Charlie Frowd, a psychology professor at the University of Central Lancashire in England, has developed a new criminal composite sketch system called EvoFIT. An example of one of the computerized faces is on the monitor.
By Mark Roth Pittsburgh Post-Gazette
When Brian Coates meets with police officers to tell them about a new computerized system he is using to produce images of criminal suspects, he'll start with some questions.
"I'll say, 'How many of you have partners?' " the detective constable said in an October interview in Driffield, England, "and most of them will hold up their hands."
Then he'll pick out one officer and say, "tell me about your partner's ears."
"He'll say, 'What?' "
"Describe her ears to me, I'll say. Do they have long lobes or short ones?"
" 'I don't know,' he'll say."
"Well what about her nose?"
" 'Well, it's normal,' he'll say."
"But what about the shape of the nostrils?"
" 'Well, I don't know,' he'll say, and by this time, of course, they're all laughing."
"Finally I'll say, 'Well if I held up a picture of her, would you know her?' And of course he'll say yes."
This little party game is Mr. Coates' way of making a point: We don't recognize faces -- even the ones of people we live with -- by being able to recall their eyes, noses, ears or eyebrows.
Instead, we see and know their faces as a whole. And that's the approach embodied in the computerized composite system Mr. Coates is field testing in the Humberside region on the east coast of England.
Known as EvoFIT, it has been developed over the past several years by a team at the University of Central Lancashire in western England, led by psychology professor Charlie Frowd.
EvoFIT tries to mimic the brain's face recognition system by showing witnesses a series of face images that have been morphed by the computer, and then asking them to select the ones that most closely match their recollection.
That's a sharp departure from older computerized systems like E-FIT, which asked witnesses to select individual noses, eyes, ears and other features and place them on a facial template.
In early tests, both E-FIT and EvoFIT showed about a 20 percent success rate of someone being able to identify a person from a composite, but those were lab tests that allowed the artists to create images immediately after a witness had seen a face.
When witnesses waited two days between seeing a photo and helping prepare a composite, though, the E-FIT method dropped to 5 percent identification rates, Dr. Frowd said.
Today, in real-world field tests, EvoFIT is getting naming rates on its composites of about 25 percent, and lab tests he is doing now have given him hope that soon the system will achieve rates of almost 60 percent.
That may not seem highly effective, but it means that composite sketches have gone from being virtually useless to having enough power now that a member of the public -- or a police officer -- will be able to identify a likely suspect in many cases, he said.
Many police departments in the United Kingdom are already testing EvoFIT, Dr. Frowd said, and he hopes to start a trial soon in America with the Boston police department.
The EvoFIT system, like other "holistic" approaches, tries to give witnesses a variety of whole faces to choose from, rather than asking them to construct a face, Mr. Potato Head-style.
The EvoFIT operator starts by having witnesses choose a hairstyle, because that is what they tend to remember best.
But then, it immediately blurs the hair, ears and neck so that the witness can start focusing on the face's internal features, which are what actually distinguish one visage from another.
Witnesses are asked to pick a basic face shape, then a basic texture, and from those, the computer generates a whole set of faces whose features are varied in shape, size and their relationship to each other.
By selecting the best of those examples and letting the computer breed new images based on those choices, Dr. Frowd said, witnesses can come much closer to the likeness of the person they saw than by using the older parts-based systems.
Even the E-FIT system itself, which is now under new ownership, is converting to this more holistic approach.
Composite images of suspects are most useful when the police either don't have a likely suspect, or the crime was committed by someone who doesn't have a record.
Detective Constable Roz Greening, of the Devon & Cornwall major crimes unit covering southwestern England, said composites have a unique advantage.
"If you hold up a strand of DNA," she said in an October interview, "you don't know who that is unless it's in the system. If you hold up a fingerprint, you don't know who that is unless you have it in the system. But if you hold up a picture, a lot of people will know who that is."
There are several reasons why it's still a challenge to get a good witness description, though, experts say.
First, many people develop a kind of tunnel vision when they see a crime, said Kareem Johnson, a face researcher at Temple University.
"One of the reasons people are bad at identifying suspects is, if someone holds a gun on you, you pay attention to the gun and you don't really pay attention to the face," he said.
Most people also are not very skilled at recognizing faces they've seen only once, said Vicki Bruce, a pioneering face researcher at Newcastle University in England.
"We're good at recognizing familiar faces," she said, but most people's ability to recognize unfamiliar faces "is lousy."
For that reason, EvoFIT tries to make the composite creation process as intuitive as possible.
What makes EvoFIT easier to use, especially with elderly witnesses, is it doesn't put a great burden on them to describe the features of the person they're trying to remember, said Mr. Coates' partner, Investigating Officer Colin Prosser.
"It doesn't matter that they can't describe the person," Mr. Prosser said, "as long as they can look at the screen and can recognize the faces that are nearest to their memory."
Neither the Humberside nor Devon & Cornwall departments can cite a high rate of convictions based on composite sketches, but that has more to do with the United Kingdom's rules of evidence than with EvoFIT itself, Ms. Greening said.
"Eyewitness identification in our country is almost like secondary intelligence; it's not prima facie evidence at all," she said. If prosecutors don't have independent forensic evidence or a confession, they often won't take the case to court.
That happened in one recent rape case, she said, even though EvoFIT proved its worth.
A young woman said she had been sexually assaulted about a week earlier by a man she had met in a bar, and she helped produce an EvoFIT composite of his face that was then publicly broadcast.
Almost immediately, calls began coming in to the station, Ms. Greening said. "He was a member of a local football team," she said, "and all his mates had seen the image and said, 'That looks like you and you need to get your backside into the police station, and if you don't go in, then we'll give your name to the police.' "
The man turned himself in, but because he argued the sex was consensual and the woman had not gone to the hospital to provide medical evidence, the prosecutor would not proceed with the case, Ms. Greening said.
Dr. Frowd is now pouring a lot of energy into trying to improve EvoFIT's accuracy.
He is using two approaches to do that -- holistic tools and caricatures.
The holistic tools come into play after the witness has selected the best image he or she can from the computerized faces. Using various sliders, the witness can then tweak the image to change its weight, age and masculinity, plus some other personality dimensions that aren't named.
The caricatures are developed after the witness has signed off on the final image. By varying the face more toward an average set of features, known as a negative caricature, and then exaggerating the differences from average, known as a positive caricature, the EvoFIT system can create a set of images that can be animated on TV or a website.
Lab tests have shown that the moving caricatures of a face help people to identify the image, possibly because each of us may remember a person's face in a slightly different form than other people do, Dr. Frowd said.
While composite systems are obviously far from perfect, Mr. Coates said that when he asked his officers recently if they would want to go back to the previous system, "there were almost gasps of consternation at the very idea we might want to go back."