Semi-parametric Distributional Models for Biometric Fusion

The performance of biometric recognition systems can become limited in operating environments due to the presence of extraneous noise factors. Biometric fusion alleviates this problem by consolidating information from various biometric sources, thereby achieving a higher recognition rate. One challenge faced in fusion is how to optimally combine information originating from the different sources. We present a framework for fusing matching scores from multiple biometric sources using multivariate t-copulas. Our approach involves eliciting suitable models for the joint distribution of matching scores via the t-copulas to accommodate generalized marginals (densities having a mixture of both discrete and continuous components). For estimating the unknown parameters in our models, we show that the discrete components can be viewed as missing components in an ExpectationMaximization (EM) framework. The newly developed estimation technique is applied to model the distribution of genuine and impostor matching scores from multiple biometric sources. Biometric fusion is subsequently performed based on the likelihood ratio statistic.