Evaluation and Selection of Discriminating Gabor Features for Face Recognition

The appearance of a human face can be represented as a grid of Gabor jets. However, the Gabor face representation is a very high dimensional graph structure data set mixed with redundant and randomized noisy information. This paper proposes two metrics for evaluating the discriminating capabilities of different Gabor features, which are based on the discriminating margin and mutual information between correct classification and features. The discrimination of different positions, different scales and different orientations of Gabor features can be evaluated and ranked based on the corresponding metric magnitudes. Some experiment results with different level discriminating features are presented. A case study shows that the highest recognition rate is achieved after eliminating about 20 percentage lower bottom discriminating features; and taking only about 50 percentage features from top rank is enough to get a fairly good recognition rate.