Time spares no one, especially when it comes to your youthful good looks. A newly developed AI system is able to forecast how a human will look as they age in the coming decades. Perhaps even more impressively, it can also reverse the effects of aging in photographs of people’s faces, making them appear 10 years younger.
Not only is this neat little study a good reminder to cut down on sunbathing, it could be an invaluable tool for police searching for missing individuals or long-lost fugitives.
The researchers tested out their new system on photographs of celebrities, including Michael Cera, Tom Hardy, Britney Spears, Justin Timberlake, and a bunch of other recognizable faces.
Previous attempts at this kind of thing, like the viral Facebook posts you might have seen, usually just add a few superficial features to faces. Basically, they just draw on a few wrinkles, stick some bags under the eyes, and recede the hairline a little. This novel method, however, draws on huge amounts of information and other images to create a subtle (and hopefully more realistic) face age progression, as you can see in the image above.
This works just as well reversing the age, as you can see below. This approach also allowed the AI to do its work regardless of whether glasses or hair were obscuring their face.
It works using a type of algorithm called a generative adversarial network (GAN) used in unsupervised machine learning. As Science reports, researchers trained their algorithm using 50,000 criminal mugshots and 163,446 face images of 2,000 different celebrities across 10 years. The first part of the system advances a person’s face to appear a target age. A separate algorithm is then used to estimate the person’s age by comparing it to a person actually of that age. This helps to assess the initial algorithm’s work and offers information on how to fine-tune its skill. Using this method, the GAN is able to learn how to realistically “age” human faces.
Humans also seem to be convinced by the new work. The researchers gathered a number of fleshy humans and asked them if this new method or previous methods created better images. Out of 1,380 votes, 69 percent voted for the new attempt, while 20 percent voted for prior work.
The full study, presented at IEEE Conference on Computer Vision and Pattern Recognition 2018 in February, can be found on the preprint server arXiv.