Face Recognition Using Nonlinear Partial Least Squares in Reproducing Kernel Hilbert Space

Abstract

Dimensionality reduction is an important step for face recognition. A novel feature extraction method using nonlinear partial least squares discrimination (NPLSD) for facial images is proposed in this paper. The method is constituted by two consecutive steps, in which the original facial features are projected onto the reproducing kernel Hilbert space to solve the linear inseparable problem, then the iterative and supervised partial least squares method is used for calculating subsequent discriminant components, wherein an approach for modeling a relationship between a set of input variables and response variables, while maintaining most of the variance in the input variables, and then extract the discrimination features from different classes, until convergence of latent vectors by more iterations. This will result in a dramatic reduction for the computational time since the calculation for the covariance matrix is not required. The experiment results using the ORL data demonstrates the effectiveness of the proposed method.