Predicting biometric facial recognition failure with similarity surfaces and support vector machines

The notion of quality in biometric system evaluation has often been restricted to raw image quality, with a prediction of failure leaving no other option but to acquire another sample image of the subject at large. The very nature of this sort of failure prediction is very limiting for both identifying situations where algorithms fail, and for automatically compensating for failure conditions. Moreover, when expressed in a ROC curve, image quality paints an often misleading picture regarding its potential to predict failure. In this paper, we extend previous work on predicting algorithmic failures via similarity surface analysis. To generate the surfaces used for comparison, we define a set of new features derived from distance measures or similarity scores from a recognition system. For learning, we introduce support vector machines as yet another approach for accurate classification. A large set of scores from facial recognition algorithms are evaluated, including EBGM, robust PCA, robust revocable PCA, and a leading commercial algorithm. Experimental results show that we can reliably predict biometric system failure using the SVM approach.