Nearest neighbourhood classifiers in a bimodal biometric verification system fusion decision scheme

Identity verification systems that use a mono modal biometrics always have to contend with sensor noise and limitations of feature extractor and matching. However combining information from different biometrics modalities may well provide higher and more consistent performance levels. A robust yet simple scheme can fuse the decisions produced by the individual biometric experts. In this paper, k-Nearest Neighbourhood (k-NN) based classifiers are adopted in the decision fusion module for the face and speech experts. k-NN rule owes much of its popularity in pattern recognition community to its simplicity and good performance in practical application. Besides that, k-NN may also provide a ternary decision scheme which is rarely found in other classifiers. The fusion decision schemes considered are voting-, modified- and theoretic evidence of k-NN classifiers based on Dempster-Shafer theory. The performances of these k-NN classifiers are evaluated in both balanced and unbalanced conditions and compared with other classification approaches such as sum rule, voting techniques and Multilayer Perceptron on a bimodal database.