Combination of score level fusion methods in receiver operating characteristic space

In fingerprint verification systems impressive improvements have been achieved through multi sample fusion methods. Among fusion methods, score level fusion with its simplicity and high performance is the most common and useful fusion method. But the quality of fingerprints has direct effect on performance and accuracy of these systems. In this paper, we present a combination approach in Receiver Operating Characteristic space using Support Vector Machine to combine score level fusion methods on multi sample fingerprints with small sample sizes. This approach uses False Match Rate from genuine class and True Match Rate from impostor class as one feature on training. For testing based on Bayesian decision theory one of the FMR or TMR is selected as a feature. We compared our combination approach with score fusion methods and combination based on FMR and FNMR as two features. The experimental results on Iran University of Science and Technology fingerprint Database show that the proposed approach with no need to normalization has doubled the distance between distribution of two classes and the accuracy has been improved to 0.997 for Equal Error Rate 0.129.