Research On Multi-modal Feature Extraction Based On Subclass Discriminant Analysis And Generalized Singular Value Decomposition

Single-mode biometric feature extraction technology has been applied to many aspects of life, but also it has reached full research and development. With the development of science and technology, requirements of authentication and public safety are getting higher. The limitations of single-mode biometric identification technology is not satisfactory for practical needs, so multi-modal feature fusion technology becomes a relatively hotspot of biometric identification at present.In this paper, we study the extraction algorithm on the feature level for multi-modal feature fusion, and propose the theory of multi-modal feature extraction based on subclass discriminant analysis (SDA) and generalized singular value decomposition (GSVD), that is multi-modal SDA-GSVD. Palmprint, face and knuckle are treated as three subclasses of a class base on the theory of subclass discriminant analysis. All samples through the projection are taken to get identification feature and we reconstructed within-class scatter matrix and between-class scatter matrix. The method minimizes difference between subclasses and maximizes the difference within each subclass of class. Generalized singular value decomposition method is used in the calculation process to solve the singularity problem, and two-modal fusion is extended to the multi-modal feature fusion.And then, we extend multi-modal integration of data through the kernel functions in the nonlinear space. We use the features of high dimensional data space through the kernel expansion, so this paper proposes the theory of multi-modal feature extraction based on kernel extension of SDA-GSVD, namely, KSDA-GSVD. The effective using of a multi-modal features in the nonlinear space makes it easier to distinguish between the subclasses.Experimental results on the fusion of AR, palmprint and FKP (Finger Knuckle Print) database show that the proposed approaches get better accuracy and validity than any other single-mode theory.