A Picture Paints A Thousand Words
Try this now. Think of your best friend’s face. Now think how you would describe it to a stranger who Had never met them, so they could be recognised in a police line-up.
If this sounds like an unlikely party game then you’d be right. But imagine it wasn't your best friend you were describing but the person who had just knocked you down and stolen your wallet, and you were describing them to the police. I
f you’ve watched any television at all over the past two decades you’ll be familiar with the composite pictures the police use to build up a picture of suspects. These were introduced in the 1970s, and have become a part of standard police procedure, but now a Canterbury Enterprise Hub company think they have found a better way to help witnesses describe what they saw.
Dr Chris Solomon, Technical Director of VisionMetric explains: “The problem with the current system is that it relies on the ability of a witness to provide a detailed description of the suspect’s face. This is a process that people find very difficult – try describing the face of close friend! We wanted to develop an accurate facial composite system that was easy for people to use.”
The VisionMetric system uses a system of point mapping (see inset) to reduce the human face to a finite number of variables (60 in fact) representing each of the different possibilities for a feature.
Chris: “Given that you can create a very realistic face using just 50 or 60 numbers we wondered if we could manipulate those numbers to produce a composite. The key idea is that we can use the numbers to produce a good approximation of anybody’s face.
“It’s quite a powerful idea, that you could have a sort of facial pin number for everyone that’s about 60 digits long.”
The witness selects a face from a set of templates, which represents the correct gender and ethnic group. From that starting point the witness proceeds by selecting an appropriate hairstyle from an integral database.
So far it’s not so different from current techniques. It’s what happens next that makes the VisionMetric system unique.
Stuart: “We’ve built a statistical model that tells us how the shape and skin colour of faces varies over the whole population. This model enables us to create ‘virtual’ computer-generated faces that exhibit a high degree of realism.”
Chris: “The trouble is that the process involves manipulating 60 variables which would be overwhelming for the witness. We needed a more intuitive approach.”
MIMICKING EVOLUTION
The approach they found uses evolutionary algorithms. An evolutionary algorithm is a mathematical process that mimics the Darwinian process of natural selection. What that means is that the system generates nine random faces that the witness ranks in order of accuracy.
The algorithm uses these selections to ‘breed’ a new generation of faces. The higher the ranking the more likely the computer is to select the face to be one of the breeding pairs.
“This process closely mimics the biological process,” says Chris, “In the sense that the numerical codes act as genes as they swap to produce a new set of faces that are the descendants of the original choices. Over a number of generations you are tending to favour the sorts of facial characteristic that the witness is looking for, without them having to describe it.”
Fitting in with how the Police work is a big part of developing the software: there are legislative issues (they have to work within the bounds of the Police and Criminal Evidence Act (PACE). But Chris is hopeful that the system might change how the police work, if only a little. “There are several benefits we hope. It’s more accurate. So it should improve the likelihood of a positive identification.
“But it’s also faster. If we can create a composite more quickly then there are a lot of benefits, not only financially but also because the quicker a suspect is identified the more likely an arrest is.
“Our other issue was usability. Although E-fit is a fantastic piece of software, it’s not very easy to use. Typically a police force will have two trained operators. The new system is much more intuitive to use, so we hope that forces will have many more officers who will be able to operate it. This can actually be used by officers on the beat or on a laptop in a police car, so that will increase the size of the market.”
MAKING SCIENCE INTO BUSINESS
Although the VisionMetric team has been working on the intellectual puzzle for over three years, making a business out of the science has been just as taxing.
Stuart: “The first thing to note is that selling to the Police is not like selling to the general public. It is important to build up a trust relationship and involve the end user in the design process. We were fortunate in having an Angel Investor, which enabled us to acquire Aspleys (the company that produces E-fits), which has improved our access to the police market: 84% of police forces in the UK at the moment use E-fit.”
Chris: “But we are looking for additional finance right now – we have plans to expand internationally. The advantages we have there are that Aspley’s have existing sales relationships with clients in over 20 countries; and that the UK Police is widely admired abroad, and other countries often copy what it does. We need the money to launch a concerted sales and marketing effort.”
“The IP for the business came from our research group, but we’ve had a lot of help with the business development, in part from the Enterprise Hub. In particular they’ve helped us with business development, and we’ve had an office here within the Hub for a little over a year. It’s often difficult to make the transition from research to business, but the University and the Hub have been extremely helpful and encouraging – we’ve had some excellent business advice.”
Stuart: “We’ve also benefited from the Hub Network. We’ve recently begun a collaboration with another Canterbury Hub client, Video Vest Ltd. It is developing a product that uses a small video camera placed in a Police Officer’s jacket to record incidents. We will be developing software to enhance the functionality of their device, which is exciting for both companies.”
Canterbury Enterprise Hub Director, David Butler, commented: “The future challenge to the company will be to develop the point-mapping process so that it can be used by police forces throughout the world across the whole range of human facial characteristics – an exciting prospect!”
CRIMINAL FIGURES - BRITISH CRIME SURVEY OFFENCES ENGLAND AND WALES
The British Crime Survey (BCS) showed that there were 10.9 million crimes committed against adults living in private households in England and Wales in 2004|05.
Vehicle-related theft was the most prevalent type of crime in the 2004|05 BCS with 1.9 million offences, 17 per cent of all offences in England and Wales.
Men in England and Wales were almost twice as likely as women to be a victim of violent crime (5 per cent compared with 3 per cent) with young men aged 16 to 24 most at risk in 2004|05.
Between 1993 and 2004 the average prison population in England and Wales rose by 67 per cent, to 75,000 – on 30 September 2005 it was 77,300.
Source: Office for National Statistics
HOW POINT-TO-POINT MAPPING WORKS
Point mapping is a method of comparing variations in the characteristics of similar objects. In this case faces.
Chris: “We’ve moved away from the idea of making a face simply as a collection of individual features. We take real photos of people and then delineate the main features using what we call a point model. If I look at your face I can place a point on the corner of your right eye, and on different parts of your face. I place points on identical places on pictures of other people’s face to generate what we call ‘points of correspondence’. We do that with points all over the face in such a way that you delineate the shape of the whole face and also specific features.”
The computer can then understand how the points vary as we move from one face to another, and it learns the ways in which the points can plausibly vary – “what we sometimes call the “legal variation”.
Chris: “The benefits are that it’s a generative model. This treats the entire face as a group of numbers that obey statistical rules. That allows us to use our evolutionary approach to generating composites – and it’s that that makes this system so powerful.”
www.visionmetric.co.uk
www.canterburyhub.co.uk
Contact: Dr Chris Soloman
Published: 23rd March 2006