Incorporating user specific normalization in multimodal biometric fusion system

The aim of this paper is to investigate the user-specific two-level fusion strategy in the context of multimodal biometrics. In this strategy, a client-specific score normalization procedure is applied firstly to each of the system outputs to be fused. Then, the resulting normalized outputs are fed into a common classifier. The logistic regression, non-confidence weighted sum and the likelihood ratio based on Gaussian mixture model are used as back-end classifiers. Three client-specific score normalization procedures are considered in this paper, i.e. Z-norm, F-norm and the Model-Specific Log-Likelihood Ratio MSLLR-norm. Our first findings based on 15 fusion experiments on the XM2VTS score database show that when the previous two-level fusion strategy is applied, the resulting fusion classifier outperforms the baseline classifiers significantly and a relative reduction of more than 50% in the equal error rate can be achieved. The second finding is that when using this two-level user-specific fusion strategy, the design of the final classifier is simplified and performance generalization of baseline classifiers is not straightforward. A great attention must be given to the choice of the combination normalization-back-end classifier.