Statistical dimension reduction methods for appearance-based face recognition

Two novel moment-based methods which are insensitive to large variation in lighting direction and facial expression are developed for appearance-based face recognition using dimension reduction methods in statistics. The two methods are based on Sliced Inverse Regression (SIR) (Li, 1991) and Sliced Average Variance Estimate (SAVE) (Cook and Weisberg, 1991) and termed as the Sirface method and the Saveface method, respectively. The Sirface method estimates the mean dierence subspace while the Saveface method estimates the mean and covariance dierence subspace. They produce well-separated classes in a low-dimensional subspace, even under severe variation in lighting and facial expression. In the subspace sense, the Sirface is equivalent to the Fisherface (Belhumeur et al., 1997) and the Saveface is even more comprehensive. Since both methods produce the optimal (smallest) image subspaces, they can lower both the error rate and the computational expense.