A New Approach on Multimodal Biometrics Based on Combining Neural Networks Using AdaBoost

Abstract

In this paper, we propose a new approach in the field of multimodal biometrics, based on AdaBoost. AdaBoost has been used to overcome the prob-lem of limited number of training data in unimodal systems, by combining neu-ral networks as weak learners. The simplest possible neural network, i.e. only one neuron, plays the role of a weak learner in our system. We have conducted different experiments with different number of AdaBoost iterations (experts) and input features. We compared the results of our AdaBoost based multimodal system, using features of three different unimodal systems, with the results obtained separately from these unimodal systems, i.e. a GMM based speaker veri-fication, an HMM based face verification and a SVM based face verification. It has been shown that even the average FAR of the multimodal system (%0.058) is much less than the lowest FARs of each of the unimodal systems (%0.32 for SV, %4 for HMM based face verification and %1 for SVM based face verification systems), while the average FRR of the multimodal system is acceptable (%2.1).