A Fusion-Based Approach to Enhancing Multi-Modal Biometric Recognition System Failure Prediction and Overall Performance

摘要

Competing notions of biometric recognition system failure prediction have emerged recently, which can roughly be categorized as quality and non-quality based approaches. Quality, while well correlated overall with recognition performance, is a weaker indication of how the system will perform in a particular instance - something of primary importance for critical installations, screening areas, and surveillance posts. An alternative approach, incorporating a Failure Prediction Receiver Operator Characteristic (FPROC) analysis has been proposed to overcome the limitations of the quality approach, yielding accurate predictions on a per instance basis. In this paper, we develop a full multi-modal recognition system integrating an FPROC fusion-based failure prediction engine. Four different fusion techniques to enhance failure prediction are developed and evaluated for this system. We present results for the NIST BSSR1 multi-modal data set, and a larger "chimera" set also composed of data from BSSR1. Our results show a significant improvement in recognition performance with the fusion approach, over the baseline recognition results and previous fusion approaches.