Evaluation of LDA based face verification with respect to available computational resources

In this study, we are interested in a face verification or authentication system based on Principal Component Analysis and Linear Discriminant Analysis (known as Fisherfaces). We evaluate the tradeoff between performance and computational requirements of making a decision in such a biometric system. This kind of evaluation is useful when implementing a practical system where the CPU power and storage space are of concern, such as a biometric system coupled with smart cards. Our results, on the standard XM2VTS database, show that the performance stays constant in a broad range of image resolutions (down to 256 pixel gray-level images) and client model sizes (down to 20 real numbers). When using 64 pixel images, the error rates are still acceptable for a mediumto low-security application or to be combined with other biometric algorithms. These results suggest that this method uses only low-frequency information to make a decision.