False Matches and Non-independence of Face Recognition Scores

Although face recognition algorithms have made significant progress over the years, they still lack the accuracy to accomplish many of the more demanding tasks that have been proposed. To address this problem, several authors have suggested supplementing biometric systems with soft biometric information such as age, height, and sex to improve the overall accuracy of the system. However, the improvement is contingent on the matcher not already intrinsically factoring the soft biometric information into its comparison score. In fact, since soft biometric traits tend to be reflected in the physical characteristics of the face, such traits are expected to correlate with face matcher scores. In this paper, two methods are used to explore the statistical relationship between soft biometric traits and non-match scores generated by a leading commercial face recognition algorithm. The first method uses a generalized linear model (GLM) to determine how age, sex, and nationality covary with the probability of a false match. The second method uses sample partitioning for a more direct presentation of how the probability of a false match varies for different combinations of soft biometric values. Results indicate that age, nationality, and sex have a statistically significant effect on non-match scores. We explore what makes soft biometric information useful for the purpose of improving the accuracy of a biometric system under the assumption that the additional subject information is used to filter comparisons between subjects that have incompatible soft biometric characteristics (e.g. comparisons between men and women).