1 Introduction
Global appearance based recognition is a popular scheme in recognizing facial identities [23]. With certain subspace representations, such as Eigenface [19], Fisher-face [2], a global face appearance can be effectively “encoded” into a very low dimensional vector for recognition, and thus brings appealing efficiency in inquiring face images. Extensive previous study has shown that applying subspace methods to global appearances can achieve high recognition efficiency with satisfying recognition accuracy, if faces do not contain significant local outlier appearances. Global approaches require accurate face alignment, which can be obtained by eye localization. Automatic eye localization has received extensive study [22], [14], [20], [1]; however, it still is a challenging problem to locate eyes under varied imaging conditions, such as severe reflection of eye-glasses and sun-glasses occlusions. Furthermore, severe local outlier appearances (such as the two above) can also significantly degrade the effectiveness of subspace methods in encoding the global appearances for face recognition [8].