Police searching for a long-lost person or fugitive sometimes have little more to go on than an old photograph. Artists or computer programs can attempt to predict what these individuals look like today, but both approaches have flaws. Now, scientists have harnessed advanced artificial intelligence (AI) to render artificial aging that’s more realistic (and depressing) than ever.
The system uses a two-part AI algorithm called a generative adversarial network (GAN). The first part takes a face and produces another face of the same individual at a target age. During training, a second part compares this image with a real image of someone at that age and with the original image and provides feedback, encouraging the first part to improve its abilities. Other artificial aging systems have used GANs, but this one differs by focusing not just on getting the age right, but also on maintaining the individual’s identity. Unlike others, it also renders foreheads and (lack of) hair, as seen in the photos of Justin Timberlake and Kirsten Dunst above.
The researchers trained their AI on more than 100,000 images from two databases, including mugshots and celebrities at different ages. A separate computer program then judged how the AI performed on a novel set of images. When the AI aged photos of people more than 20 years, so that people under 30 were meant to look between 50 and 60, for example, the computer program saw them (on average) as a 60-year-old (for mugshots) or a 52-year-old (for celebs). This analysis was not performed on prior work, but human participants deciding whether the new results or images from prior attempts to age people looked most like a younger source image chose the new images 70% to 9%, the researchers report in a paper to be presented this month at the Conference on Computer Vision and Pattern Recognition in Salt Lake City.
And yes, you could use the method on yourself. But don’t despair at your thinning hair line: You can also use it to subtract the decades.
By Matthew Hutson
