Research On Finger Vein Recognition Based On Manifold Learning And Extension Classifier

As a new kind of identity authentication technology, finger vein identification technology has achieved more advantage of security and high-precision. Therefore, it has very broad application prospects. Authentication is through the body characteristics of the human finger vein. In this paper, the finger vein image collection system, feature extraction of finger vein and classifier are researched.Firstly, An acquisition device of finger vein image was designed and then produced. According to the finger vein images of many literatures, the disadvantages of these finger vein acquisition devices that were designed in china are analyzed. In order to get consistency quality finger vein images, a new acquisition device of finger vein image was designed. According to the finger thickness, auto-dimming circuit is designed which can automatically adjust the intensity of the crossing infrared light and keep the image relatively stable. And the device'interface is Windows USB2.0. The device is designed based on USB2.0 and CMOS. Experimental results show that the finger vein image collected by this system can be clear and stable quality. The hardware is made of CMOS image sensor module, auto-dimming module, USB controller module, other supporting module and so on. The software is made of USB firmware program, USB driver and image acquisition program.Secondly, finger vein feature extraction methods based on linear manifold learning are studied. The typical feature extraction methods of linear manifold learning are analyzed in detailed. The typical feature extraction methods are principal component analysis (PCA), linear discriminant analysis (LDA), bi-directional two dimensional PCA (B2DPCA) and bi-directional two dimensional LDA (B2DLDA). On the basis, the shortages of the traditional weighted method with one directional are analyzed, and two improved algorithms are proposed. The two algorithms are bi-directional weighted B2DLDA (BWB2DLDA or (W2D)2LDA) and bi-directional weighted B2DPCA (BWB2DPCA or (W2D)2PCA) with eigenvalue normalization. The characteristics of modular PCA are analyzed. And combining the advantages of the image modular and the bi-directional weighted, a new improved algorithm is proposed, which is bi-directional weighted modular B2DPCA (BWMB2DPCA) with eigenvalue normalization. The comparing experimental results of finger vein show that the proposed algorithms are better than the traditional algorithms.Thirdly, finger vein feature extraction methods based on nonlinear manifold learning are studied. The typical feature extraction methods of nonlinear manifold learning are analyzed in detailed. The typical feature extraction methods are locally linear embedding (LLE), isometric mapping (ISOMAP), Laplacian Eigenmap (LE). However, these algorithms can only formed mapping in the training samples, and the mapping in a new test sample can not be formed. So locality preserving projections (LPP), supervised locality preserving projection (SLPP) and bi-directional two dimensional LPP (B2DLPP) are studied. On the basis, a new improved algorithm is proposed. The algorithm is bi-directional weighted modular B2DLPP (BWMB2DLPP). The comparing experimental results of finger vein show the validity of the algorithm.Finally, the classifiers based on extension are studied. The extension distance and the correlation function of extension are deeply studied. In order to solve the problem that the traditional extension classifier is only suitable for little pattern, three classifiers based on the extension distance and the extension correlation function are proposed. These classifiers are average correlation function, K-maximum correlation function and maximum correlation function. Including finger vein images, three kinds of experimental data are used to compare and analyze for the extension classifiers. The test results show that the classification results of the extension classifiers can reach the level of k-nearest neighbor or nearest neighbor.