Likelihood Ratio in a SVM Framework: Fusing Linear and Non-Linear Face Classifiers

Abstract

The performance of score-level fusion algorithms is of-ten affected by conflicting decisions generated by the con-stituent matchers/classifiers. This paper describes a fusion algorithm that incorporates the likelihood ratio test statis-tic in a support vector machine (SVM) framework in order to classify match scores originating from multiple match-ers. The proposed approach also takes into account the precision and uncertainties of individual matchers. The re-sulting fusion algorithm is used to mitigate the effect of co-variate factors in face recognition by combining the match scores of linear appearance-based face recognition algo-rithms with their non-linear counterparts. Experimental results on a heterogeneous face database of 910 subjects suggest that the proposed fusion algorithm can significantly improve the verification performance of a face recognition system. Thus, the contribution of this work is two-fold: (a) the design of a novel fusion technique that incorporates the likelihood ratio test-statistic in a SVM fusion framework; and (b) the application of the technique to face recognition in order to mitigate the effect of covariate factors.