Biometric score calibration for forensic face recognition

Abstract

When two biometric specimens (such as two faces or two fingerprints) are compared using an automatic biometric recognition system, a similarity metric called “score” can be computed. In forensics, one of the biometric specimens is from an unknown source, for example, a face image from CCTV footage or a fingermark found at a crime scene and the other biometric specimen is obtained from a known source, for example, from a suspect. Automatic biometric recognition systems are gradually replacing the forensic examiners' manual comparison of the two biometric specimens. In forensics, there is a huge interest to use a suitable metric to report the output of the comparison of the two biometric specimens. This has led to the use of the likelihood-ratio P(s|Hp)/(P(s|Hd)), where s is the score computed by an automatic biometric recognition system, Hp is the hypothesis of the prosecution (which states that the two biometric specimens are obtained from a same-source) and Hd is the hypothesis of the defense (which states that the two biometric specimens are obtained from different sources). Generally, two sets of training scores, one under Hp and the other under Hd, are needed to compute a likelihood-ratio from a score. In this thesis, we review several methods of likelihood-ratio computation focusing mainly on the issue of the sampling variability in the sets of the training scores and the specific conditioning imposed on the pairs of the biometric specimens to compute the training scores. We propose a simulation framework which can be used to study several properties of likelihood-ratio computation methods. This is useful for an appropriate and informed choice of a likelihood-ratio computation method. It is shown that sampling variability is a serious concern when small sets of the training scores are available for likelihood-ratio computation. Training sets can be suspect-specific or generic. An empirical study is carried out to quantify the effect of the two different kinds of training sets considering a speaker, a face and a fingerprint recognition system. The work is more focused on forensic face recognition systems. The concept of likelihood-ratio is applied to several existing biometric face recognition systems in order to study how they perform in forensic evidence evaluation. Furthermore, a study of the discriminating powers of different facial features such as eyes, eye brows, nose, etc. is carried out. This kind of regional comparison is inspired by the forensic examiners’ practice when they manually compare two face images. It proves very useful in situations where a part of the face is available for comparison. Analysis of the discriminating powers of facial features is also useful for future automation of the practice of forensic examiners.