This is an evaluation of different score-level fusion techniques, and the results of a variety of fusion experiments using face and fingerprint data from 187,000 individuals, with matcher scores from three fingerprint and three face recognition systems. Eight score-level fusion techniques were implemented and evaluated. These differed in effectiveness, in the types of training data required, and in the complexity of modeling of genuine and imposter distributions. The most effective fusion techniques were product of likelihood ratios and logistic regression. Techniques that were as effective were product of FARs and optimized linear. Multi-modal fusion is highly effective; fusing one fingerprint and face resulted in a 64-85% reduction in false reject rate at a constant false accept rate of 0.0001. Multi-instance fusion using fingerprints from multiple fingers is also highly effective; fusing two fingerprints resulted in a 48-90% reduction in false reject rate. Multi-sample fusion using the enrollment of two samples rather than one resulted in a 45-72% reduction in false reject rate. Multi-algorithm fusion using different matchers on the same data resulted in an 8-33% reduction in false reject rate.
Citation: NIST Interagency/Internal Report (NISTIR) - 7346
NIST Pub Series: NIST Interagency/Internal Report (NISTIR)
Pub Type: NIST PubsReport Number:
Biometric accuracy: Biometric fusion, Biometrics, Multi-modal matching