Based On Bayesian Decision Fusion Of Multiple Biometric Identification System

Biometrics is an attempt to imitate the elegant sensor fusion network of humans in identifying and verifying other humans by their behavioral and physiological characteristics. Compared to something you possess or something you know, these characters are inherently secure for they are unique in you and hard to be forged. The modern technology of Biometrics consists of face recognition, fingerprint recognition, speaker recognition and so on. However, each modality has its own issues and limitations. Therefore, it is an inevitable trend to improve the security level of a system by fusing several biometrics modalities together, which has far-reaching research significance and vast application prospect.This thesis realizes the algorithms of face recognition, fingerprint recognition respectively and proposes a novel adaptive algorithm to integrate multiple biometrics into one system in terms of accuracy, reliability and universality by adopting related Pattern Recognition methodology. What's more, a multiple biometrics recognition system is developed accordingly. The main work is as follows.Firstly, a face recognition algorithm based on feature subspace is realized by combing PCA (Principle Component Analysis) and FLDA (Fisher Linear Discriminant Analysis). PCA is used to extract the dimensions with the largest variance, while FLDA further compresses the features, which minimize the within-class scatter and maximize the between-class scatter at the same time. As a result, better classification ability is obtained.Secondly, a fingerprint recognition algorithm based on Gabor filterbank is improved and realized by combing reference point localization, feature region tesselation and Gabor filtering model. Translation invariance is implemented by locating the reference point. Rotation invariance is implemented by tesselating the feature region and extracting features of both original and rotated pictures. As a result, better matching ability is obtained.Thirdly, an adaptive multimodal recognition algorithm at the decision level is improved and realized by combing VL-BPSO (Binay Particle Swarm Optimization with Velocity Limitation) and the minimal risk Bayes decision model. VL-BPSO is used to seach for the best fusion rule, while Bayes decision model supplys the constructed risk function as PSO's objective, then multiple biometrics can be fused based on such fusion rule and results of all single modalities subsequently. As a result, better fusion ability is obtained.The experimental results on ORL and UMIST face database and FVC2004 fingerprint database show that the algorithms perform well on both accuracy and efficiency. Moreover, the fusion algorithm of multiple biometrics is able to choose the optimal fusion rule dymatically so that a better decision-making is achieved according to the very rule and results of each modality, which further optimizes the overall performance of the recognition system.