Selecting discriminant eigenfaces by using binary feature selection

Abstract

Eigenface (PCA) and Fisherface (LDA) are two of the most commonly used subspace techniques in the area of face recognition. PCA maximizes not only intersubject variation but also the intrasubject variation without considering the class label even if they are available. LDA is prone to overfitting when the training data set is small, which wildly exists in face recognition. In this work, we present a binary feature selection (BFS) method to choose the most suitable set of eigenfaces for classification when only a small number of training samples per subject are available. In the proposed method, we make use of class label, look on two subjects as a group, and then the most suitable eigenfaces that help to identify these two subjects are picked out to form the binary classifier. The final classifier is the integration of these binary classifiers by voting. Experiments on the AR and AT&T face databases with small training data set prove that our proposed method outperforms not only traditional PCA and LDA but also some state of the art methods.