A facial composite is a picture of a suspect to crime (as seen on TV crime programs, newspapers, etc.) Current composite systems used by the police require witnesses to describe an assailant's face and then select individual facial features (producing a "composite" face). This process does not work well: we are not good at describing faces nor selecting individual features. We have developed a different approach here in Stirling. It is called EvoFIT and is showing considerable promise. Unlike current systems,
faces in EvoFIT are modelled in their entirety and thus not separated into
component parts. A facial composite (the new term for 'photofit') is created by first displaying 18 faces containing
random eyes, noses, mouths, etc. A witness selects a few of these faces
that are most similar to the target. The selected faces are then mixed
or "bred" together to produce a new set of 18 faces that more closely resemble
the target. The witness then selects a few of these faces to be bred together
again. The process of selection and breeding continues until an acceptable
likeness is reached. A facial composite is therefore created by evolution.
Recent tests have demonstrated that EvoFIT outperforms several current computerised composite systems. In the most recent study involving a Cognitive Interview and a (normal) 2 day delay to construction, EvoFITs were named roughly three times that of another UK composite system. EvoFIT also produces highly realistic-looking faces, as the image above demonstrates (it is of the pop singer Robbie Williams).
The problem The construction of a composite by assembling facial features (e.g. eyes,
nose and mouth) is common in criminal investigations. Photofit and Identikit are old fashioned systems that enable a composite to be assembled manually. Although these systems have now
been computerized (e.g. E-FIT and CD-Fit/PROfit), performance is still rather
poor when used to identify suspects. A shortcoming is believed to lie at
the heart of these methods: the selection of individual facial features.
This idea is backed up by considerable research suggesting that we recognize
faces by facial features seen at the same time (i.e. faces are perceived
"holistically"). This finding is easily demonstrated by the difficulty
identifying a famous face given just the eyes (as seen in magazines from
time to time).
Part of the problem is that current practices build composites with
a single face. Operating on only one face means that a witness must continuously
describe changes necessary to make the composite better. This "recall"
process is a naturally hard task for anyone to do. Recognizing a face,
on the other hand, tends to be both easy and accurate.
'Evolving' a composite A natural alternative then is to present groups of faces and allow a
witness to select a few of these based on their similarity to a suspect.
Selecting in this way is rather like picking criminals from a mugshot album.
The task can be carried out without having to describe a face. What is
required then is a method of combining these similar looking faces. Fortunately,
one can turn to nature for an answer.
In nature, plants and animals compete for resources in their surroundings.
This competition tends to result in the best of a species surviving to
maturity, mating and passing on genes that have been beneficial to their
survival. In this scenario, the likelihood of survival can be measured
by how well organisms are suited or adapted to their environment. For example,
the ability to be both smart and swift could be very beneficial to a wolf
and affect its longevity.
This idea can be applied to the solution of certain complex mathematical
problems. The approach (known as Genetic Algorithms) begins with a set
of random solutions to a problem. The best solutions are selected and then
"bred" together. This process is repeated until a good solution is found.
This is the basic approach used to create a composite with the EvoFIT system.
Here is an example
of Michael Owen being evolved over 4 generations.
Constructing an EvoFIT In more detail, EvoFIT begins by creating a set of faces with random
facial shapes and textures (an example
screen). A witness would normally select six of these shapes and textures
that most resemble a suspect. These selections then become the "parents"
of the next population; to produce another generation, the components of
the selected faces are mixed together. The "offspring" faces are then selected
and bred together as before. The selections enable the set of faces to
become more like the suspect. Evolution is completed when an acceptable
likeness is found; the relevant image is then saved on disk as the composite.
The EvoFIT system also contains a number of software tools to assist
in the creation of a composite. These include a utility allowing facial
features to be directly manipulated (Feature Shifter); an interface to
the major computerized composite systems (E-FIT and CD-FIT) allowing a range
of hairstyles to be selected; and software allowing facial adornments and
scars to be added in a number of photographic editing programs (e.g. Adobe
Photoshop). Details of the early development may be found in my Ph.D. at http://www.stir.ac.uk/staff/psychology/cdf1/Thesis.txt (a more recent evaluation is available below).
Evaluation
We have carried out several formal evaluations of the system, and the most recent ones have been positive. This work typically involves a 'witness' looking at a face for a short time and constructing a composite a couple of days later with either EvoFIT or another modern system (control). A recent popular magazine article on EvoFIT may be found here. Other research papers are available on my personal webpage.
Police work EvoFIT has also been used in a criminal investigation. Here are brief
details of Operation Mallard and the composite.
contact
Send any correspondence regarding EvoFIT to
Links
ABM is our industrial partner who are distributing EvoFIT and providing training and customer support.
A facial composite is a picture of a suspect to crime (as seen on TV crime programs, newspapers, etc.) Current composite systems used by the police require witnesses to describe an assailant's face and then select individual facial features (producing a "composite" face). This process does not work well: we are not good at describing faces nor selecting individual features. We have developed a different approach here in Stirling. It is called EvoFIT and is showing considerable promise. Unlike current systems,
faces in EvoFIT are modelled in their entirety and thus not separated into
component parts. A facial composite (the new term for 'photofit') is created by first displaying 18 faces containing
random eyes, noses, mouths, etc. A witness selects a few of these faces
that are most similar to the target. The selected faces are then mixed
or "bred" together to produce a new set of 18 faces that more closely resemble
the target. The witness then selects a few of these faces to be bred together
again. The process of selection and breeding continues until an acceptable
likeness is reached. A facial composite is therefore created by evolution.
EvoFIT has also been used in a criminal investigation. Here are brief
details of 
