Research On Biometrics Recognition Based On Manifold Learning And Kernel Learning

Manifold learning is a nonlinear dimensionality reduction method developed in recent years which aims to preserve the neighborhood relations and reduce the dimensions of data. In this paper, the traditional manifold learning methods were promoted that solved the problem of parameter selection, insufficient discriminative ability and sensitivity to outliers by mapping the data to kernel space, directly project the 2D image data and increase the L1-norm constraint to improve the discriminative ability and the robust to noises. Then the improved algorithms were used for the recognition of faces, palm prints and Movement-related Coritical Potentials (MRCPs). The main contribution of this paper is as follows:(1) A method named Supervised Kernel Optimized LPP (SKOLPP) is proposed in this paper. Kernel Locality Preserving Projection algorithm (KLPP) can effectively preserve the neighborhood structure of the database using the kernel method. We have known that supervised KLPP (SKLPP) can preserve within-class geometric structures by using label information. However the conventional SKLPP algorithm endures the kernel selection which has significant impact on the performances of SKLPP. SKOLPP maximized the class separability in kernel learning using a data-dependent kernel and Fisher criterion in order to solve this problem. Experiments results on ORI?AR and Palmprint databases showed the effectiveness of the proposed method.(2) A novel method named two-dimensional discriminant neighborhood preserving embedding (2DDNPE) is proposed.2DDNPE benefits from two important strategies, namely, the fisher-like criterion and matrix based projection First, a novel fisher-like criterion is proposed which can conduct dicriminant analysis across the feature spaces constructed by Laplacian Eigenmap and LLE, and improve the discriminant ability by minimizing the within-class distance, while maximizing the between-class distance. Second, the matrix based projection is adopt to directly extract optimal projections from 2D image matrices rather than 1D vectors, and thus enabling 2DDNPE to avoid the small sample size problem. The experiments conducted on UMIST, Yale and AR face databases showed that 2DDNPE can effectively extract face features and classification.(3) An LPP-L1 classifier for detecting MRCPs is proposed in this paper. It is demonstrated that LPP-L1 based L1-norm distance is more robust than LPP based on L2-norm with the presence of outliers. Consequently, LPP-L1 reduced the false positive rate significantly for detecting MRCPs while maintaining the true positive rate substantially unchanged. The method can be a promising candidate for developing BCI neurorehabilitation systems based on MRCP detection.