Rank-Based Score Normalization for Multi-Biometric Score Fusion

The matching score distributions produced by different biometric modalities are heterogeneous. The same is true for the matching score distributions obtained for different probes. Both of these problems can be addressed by score normalization methods that standardize the corresponding distributions. In our previous work we demonstrated that, in the case of multi-sample galleries, the matching score distributions are also heterogeneous between different subsets of matching scores obtained for the same probe. In this paper, we use this result to propose a rank-based score normalization framework for multi-biometric score fusion. Specifically, in addition to normalizing the matching scores produced for each biometric modality independently, we propose to further join them to form a single set. This set is then partitioned to subsets using a rank-based scheme. The theory of stochastic dominance demonstrates that the rank-based scheme imposes the distributions of the subsets to be ordered. Hence, by normalizing the matching scores of each subset independently, better normalized scores are produced. The normalized scores can be fused using any fusion rule. Experimental results using face and iris data from the CASIA-Iris-Distance database demonstrate the improvements obtained.