Kernel uncorrelated supervised Discriminant Projections with its application to face recognition

Feature extraction is an important step towards pattern recognition. Unsupervised Discriminant Projection (UDP) shows desirable performance for face recognition, but it is unsupervised and the features extracted are correlated; besides it is a linear method in nature. To solve these problems, a new feature extraction method called kernel uncorrelated supervised discriminant projection (KUSDP) is proposed. In the proposed algorithm, the data in the original space are first mapped into one high dimensional space by kernel trick, then one supervised discriminant method is performed in this high dimensional space, meanwhile an uncorrelated constraint is imposed. As a result, the proposed algorithm can handle the nonlinearity, and the locality of the intra-class can be preserved and the separability of inter-class is enlarged, also the uncorrelated vectors reduce the redundancy to its minimum, so it has more discriminative power. Experiments on face recognition demonstrate the correctness and effectiveness of the proposed algorithm.