Optimising multimodal fusion for biometric identification systems

Biometric systems are automatic means for imitating the human brain’s ability of identifying and verifying other humans by their behavioural and physiological characteristics. A system, which uses more than one biometric modality at the same time, is known as a multimodal system. Multimodal biometric systems consolidate the evidence presented by multiple biometric sources and typically provide better recognition performance compared to systems based on a single biometric modality. This thesis addresses some issues related to the implementation of multimodal biometric identity verification systems. The thesis assesses the feasibility of using commercial offthe-shelf products to construct deployable multimodal biometric system. It also identifies multimodal biometric fusion as a challenging optimisation problem when one considers the presence of several configurations and settings, in particular the verification thresholds adopted by each biometric device and the decision fusion algorithm implemented for a particular configuration. The thesis proposes a novel approach for the optimisation of multimodal biometric systems based on the use of genetic algorithms for solving some of the problems associated with the different settings. The proposed optimisation method also addresses some of the problems associated with score normalization. In addition, the thesis presents an analysis of the performance of different fusion rules when characterising the system users as sheep, goats, lambs and wolves. The results presented indicate that the proposed optimisation method can be used to solve the problems associated with threshold settings. This clearly demonstrates a valuable potential strategy that can be used to set a priori thresholds of the different biometric devices before using them. The proposed optimisation architecture addressed the problem of score normalisation, which makes it an effective “plug-and-play” design philosophy to system implementation. The results also indicate that the optimisation approach can be used for effectively determining the weight settings, which is used in many applications for varying the relative importance of the different performance parameters.