Study Of Several Key Problems In Biometrics Based On Machine Learning Techniques

Biometrics is a novel identity authentication technique of which the research value and application prospects have been accepted in both academic and business fields. Through years of endeavors form academic and enterprise counterparts, great progress has been made in the research of biometrics, and a great number of biometric authentication techniques have been widely applied to many application fields since the 90th of last century. Nevertheless, the imperfection in the performance of current biometric system affects the popularization and application of biometric authentication techniques to a certain degree. How to further improve the performance of the biometric system, so that it can better meet the requirements of practical applications, is still challenging and needs in-depth researches. Besides, in recent years, machine learning techniques have become a hot research issue in computer science, been developed constantly, and applied successfully in many respects. In the research of biometrics, to solve problems through machine learning techniques has become a mainstream trend. Successful applications include utilazting the ensemble learning technique to solve the multi-template integration problem, and applying the semi-supervised learning techniques to relieve the inta-class variation problem, etc.This thesis analyses several key problems that influence the-recognition performance and application prospects of the fingerprint recognition, face recognition and multimodal biometrics systems, based on which, proposes appropriate machine learning techniques to effectively solve the problems to promot the recognition accuracy of the face and fingerprint recognition systems, and boost the convenience and recognition accuracy of the multimodal biometric system simultaneously. The main works and contributions of this thesis include:1. Minutiae based matching methods have always been the mainstream methods in fingerprint recognition. In such methods, minutiae that cannot be successfully matched are all treated as useless information. This thesis analyses that actually, since the sources and characteristics of unmatched minutiae are different in the matching of two homologous and two heterologous fingerprints, unmatched minutiae contain a certain amount of discriminative information. How to mine and apply such discriminative information reasonably is a key problem to promote the performance of the fingerprint recognition system. Based on this, the thesis defines and extracts 7 features from the unmatched minutiae. Accordingly,7 auxiliary matching scores are obtained. The ensemble learning technique is applied to fuse the auxiliary scores with the main marching score of the traditional minutiae based matching method to realize the recognition task. Experimental results on the databases of FVC2000, FVC2002 and FVC2004 for international fingerprint verification competition validate the discrimination ability of the unmatched minutiae, and the electiveness of the proposed methodologies in boosting the performance of the fingerprint recognition system. The proposed method boosts the performance of the fingerprint recognition system by at least 33.6% and at most 77.0% on the 12 sub-databases in the three tested verification competition databases.2. This thesis analyses that in present face recognition, whether two face samples are from the same individual is only decided by the similarity of the two samples, while other useful information is ignored. Actually, the similarity of two face samples from different individuals is of stable regularity, and the minor significant similarity information is of certain discrimination ability, thus, how to mine and apply such useful information to face recognition is a key problem to promote the recognition performance. Based on this, the thesis proposes a Sparse Similarity Sequence based face recognition Method (S3M). Firstly, the concept of similarity sequence is defined to extract the similarity information of faces belonging to different individuals. Next, Lasso sparse learning method is applied to select discriminating information. In this process, a discrimination criterion is proposed to optimize the sparse degree. Finally, the face recognition method is realized based on the optimal sparse similarity sequence. The EER (Equal Error Rate) of the proposed method reaches 0.05% and 0.01% respectively on two public face databases, showing outstanding advantage on face recognition performance.3. The present multimodal biometric systems mainly adopt parallel fusion mode in which all biometric traits involved in the system need to be captured on both the enrollment and recognition phases, which brings great inconvenience to users. The defect in convenience and efficiency severely influences the scope of applications of multimodal biometric systems. This thesis proposes to use a serial fusion framework in which the more convenient traits are positioned earlier in the fusion chain, and the use of less convenient traits is saved whenever possible, so that the system provides the maximal convenience to users and meets the requirements of most applications. However, the thesis points out that more convenient traits are always of less discrimination ability, and the contradiction of convenience and performance leads to the fact that if we put more user convenient traits earlier in the chain, in order to guarantee high recognition accuracy, it may turn out that we still have to heavily use inconvenient traits for most users, which will discount the advantage in user convenience. How to solve the contradiction is the key problem to boost the convenience and accuracy of the multimodal biometric systems simultaneously. Based on the analysis, this thesis proposes to integrate Semi-Supervised Learning (SSL) techniques into the framework to solve the contradiction by enhancing the performance of the more convenient but weaker trait. The core SSL thechnique is the Dependence Maximum Dimensionality Reduction (DMDR) method by which the weaker trait is promoted under the supervision of the stronger trait. In DMDR, the unlabeled and labeled samples and the tight coupling of the weaker and stronger traits are used to transform the weaker trait into a less dimensional feature space so that the dependence of information of the stronger and weaker traits is maximized. The experiments on two prototype systems verify the effectiveness of the SSL method in enhancing the weaker trait and the ability of the proposed framework in boosting the convenience and the recognition accuracy of the system at the same time.