THE German police were stumped. In January 2003 they discovered a skull in a forest near Celle, north of Hannover. They could tell that its owner was a man and that he had been about 56 when he died. He still had most of his teeth, had undergone extensive dental work, and had broken his nose long enough before his death for it to have healed. DNA samples indicated that he had dark hair and a southern European complexion. Yet two years later, they still had no idea who the man was.
As luck would have it, experts in the forensic art of reconstructing how a person鈥檚 face would look based only on their skull were due to meet in Germany that year for a conference. So the police offered to pay the conference fee for any of the specialists who were willing to conjure up the face of the man behind the Celle skull. In the end, 21 researchers took up the challenge.
Their results, presented at this year鈥檚 International Conference on Reconstruction of Soft Facial Parts, held last month in Leuven, Belgium, were not what the police had hoped for. The 21 faces they came up with are all completely different, with none more likely than any other to be accurate. The Celle skull鈥檚 identity remains a mystery.
Advertisement
Even though facial reconstruction is widely used for difficult cases, this represented the largest experimental test to date for the technique. One thing it clearly shows is that the field, which has historically been the province of police artists, remains more art than science. That needs to change, says Caroline Wilkinson, a forensic anthropologist at the University of Dundee in the UK. 鈥淲e need to work by rules, not intuition.鈥
All facial reconstruction techniques start with the bare skull, and build up the soft tissue layer by layer. Traditionally, artists have used strips of clay for each muscle, but now many manipulate computer images instead.
What the resulting face looks like is determined by a combination of the shape of the skull and the thickness of the tissue layers above it. To decide how much to add, reconstructors use average thicknesses of skin, fat and muscle at different points on the skull. These parameters used to have to be measured on cadavers, but more recently ultrasound and CT scanners have been used to take measurements from living people.
This is where the trouble begins. Some of the standard measurements used in the past are simply wrong, says Carl Stephan, a biological anthropologist at the University of Queensland in Brisbane, Australia. From an analysis of the anatomical literature, Stephan has found that people鈥檚 eyes stick out of their sockets an average of 4 millimetres more than the reconstruction guidelines say they do. The idea that roughness on the skull indicates the location of a person鈥檚 hairline is another long-held idea that fails to stand up to scrutiny.
A more fundamental problem results from the way we recognise faces. 鈥淲e carry an average or prototype face in our minds, and recognise individuals from the subtle ways they diverge from it,鈥 Stephan says. If police trying to put a name to unidentified remains want to show the public a reconstructed face that someone may recognise, that reconstruction must show these divergences from the norm.
鈥淲e carry an average face in our minds, and recognise people by how they diverge from it鈥
This is precisely what a reconstruction based on average tissue thicknesses will lose, and it doesn鈥檛 take much averaging to lose individuality. John Clement of the University of Melbourne in Australia has been compiling large digitised databases of Caucasian and Japanese faces. When the researchers created faces by combining as few as 15 Caucasian or Japanese individuals, the result was always the same 鈥渁verage鈥 Caucasian or Japanese face, which did not change as more faces were added. 鈥淲e were staggered by how few it took,鈥 he says.
So how can faces be reconstructed in a way that recaptures their individuality? Dirk Vandermeulen and colleagues at the Catholic University of Leuven (KUL) believe the answer lies in using real faces to help them reconstruct unknown ones. They performed ultrasound scans on the faces and skulls of nearly 400 people, and on each one they measured the distances between 52 landmarks such as the tip of the chin or the corner of the eye, plus the thickness of the soft tissue at each of those points.
All these measurements, plus the body mass, age and gender of each person, are now stored in a computerised 鈥渇ace space鈥. Each face is assigned a point within this space that reflects the value of each of these variables. The resulting map will show young males, for example, clustered at one part of the face space, with elderly females at another. Within these groupings those with the most similar features will be closest together. 鈥淚t鈥檚 like a three-dimensional space,鈥 says team member Peter Claes, except there are many dimensions, corresponding to each of the variables.
When an unknown skull is entered into the 鈥渟pace鈥, the system picks out those faces which lie closest to it. The tissue from these faces can then be altered to fit the measured differences in the skull to give it a face of its own. Claes says that because the system starts with real faces, it yields a more plausible reconstruction than earlier computerised methods that simply fit a standard digital sheet to a skull using average tissue thicknesses. The system can be used to generate several plausible reconstructions of each skull.
The researchers are now testing the system by reconstructing the face for one of the skulls in the database. Observers are then shown a set of 11 similar faces, including both the reconstructed face and the real face that belonged to the skull, and asked to identify which two are the same person.
All reconstruction assumes that some vestige of the face鈥檚 characteristic features can be found on the skull, but there are no guarantees that they can. 鈥淒oes the shape of the skull predict the upper lip?鈥 asks Stephan. The width of the teeth may determine the tension on the lip tissue, for example, but no one knows how closely that relationship holds. Generating several plausible faces might get around such variability.
Wilkinson argues that advances such as these will never do away with the need for artists. However theoretically accurate a reconstructed face is, it will not help if it is clumsily produced and looks artificial, says Richard Neave of the University of Manchester in the UK, who was one of the pioneers of clay-based reconstruction. 鈥淭he artist brings years of skill and training to producing a face,鈥 he says. 鈥淭here are many spaces in a face you don鈥檛 have the science to fill in. What do you do with the corners of the mouth? They鈥檙e important.鈥
I鈥檝e seen that face before
Despite years of research into facial reconstruction, it is still unclear which features play the most important role in allowing us to recognise each other.
In the past the accuracy of facial reconstruction techniques has been gauged by creating a face from the skull of a known person and then comparing the result with the person鈥檚 real face. One problem that arises is how to make the comparison, as people are notoriously bad at matching up pictures of faces. When asked whether or not two different pictures were of the same face, people have been found to correctly identify pictures of the same person only 70 per cent of the time. To make a more objective assessment, scientists now take a CT scan of a living person鈥檚 skull, and then create a digital reconstruction of their face. The accuracy of this reconstruction is gauged by comparing it with the real face using a large number of measurements such as the distance between the chin tip and lower lip.
The problem here is that two faces can have closely similar measurements and yet look no more recognisably similar than faces with markedly different measurements, says Carl Stephan of the University of Queensland in Brisbane, Australia. Conversely, the wildly distorted cartoons of public figures and celebrities do not actually resemble the real face, yet the cartoon is instantly recognisable because it retains and exaggerates the face鈥檚 peculiarities.
Jens Bongartz of the Koblenz University of Applied Sciences in Germany has developed a system that might make it possible to chart the relationship between resemblance and recognition. His software breaks faces down into elements based on the distances between different facial landmarks, and can also be used to create artificial faces in which some of these elements have been deliberately altered. This might make it possible to compare different faces to see which elements are crucial to recognition.
