Handling high dimensionality in biometric classification with multiple quality measures using Locality Preserving Projection

The use of quality measures in biometrics is rapidly becoming the standard strategy for improving performance of biometric systems, especially in the presence of variable environmental conditions of signal capture. It is often necessary to integrate multiple quality measures into the classification process in order to capture the relevant aspects of signal quality. The inclusion of multiple quality features quickly increases the dimensionality of the classification problem, which leads to the risks of overfitting and dimensionality curse. So far, no mature strategy of coping with multiple quality measures has been developed. In this paper we propose to use a scheme, where the dimensionality of the vector of quality measures is reduced using the Locality Preserving Projections. We show that the proposed technique offers higher accuracy and better generalization properties than existing techniques of classification with quality measures, in same- and cross-device biometric matching scenarios.