Square Loss based regularized LDA for face recognition using image sets

In this paper, we focus on face recognition over image sets, where each set is represented by a linear subspace. Linear Discriminant Analysis (LDA) is adopted for discriminative learning. After investigating the relation between regularization on Fisher Criterion and Maximum Margin Criterion, we present a unified framework for regularized LDA. With the framework, the ratio-form maximization of regularized Fisher LDA can be reduced to the difference-form optimization with an additional constraint. By incorporating the empirical loss as the regularization term, we introduce a generalized Square Loss based Regularized LDA (SLR-LDA) with suggestion on parameter setting. Our approach achieves superior performance to the state-of-the-art methods on face recognition. Its effectiveness is also evidently verified in general object and object category recognition experiments.