Authors: Lorène ALLANO
Keywords: Multimodal biometrics
Abstract: Multimodal biometrics, that is the combination of several biometric systems, is a widely studied area. Multimodality is interesting because it overrides most of the limitations of biometric systems taken alone. It provides a solution when enrolment is impossible for one modality and is more robust against intentional forgeries, while enhancing recognition performance. The advantages of multimo-dality compared to "monomodal" biometric systems are obtained in combining several systems.
The first part of this work consists of comparing several well-known score fusion methods, in the mobility context. We have analysed and compared these methods on degraded data corresponding to real-life mobile applications like identity verification on a handheld device or a mobile phone. In a second step,
we proposed an original sequential fusion scheme. This strategy exploits serial fusion of systems, in order to take the final decision with the optimal quantity of information. The proposed sequential fusion strategy allows keeping the same level of performance than systems performing parallel fusion. But the major advantage of this strategy is that it counterbalances the major drawback of multimodality coming from the fact that several modalities are used : processing time and usability.
In the second part, statistical dependence measures are presented and particularly a measure based on entropy called mutual information. We propose to
use those measures in order to justify the use of "virtual" databases. "Virtual" databases are composed of "non-real" users, artificially created by associating a modality from one person and another modality from another person. "Virtual" databases are frequently used in multimodal systems evaluation because "real" multimodal databases are not often available. This practice is justified by the claim of independence assumption between modalities. Our proposed measures of dependence will allow estimating whether modalities are dependent or not, and proving that, in case they are independent, virtual databases (generated randomly) can be used to estimate fusion performance. In case modalities are dependent, we propose a novel approach based on clusters of persons. Virtual databases are then generated by combining persons inside each cluster, and we prove that it allows preserving dependence between modalities and then estimating fusion performance on such databases.