Decision-level fusion strategies for correlated biometric classifiers

The focus of this paper is on designing decision-level fusion strategies for correlated biometric classifiers. In this regard, two different strategies are investigated. In the first strategy, an optimal fusion rule based on the likelihood ratio test (LRT) and the chair Varshney rule (CVR) is discussed for correlated hypothesis testing where the thresholds of the individual biometric classifiers are first fixed. In the second strategy, a particle swarm optimization (PSO) based procedure is proposed to simultaneously optimize the thresholds and the fusion rule. Results are presented on (a) a synthetic score data conforming to a multivariate normal distribution with different covariance matrices, and (b) the NIST BSSR dataset. We observe that the PSO-based decision fusion strategy performs well on correlated classifiers when compared with the LRT-based method as well as the average sum rule employing z-score normalization. This work highlights the importance of incorporating the correlation structure between classifiers when designing a biometric fusion system.