Class-dependent LDA for feature extraction and recognition

Motivated by the fact that a class has its own optimal feature vector discriminating itself from the other classes, we propose a new feature extraction algorithm for face recognition. In the scheme of the proposed class-dependent LDA (CDLDA), the evaluation criterion discriminating the samples of one class from those of the other classes is proposed. Different linear transformation spaces are constructed for different classes. It is contrary to the class-common methods that have only one identical space for all classes. A classification method based on the nearest neighbor classifier is used to identify new samples. The proposed method is compared with three algorithms on 4 widely used face databases. The experimental results show that CDLDA outperforms the other methods.