Adapted Fusion Schemes for Multimodal Biometric Authentication

This Thesis is focused on the combination of multiple biometric traits for automatic person authentication, in what is called a multimodal biometric system. More generally, any type of biometric information can be combined in what is called a multibiometric system. The information sources in multibiometrics include not only multiple biometric traits but also multiple sensors, multiple biometric instances (e.g., different fingers in fingerprint verification repeated instances, and multiple algorithms. Most of the approaches found in the literature for combining these various information sources are based on the combination of the matching scores provided by individual systems built on the different biometric evidences. The combination schemes following this architecture are typically based on combination rules or trained pattern classifiers, and most of them assume that the score level fusion function is fixed at verification time. This Thesis considers the problem of adapting the score fusion functions in multimodal biometric authentication, with application also to other multibiometric scenarios. The term adapted in this Thesis refers to fusion approaches that are trained using background information, for example a pool of users, and then adjusted considering input information such as user-dependent scores or test-dependent quality measures. In this regard, the user-dependent score fusion methods found in the literature are not adapted to the users, but trained either on the pool of users or on the particular user being tested. On the other hand, the idea of adapted fusion from quality information was already embedded in some previous works, but not in an explicit way as developed in this Dissertation. After a summary of the state-of-the-art in fusion strategies for multimodal biometrics, a number of novel adapted fusion schemes are proposed. These approaches adapt either to individual users through a reduced number of user-specific matching scores or to the input biometric quality. User-dependent fusion methods are further classified into three groups: 1) user-dependent score normalization plus simple fusion, 2) user-dependent score fusion, and 3) user-dependent decision. For most of the proposed approaches, two implementations are given, one based on statistical assumptions and the other one based on discriminative criteria using Support Vector Machines. We then consider the issue of performance evaluation in multimodal biometric authentication systems, and introduce the experimental framework and the biometric databases used in this Dissertation. This is followed by the application of the proposed methods to competitive multi-algorithm approaches for three individual biometrics, namely: signature, voice, and fingerprint; using standard biometric data and experimental benchmarks. The experimental part of the Thesis starts with a study of user-dependent score normalization and decision in multi-algorithm on-line signature verification. For this study we introduce two new systems based on local and global information, respectively. The local system is also used to study various practical aspects of system development including feature extraction and modeling, and to demonstrate the benefits of incorporating user-dependent score normalization. We finally combine the local and global systems using simple score level fusion rules, demonstrating both the complementarity of the two approaches and the benefits of incorporating user-dependent decision thresholds. We then study the application of adapted user-dependent fusion to multi-algorithm speaker verification using third party systems. We compare user-independent, user-dependent, and adapted user-dependent versions of score level fusion. It is shown that the proposed adapted approach outperforms both user-independent and user-dependent traditional fusion schemes. After that, we study the effects of image quality on the performance of two common approaches for fingerprint verification. It is observed that the approach based on ridge information outperforms the minutiae-based approach in low image quality conditions. This is exploited by an adapted score-level fusion approach using quality measures estimated in the frequency domain. The proposed scheme leads to enhanced performance over the best matcher and the standard sum fusion rule over a wide range of operating points. Finally, a comparative study of the proposed adapted fusion schemes, both user-dependent and quality-based, is given for the case of multimodal authentication based on signature and fingerprint on the real bimodal database MCYT. The proposed approaches are demonstrated to outperform traditional non-adapted fusion schemes. The experimental results favor the adapted fusion schemes based on discriminative formulations with respect to the Bayesian approaches in the case of small training set sizes. The opposite occurs for large training set sizes.