Evaluation and performance prediction of multimodal biometric systems

Multibiometric systems fuse the evidence presented by different biometric sources in order to improve the matching accuracy of a biometric system. In such systems, information fusion can be performed at different levels; however, integration at the matching score level is the most commonly used approach due to the tradeoff between information content and accessibility. This work develops a tool in order to analyze the impact of various normalization schemes on the matching performance of score-level fusion algorithms. The tool permits the systematic evaluation of different fusion rules after employing normalizing and mapping the match scores of different modalities into a common domain. Furthermore, it provides a method to fit various parametric models to the score distribution and analyze the goodness of fit statistic based on the Chi-Squared and Kolmogorov-Smirnov tests. Experimental results on multiple datasets indicate the benefits of normalization, the role of parametric distributions and the variations in matching performance on different databases.