Two-step calibration method for multi-algorithm score-based face recognition systems by minimizing discrimination loss

Abstract

We propose a new method for combining multialgorithm score-based face recognition systems, which we call the two-step calibration method. Typically, algorithms for face recognition systems produce dependent scores. The two-step method is based on parametric copulas to handle this dependence. Its goal is to minimize discrimination loss. For synthetic and real databases (NIST-face and Face3D) we will show that our method is accurate and reliable using the cost of log likelihood ratio and the information-theoretical empirical cross-entropy (ECE).

Semiparametric likelihood‐ratio‐based biometric score‐level fusion via parametric copula

“…2. In order to reduce this calibration loss, we proposed in our previous work [7] a method called the two‐step calibration method . Briefly, the first step of this method is computing both training and testing sets to their fused scores once the best copula pair has been found and the second step is calibrating the fused scores by the PAV algorithm trained based on the fused scores of the training set.…”

“…Briefly, the first step of this method is computing both training and testing sets to their fused scores once the best copula pair has been found and the second step is calibrating the fused scores by the PAV algorithm trained based on the fused scores of the training set. Readers who are interested in the detailed explanation of the two‐step calibration method may refer to [7].…”

“…These parametric copulas are the same as used in [6,7]. Suppose that we have genuine and impostor scores as given in (8) and 9, respectively, that will be used to train our method with respect to an evaluation measure e. Our procedure to choose the best copula pair is simple.…”

“…To approximate the correction factor, we will use the following parametric copulas: ind, Gaussian copula (GC), Student's t (t), Frank (Fr), Clayton (Cl), flipped Clayton (fCl), Gumbel (Gu), and flipped Gumbel (fGu). Therefore, the copulas and are chosen from the copula family These parametric copulas are the same as used in [6, 7].…”

“…This paper uses the copula concept to handle dependence between the scores. Although the copula model has already been used in [4][5][6][7] for some different scenarios, none of them provides analytically how this model is built and why the estimation of parameters determining the model is reliable. After explaining some related works in Section 2, this paper will explain how a copula model splits the LR computation for two or more classifiers into a product of the individual LRs and a correction factor in Section 3.…”