Oct. 14, 2008 -- They say it's not how old you are, but how old you feel. With new age recognition software, it's actually how old you look.
Developed at the University of Illinois, Urbana-Champaign, computer software could analyze an image of your face to verify your identity or run a commercial that would appeal to your, ah, level of life experience.
"Age measurement is very difficult," said Thomas Huang, the lead developer. "If you use the face to estimate age we can really get the apparent age, or how old a person looks."
The researchers trained their computer algorithm using 1,600 different people with five pictures of each person, for a total of 8,000 images. The age of the people in the pictures ranged from one year to 93 years old.
Huang didn't tell the computer what to look for. The computer searched the faces and used its own software to determine which features best determined the person's apparent age.
Related Content:
One feature the computer looked at was gray scale. If, for instance, the face in the picture measured 100 pixels by 100 pixels, all 10,000 separate pixels has its own level of gray (no color yet). By comparing how dark or how light each pixel was compared to other pixels, the software estimates the apparent age of the individual.
"A woman wearing makeup should get a younger age," said Huang. "A smoother skin texture will register as younger-looking."
The computer also looked at the shape of the face. The relative positions of the eyes, nose, ears, the shape of the mouth, all change over time and can help indicate a person's age.
"If you use the real age as the 'ground truth', then the accuracy is quite low," said Huang. "But if we estimate [a person's age] to within 10 years, then the accuracy is about 80 percent."
Shape, position, color and texture don't just reveal age. They also give away ethnicity, gender, even emotions, all of which Huang is studying with computer software.
"We are more accurate in recognizing the gender of a person, about 90 percent accurate," said Huang. "We are also about 90 percent accurate at recognizing positive expressions like smiling. Negative emotions, like surprise, aren't as easy though."
This isn't just an academic exercise in computer science. Several companies have approached Huang about using his software. While he won't give details, he says that face-recognition software would be useful for, say, fast food companies that want to know how many male teenagers buy a particular sandwich or clothing companies that could run a different ad to each person who walks into their store.
"If you can estimate the gender or the age of the viewer you can change the display," said Huang. "For a younger viewer you might want to display one type of commercial, and show a different commercial for an older viewer."
Face-recognition software would also be useful for security, says Jonathon Phillips of the National Institutes of Standards and Technology, who works on face recognition software.
"Soft biometrics" like age or gender (as opposed to "hard biometrics" like fingerprints and palm vein scans) could be use as an additional safeguard to a PIN number to verify a person's identity at automatic teller machines (ATMs) or at building entrances.
Whatever the software is used for, expect to see age recognition systems within the next few years. The system could tap into existing security cameras to capture images which would then be run through the software.
Huang and Phillips are trying to ensure that no additional hardware, such as cameras or special lighting, is necessary. One difficulty with Huang's software is that it works best when the face is turned towards the camera. Camera side angles make it more difficult to predict age, gender, ethnicity or emotion. Huang is refining his software to ibe able to read face images at different angles.
The ultimate test of the recognition software is how well it guesses age, compared with people. Huang thinks that his software is just as good as humans at age estimation. Phillips says that humans are exceptionally good at recognizing the ages of our friends and loved ones, but we are bad at assessing the age of strangers.
"We do it all the time," said Phillips. "When people design any kind of service they want it to be a positive experience. If the computer is more accurate than a human at estimating age, emotion, etc. then it will improve the ability of computers to interact with humans in a more positive manner."
Related Links:
Dept. of Electrical and Computer Engineering, University of Illinois
our sites
video
mobile
shop
stay connected
corporate
