Abstract
Nonnegative matrix factorization (NMF) tends to characterize local feature variation and has been shown to be better interpretable on facial images. In this thesis, to solve the facial age estimation problem, we propose to extract age-related features by using a supervised NMF. In addition, since different people usually have very different aging tendency, it is by no means an easy task to find a set of good age-related features feasible for all individuals. To overcome this difficulty, we further include a person-independent constraint and propose a new approach called person-independent supervised NMF (PISNMF) to characterize the aging properties. In addition, we also extend our proposed PISNMF to handle the age-invariant face recognition problem. We conduct PISNMF on FG-NET database and successfully extract the aging-related features. Our experiments show that the derived facial bases indeed characterize age-related local variations and the results of both age estimation and age-invariant face recognition outperform most existing methods.