Abstract: To overcome the disadvantage that the penalty graph constructed by marginal Fisher analysis (MFA) can't sufficiently describe interclass separabihty, this paper proposed a novel feature extraction method, called local and global margin embedding (LGME). In LGME, all interclass data pairs are used to construct penalty graph, whereas the importance of limited interclass data pairs with minimal margins is emphasized properly. Compared with MFA, I_GME simultaneity uses local and global interclass margin to characterize interclass separability, so the data features extracted by LGME have more discriminative power. hhe experimental results show that the face image features extracted by LGME for face recognition have higher recognition rate and more robust.

Key words: Face recognition, Feature extraction, Marginal Fisher analysis (MFA)，Local and global margin embedding(LGME)