Research On Multimodal Biometric Algorithms Based On SVM Fusion

Biometrics is an emerging technology, which has unique advantages over traditional authentication technologies and has wide applications such as in financial, public security areas and also in everyday life. This paper focuses on both face and voice biometric algorithms and their support vector machine (SVM) fusion algorithm under verification mode and the main content is as follows:First, due to the problem that face recognition can not be guaranteed to be real-time in practical applications with a huge number of clients, face verification algorithm based on eigenfaces is studied. Above all, individual eigenface subspaces are constructed for each subject, under which there are more samples in different illumination, facial expressions and pose situations. Then, four classifiers namely, Euclidean distance, Mahalanobis distance, normalized correlation and reconstruction error are introduced to classify experimental results. In the end, fusion is conducted on the score matching level with generalized discriminant analysis (GDA). It is observed that eigenface algorithm in verification mode is robust, and that reconstruction error classifier functions best among the four classifiers at hand while fusion improves the overall performance again, which show certain advantages compared to relevant algorithms.Second, text-independent speaker verification is based on Gaussian mixture model, which employs linear combination of dozens of single Gaussian probability density function to describe the distribution of speaker's voice feature and overcome short-session test problem. In the experiment, only durations of 0.20s and 0.52s speech sessions are used for test and small equal error rates are obtained.Third, an alternative to construct support vector machine (SVM) kernels from orthogonal polynomials is presented. In addition, by introducing fuzzification of the penalty and different cost algorithm, it can alleviate over-fitting problem and correct the class-boundary-skew problem arising from the imbalanced training data set respectively. The elegant and fascinating characteristics of the orthogonal polynomials promise the minimal data redundancy in feature space and make it possible to represent the data with less support vectors. Experimental results show that the SVMs with orthogonal polynomial kernels outperform that with traditional kernels in terms of generalization power and less support vectors.In the end, a hybrid personal identity authentication strategy with flexible security levels is proposed after combining traditional personal identity technology and biometric technology.