Face Recognition Based on a New Nonlinear Feature Fusion Method

摘要

Compared with the single facial feature, the fusion feature is better in the comprehensive representation of the visual content of face image and in improving the accuracy of face recognition. However, how to fuse these different face features is still an open question. A new multiple feature fusion method is proposed based on the enlightenment of Multi-kernel learning theory. Firstly, these different features are mapped to their corresponding kernel feature spaces. Then, the fusion feature, obtained by the linear combination of all the features coming from different kernel feature spaces, is applied to face recognition. The new method cannot only effectively capture the nonlinear discriminant features of the faces, but also avoid the curse of dimensionality which often happens in the linear combination method. Further more, based on the maximum-entropy principle, a non-Gaussian measurement is presented for optimizing the best combination coefficients and the kernel parameters. The experiments on ORL and Extended YaleB face databases demonstrate the recognition accuracy of the proposed method is superior to that of those methods which are based on the single feature or the linear combination features.