Optimal Gabor Features for Face Recognition

2D Gabor features have been recognized as one of the most successful face representations. But they often result in very high dimensional feature vectors, which rend them impractical for real applications in resource limitation environment (such as mobile phone, PDA etc). This paper proposes a two-level supervised feature selection algorithm for Gabor feature-based face recognition. In the first stage, a non-parametric measure of discrimination performance is used as criterion to reduce Gabor features and formed a sub-optimal subset. In the second stage, the most informative Gabor features are determined using Adaboost feature selector. These most discriminative Gabor features are then subjected to the linear discriminant analysis (LDA) process for further class separability enhancement. Experimental results on a CAS-PEAL large-scale Chinese face databases show that the proposed method achieves high recognition accuracy, whilst the dimensionality and computation cost of Gabor features have been effectively reduced. It finished face authentication with high detection rates in 0.8 s on a CASIO-W21CA mobile phone with ARM926EJ-S processor that lacks floating-point hardware