (LSPP) is proposed. Unlike the traditional dimensionality reduction methods based on sparse representation which only preserve the sparse reconstructive relationship, LSPP preserves sparsity and locality characteristics of the data simultaneously. In LSPP, a training sample could be more possibly represented by training samples from the same class and a more accurate sparse reconstructive weight matrix is obtained. Thus, LSPP has more powerful discriminative ability than traditional dimensionality reduction methods. As kernel extension of LSPP, Kernel Local Sparsity Preserving Projection (KLSPP) which is more effective for nonlinear data is also presented. In the experiments on several biometric databases, the proposed methods obtain higher recognition rates and verification rates and are computationally more efficient than the traditional dimensionality reduction methods based on sparse representation.