Using Independence Assumption to Improve Multimodal Biometric Fusion

Abstract

There is an increased interest in the combination of biometric matchers for person verification. Matchers of different modalities can be considered as independent 2-class classifiers. This work tries to answer the question of whether assumption of the classifier independence could be used to improve the combination method. The combination added error was introduced and used to evaluate performance of various combination methods. The results show that using independence assumption for score density estimation indeed improves combination performance. At the same time it is likely that a generic classifier like SVM will still perform better. The magnitudes of experimentally evaluated combination added errors are relatively small, which means that choice of the combination method is not really important.