Shape of things to comeAge-guessing software has security, commercial applications

Published 24 September 2008

Fighting Illini researchers develop an age-guessing software which can perform tasks such as security control and surveillance monitoring, and may also be used for electronic customer relationship management

We know a lovely and engaging 92-year old woman who lives in our building and who, when the subject of age comes up, describes herself as “a woman of her middle years.” People who want to keep their age a secret may not want to go near a computer running a new software being developed at the University of Illinois — a software which can fairly accurately estimate a person’s age just as an age-guesser at a carnival would do. Unlike age-guessers, who can view a person’s body, the software works by examining only the person’s face. “Age-estimation software is useful in applications where you don’t need to specifically identify someone, such as a government employee, but would like to know their age,” said Thomas Huang, the William L. Everitt Distinguished Professor of Electrical and Computer Engineering at the U. of I.

For example, age-recognition algorithms could stop underage drinkers from entering bars, prevent minors from purchasing tobacco products from vending machines, and deny children access to adult Web sites, said Huang, who leads the Image Formation and Processing group at the university’s Beckman Institute.

Estimating someone’s age is not an easy task, even for a computer. This is partly because the aging process is determined not only by a person’s genetic makeup, but by many other factors as well, including health, location and living conditions. “Human faces do convey a significant amount of information, however, and provide important visual cues for estimating age,” Huang said. “Facial attributes, such as expression, gender and ethnic origin, play a crucial role in our image analysis.”

Consisting of three modules — face detection, discriminative manifold learning, and multiple linear regression — the researchers’ age-estimation software was trained on a database containing photos of 1,600 faces. The software can estimate ages from 1 year to 93 years. The software’s accuracy ranges from about 50 percent when estimating ages to within 5 years, to more than 80 percent when estimating ages to within 10 years. The accuracy can be improved by additional training on larger databases of faces, Huang said.

In addition to performing tasks such as security control and surveillance monitoring, age-estimation software also could be used for electronic customer relationship management. For example, a camera snapping photos of customers could collect demographic data — such as how many adult men and women buy burgers, or what percentage of teenagers purchase a particular soft drink. Or, combined with algorithms which identify a person’s sex, age-estimation software could help target specific audiences for specific advertisements. For example, a store display might advertise a new automobile or boat as a man walks by, or new clothing or cosmetics as a woman walks by. “All of this can be done without violating anyone’s privacy,” Huang said. “Our software does not identify specific individuals. It just estimates their ages.”

-read more in Guodong Guo et al., “Image-Based Human Age Estimation by Manifold Learning and Locally Adjusted Robust Regression,” IEEE Transactions on Image Processing 17, no. 7 (July 2008):1178-88