Local Gabor Fisher Classifier for Face Recognition

Abstract

This paper proposes a novel Local Gabor Fisher Classifier (LGFC) for face recognition. Gabor feature vector has been recognized as one of the most successful face representations, however, its dimension is too high for fast extraction and accurate classification. In LGFC, Local Feature Analysis (LFA) is exploited to select the most informative Gabor features (hereinafter as local Gabor features) optimally. The selected low-dimensional local Gabor features are then classified by Fisher Linear Discriminant (FLD) for final face identification. We demonstrate that Gabor representation is much more robust than gray-level intensity to image variation caused by the imprecision of facial feature localization. Comparative studies of different similarity measures to local fisher classifier and local Gabor fisher classifier are also performed. The experiments on two traditional face databases, ORL and Aberdeen, have shown that compared with other face recognition schemes, the proposed method can effectively reduce the dimensionality of Gabor features and greatly increase the recognition accuracy.