Facial Comparison (also known as 'facial mapping')

In a case of disputed identification, how can the identity of a person in a CCTV image be established? The issue may arise of whether a facial image captured by CCTV is of the same person as a photograph of the suspect. There is no single method with an agreed procedure to compare the identity of two facial images. 'Facial mapping' refers to a number of techniques adopted by scientists and experts from a variety of backgrounds. The methods differ substantially in their scientific rigour and reliability.

No method of facial comparison can give absolute certainty that two images are of the same person. For example, it would be impossible to distinguish between identical twins from typical CCTV images. However, a measure of the similarity between facial images can be obtained. There are a number of problems that confront any comparison of facial images. 1). CCTV images are often of poor quality. 2) Even slight differences in the pose and lighting of faces make the comparison less reliable. 3) Facial images of different people can closely resemble each other. 4) Facial images of the same person taken with different caneras on separate occassions can look very different. 5) Human performance in matching the identity of unfamiliar faces, especially from poor quality images, is unreliable.

There are three main approaches to facial comparison. Morphological comparison, anthropometric comparison and video superimposition. Morphological comparison involved categorising facial features according to their type or shape and counting the number of similar categorisations across the comparison images. The method is based upon subjective judgements of facial features. Anthropometric comparison involves comparing measurements of distances and angles between facial landmarks. Video superimposition involves wiping from one video image to another which has been scaled and rotated to provide the best match. This method can provide compelling graphics in courts but has the weakness of requiring a human judgement of similarity with little or no formal treatment of the error possible.

Any sound method of facial comparison should include 1) A formal treatment of the error of measurement that arises due to: the quality of the images, the reliability of the measurement process, difference between measurements taken by different people, error due to minor changes in pose, changes in lighting etc. 2) Comparison of the differences in measurements, or other data, taken from the comparison images with comparable data from an appropriate database of facial images. Are the differences between the two images in question less than the differences between one of the images and the faces in the database (known to be of a different person)? 3) Evaluation of the comparison between the images in question that is informed by both the error of measurement and comparison with the relevant database. Not all methods presented in expert opinion have these features.

At Valentine Moore Associates, we are active researchers in human face recognition and identification with more than 20 years experience. We are mindful of scientific evidence that humans are error-prone when judging the identity of two facial images captured with different cameras, even when the quality of the images is good. With poor quality images human judgement of the identity of unfamiliar faces is very unreliable. For facial comparisons we use a method of measuring distances between facial landmarks which are then expressed as ratios and angles between features. Differences in the size of the images (perhaps introduced by differences in distance from camera of the type of camera) are removed from the comparison.