Generalized Locality Preserving Projection for Multimodal Biometric Recognition

Multimodal biometric recognition is a promising personal identity authentication technology which can remedy the limitation of the traditional identity authentication and the unimodal biometrics. Comparing with other three fusion levels of multimodal biometrics, feature level fusion can reduce the redundant information to avoid calculation consumption and acquire the discriminative information to improve the system performance. Complex fusion is a novel feature fusion pattern which takes two features as real part and imaginary part of a complex vector. However, the existing linear methods of complex fusion cannot consider the nonlinear factor. Meanwhile the computations of the nonlinear methods is too great. This paper extended LPP into the complex field and proposed generalized locality preserving projection (GLPP) which takes advantage of the optimal linear approximations to find the nonlinear manifold structures.Face and palm are taken as the experimental objects to conduct the fusion features. Experimental result shows the proposed algorithm achieves much better performance than two unimodal biometrics and other four conventional multimodal biometric algorithms.