A Statistical Approach to Facial Identification

This thesis describes the development of statistical methods for facial identification. The objective is to provide a technique which can provide answers based on probabilities to the question of whether two images of a face are from the same person or whether there could be two different people whose facial images match equally well. The aim would be to contribute to evidence that an image captured, for example, at a crime scene by CCTV, is that of a suspect in custody. The methods developed are based on the underlying mathematics of faces (specifically the shape of the configuration of identified landmarks) At present expert witnesses carry out facial comparisons to assess how alike two faces are and their declared expert opinions are inevitably subjective.

To develop the method a large population study was carried out to explore facial variation. Sets of measurements of landmarks were digitally taken from ≈3000 facial images and Procrustes analyses were performed to extract the underlying face shapes and used to estimate the parameters in statistical model for the population of face shapes. This allows pairs of faces to be compared in relation to population variability using a multivariate normal likelihood ratio (MVNLR) procedure. The MVNLR technique is a recognised means for evidence evaluation, and is widely used for example on trace evidence and DNA matching. However, many modifications and adaptations were required because of unique aspects of facial data such as high dimensionality, differential reliabilities of landmark identification and differential distinctiveness within the population of certain facial features.

The thesis describes techniques of selection of appropriate landmarks and novel dimensionality reduction methods to accommodate these aspects involving non-sequential selection of principal components (to avoid ephemeral facial expressions) and balancing of measures of reliability against selectivity and specificity.