GO IN DEPTH

Averaging Faces

Averaging Faces

Creating composite faces may help computers recognize people.


Transcript

Recognizing the average face. I'm Bob Hirshon and this is Science Update.

Even sophisticated computer programs aren't very good at recognizing human faces. That's because our appearance changes from day to day. But psychologist Rob Jenkins of the University of Glasgow in Scotland has a solution. In a recent study, he and colleague Mike Burton created computerized composite photos of celebrities, by averaging dramatically different photos of each person.

Jenkins:

You had changes in hairstyle, facial hair, decades of aging, changes in weight and health, and changes in the conditions under which the photo was taken.

Popular face recognition software identified the composites with 100 percent accuracy, compared to 54 percent for normal photos. Jenkins says that putting these composites on passports or other ID cards could make it possible to design rock-solid security systems. I'm Bob Hirshon for AAAS, the science society.


Making Sense of the Research

In this age of identity theft and global terrorism, it's become more and more important to create reliable forms of ID. For a long time, we've relied on humans to match a driver's license or other picture ID to the person carrying it. That's not a bad system, but it's subject to human error. Can computers do better?

In theory, yes, but so far, not really. It turns out it's actually very difficult to teach a computer how to recognize a human face. Humans are very good at recognizing someone we know, even if he or she gets a new hairstyle, wears different glasses, grows a beard, or shows up after being away for decades. We also can make adjustments for environmental conditions like lighting, which can change the way people look.

However, we're not so good at identifying people we don't know. That may sound obvious, but it might surprise you to learn that given two different pictures of strangers, taken under different conditions, you'd be hard pressed to say whether they're actually the same person or not. Unfortunately, most people checking ID's don't know the people whose identity they're trying to verify. So it would be very useful if a computer could take the guesswork out of that process.

In designing computerized face recognition systems, it's been difficult to get computers up to the accuracy level of familiar people. Generally, they work by comparing a fixed image—for example, a photo filed in its memory bank, or a photo on an ID—to an actual person's face. To make them work really well, designers have had to control lighting and other environmental conditions at the checkpoint, and ask its subjects to keep their appearance consistent and look at the computer with a blank expression. That's hard to do even in small, tightly controlled environments like a top-secret military base, to say nothing of an airport or school.

Rather than try to improve the recognition technology itself, Jenkins and Burton decided to design a better photo template. They culled photos of celebrities from the Internet, seeking out pictures that varied wildly in setting, age, and appearance. Then they averaged the photos together into a single composite. Using someone else's face recognition software, they found that computers could be up to 100 percent accurate in matching these composites to other, individual photos of the same celebrity.

In order for this to work in the real world, of course, you would have to create these averaged photos. The averaged photo could be stored inside a gatekeeping program and matched to the actual person's face to identify him or her. Alternatively, Jenkins suggests that if passport and driver's license photos were actually averaged images, a computer could verify whether the person holding the document is, in fact, who they say they are. Creating these systems would require no small amount of work, but it might be a shortcut compared to constantly refining facial recognition software to make it more and more accurate.

Now try and answer these questions:

  1. What is an “averaged” face?
  2. Why do you think averaged faces helped the computers so much? How does this relate to the fact that humans are good at recognizing familiar people in different contexts, but not strangers?
  3. How does this strategy differ from making better facial recognition software? Why might it be better? Are there potential disadvantages?

You may want to check out the February 29, 2008, Science Update Podcast to hear further information about this Science Update and the other programs for that week. This podcast's topics include: what happened to our vitamin C, new insights into childhood leukemia, and why artificial sweeteners make rats fat.


Going Further


For Educators

The National Geographic News Article Brain Has "Face Place" for Recognition, Monkey Study Confirms describes research in monkeys that relates to humans' keen face recognition talents.

The United States Government's Biometric Consortium serves as a focal point for research, development, testing, evaluation, and application of body-based personal identification technology.


Related Resources

Cell Phone Traffic
6-12 | Audio
Server Naps
6-12 | Audio
Ancient Mexican Food
6-12 | Audio

Did you find this resource helpful?

Science Update Details

Grades Themes Project 2061 Benchmarks
AAAS Thinkfinity