Face recognition using regularised generalised discriminant locality preserving projections

Discriminant locality preserving projection (DLPP) is a recently proposed algorithm, which is an extension of locality preserving projections (LPP) and can encode both the geometrical and discriminant structure of the data manifold. However, DLPP suffers from small sample size (SSS) problem which is often encountered in face recognition tasks. To deal with this problem, the authors propose a novel regularised generalised discriminant locality preserving projections (RGDLPP) method for facial feature extraction and recognition. First, locality preserving within-class scatter in DLPP method is replaced by locality preserving total scatter and all the training samples are projected into the range of locality preserving total scatter. Then the authors regularise the small and zero eigenvalues of locality preserving within-class scatter since the small eigenvalues are sensitive to noise. RGDLPP address SSS problem by removing the null space of locality preserving total scatter without loss of discriminant information. Meanwhile, RGDLPP can alleviate the problem of noise disturbance of the small eigenvalues. Experiments on the ORL, Yale, FERET and PIE face databases show the effectiveness of the proposed RGDLPP.