Improved complete neighbourhood preserving embedding for face recognition

Complete neighbourhood preserving embedding (CNPE) is a recently proposed approach to overcome the drawbacks of neighbourhood preserving embedding (NPE) which is difficult to directly apply to face recognition because of computational complexity. However, there are still disadvantages for CNPE: (i) CNPE is time-consuming when N is large, here N is the sample size; (ii) the solutions of CNPE may suffer from the degenerate eigenvalue problem, that is, several eigenvectors with the same maximal eigenvalue, which make them not optimal in terms of the discriminant ability. In this study, the authors proposed a new approach, namely improved complete neighbourhood preserving (ICNPE), to address the drawbacks of CNPE. ICNPE is more efficient than CNPE and can overcome the degenerate eigenvalue problem of CNPE. Experiments on the Olivetti & Oracle Research Laboratory (ORL), Yale, PIE (pose, illumination and expression) and Alex Martinez and Robert Benavente (AR) face databases show the effectiveness of the proposed ICNPE.