Threshold-optimized decision-level fusion and its application to biometrics

Abstract

Fusion is a popular practice to increase the reliability of biometric verification. In this paper, we propose an optimal fusion scheme at decision level by the AND or OR rule, based on optimizing matching score thresholds. The proposed fusion scheme will always give an improvement in the Neyman–Pearson sense over the component classifiers that are fused. The theory of the threshold-optimized decision-level fusion is presented, and the applications are discussed. Fusion experiments are done on the FRGC database which contains 2D texture data and 3D shape data. The proposed decision fusion improves the system performance, in a way comparable to or better than the conventional score-level fusion. It is noteworthy that in practice, the threshold-optimized decision-level fusion by the OR rule is especially useful in presence of outliers.