This page describes work by Mike Burton & Rob Jenkins, which was begun on a project funded by ESRC. It stems from an attempt to understand how people can recognise faces over a huge variation in images (see figures below). It turns out that the research also has significant applications in automatic, computer-based face recognition, for example systems used in security settings.
Recognition of faces over a range such as this is a very difficult problem. Yet, for viewers familar with this person, it is easy to recognise all the different photos. The image-variability arises because of changes in the person (age, expression, pose, hairstyle etc), change in the capture device (cameras with different focal lengths, resolution etc), change in the viewing conditions (direction and degree of illumination), and many other variables. For these reasons, most research on automatic face recognition avoids this level of variability.
The images at the centre of these blocks are an average of the other twenty. The papers cited below give details about the procedure, which involves computing simple means of photographic images.
The average of a particular face has some very nice properties. For example, any particular image is lit from a given direction - a variable which engineers sometimes attempts to "factor out" before recognition. However, the average image has no directional lighting, because over a set of images, this will cancel out. Instead, what remains in the average is only information which is present consistently across images of that person, i.e. information which is diagnostic of that person's identity.
Here's an example of my own face, which shows the photos which were used to generate the 'average Mike Burton' on my home page.
We have used these averages to construct a theory of face representation and face learning in human perceivers. We have also suggested that they could be very useful in building artifical face recognition devices. One of the appeals of the aproach is its simplicity. Constructing an average is a very simple way of achieving something very complicated, i.e. stripping away all the information in an image which is not about who that person is.
To read more about this proposal, have a look at
- Jenkins, R. & Burton, A.M. (2008). 100% accuracy in automatic face recognition. Science, 319, 435.
- Burton, A.M., Jenkins, R., Hancock, P.J.B. & White, D. (2005). Robust representations for face recognition: The power of averages. Cognitive Psychology, 51, 256-284.
- Jenkins, R., Burton, A.M., & White, D. (2006) Face recognition from unconstrained images: Progress with prototypes. Proceedings of the Seventh IEEE International Conference on Automatic Face and Gesture Recognition, pp 25-30.
Back to Glasgow Face Recognition Group
Back to Mike Burton's Home Page