Know the face

Big Brother may be watching you but it doesn't mean he recognises you, which is why researchers at University College London (UCL) are developing new automated facial recognition software. It will be able to identity faces regardless of variation in position, pose, expression, and illumination.

Current facial recognition software needs full frontal images to function properly. This usually requires the co-operation of the user, which works for access control to a building but is unsuitable for cameras acquiring images in uncontrolled conditions, such as with CCTV. The new system will use a Bayesian probability model to allow the recognition of images taken in such conditions.

'The most current commercial methods use high-resolution images to inspect little facial features,' said Dr Simon Prince, principal investigator, of UCL. 'Our technology can't work miracles but it can work with lower-resolution images. If, as a human, you think the image is useful, the computer should be able to do something with it.

'Some academics are creating 3D models of a face and they can use lighting and a geometrical model to make a comparison. We don't do that. We use statistical techniques that don't have knowledge of geometry. We collect lots of photos and the model learns the statistical irregularities between different photos of the same person.

'We built this generative statistical model that takes into account viewing conditions — specifically pose, illumination and expression and estimates of the probability that any given face matches another.'

The model is built up using photos of people in varying conditions. From this it can take into account the conditions in the image and analyse the probability of two different photos being the same person. This means for CCTV identification, the police would only require two images of the person, for example the CCTV image of the person and a mug shot.

Because it is assessing the likelihood of the picture being the same person rather than building a model of the face and matching it to another, the system doesn't require as much time or computer processing power to do its work.

'Of an admittedly small database of 300 people, normal straightforward frontal recognition has been 100 per cent accurate,' added Prince. 'With a 20º rotation of the face — which is when almost all commercial recognition systems drop to 50 per cent accuracy — the system still gives 100 per cent accuracy. At 60º accuracy is 99 per cent and at 90º we still get 92-93 per cent. It is significantly better than anyone else.'

The research is supported by the National Policing Improvement Agency (NPIA), the Home Office scientific development branch, the MoD systems analysis-system design department and Canada's York University.

'The area of automated face recognition is one that has the potential to bring great benefits to the police service,' said Geoff Whitaker, head of biometrics at the NPIA. 'But there are still a number of significant challenges to overcome before these can be fully realised, especially with respect to its use with "uncontrolled" and low-quality facial images such as those typically obtained from CCTV.

'Research projects such as this are an essential step along the way to the development of more accurate and robust facial recognition systems for use both by law enforcement agencies and in the civilian market.'

Prince believes the technology has more applications than simply security and could affect every aspect of life. 'In robotics there is a need to recognise humans and it could be useful in almost any device,' he said. 'Even your shower could use it to know what temperature you like. It has ubiquitous uses.

'It could also be used to search for images on the internet. Google images looks at the text, and sites like Flickr have thousands of pictures that are not labelled. Also, if your name is Britney Spears and you wanted to look for pictures of yourself there's no way of distinguishing you from the celebrity. This technology could distinguish individuals and make searching much easier.'