Using Random Subspace to Combine Multiple Features for Face Recognition

LDA is a popular subspace based face recognition approach. However, it often suffers from the small sample size problem. When dealing with the high dimensional face data, the LDA classifier constructed from the small training set is often biased and unstable. In this paper, we use the random subspace method (RSM) to overcome the small sample size problem for LDA. Some low dimensional subspaces are randomly generated from face space. A LDA classifier is constructed from each random subspace, and the outputs of multiple LDA classifiers are combined in the final decision. Based on the random subspace LDA classifiers, a robust face recognition system is developed integrating shape, texture, and Gabor wavelet responses. The algorithm achieves 99.83% accuracy on the XM2VTS database.

• Type: Conference
• Year: 2004
• Where: FGR
• Authors: Xiaogang Wang, Xiaoou Tang
• Keywords: Biometrics, FGR 2004, LDA Classifiers, Random Subspace, Sample Size Problem