NYU research combines subject and age sets to improve aged-face imaging

Biometrics researchers in the U.S. have managed to preserve genuine facial identifiers while aging images of people’s faces.

A team working at New York University’s Tandon engineering school say they have developed a latent diffusion model that knows how to pull off the task. Code like that could help identify people who have been missing for years or more realistically augment actors’ faces in movies. In theory, it could also be used to spoof remote facial age estimation systems.

This process has been hampered by having too few photographs of a person over time to create authentic-looking age progressions. Here, the scientists relied on two teaching datasets with relatively modest numbers of images.

The first set of about 20 images contained all available images of the subject over time, teaching the algorithm to identify the person using facial recognition. The second, regularization, set of images – about 600 in experiments carried age captions, teaching the code the ways human faces age.

False non-matches were cut about 44 percent in experiments compared to modern baseline generative models, according to the researchers.

In fact, the researchers say, their method outperforms self-organizing tree algorithms for image editing including face-aging biometrics-preserving conditional (IPCGAN), attribute-editing (attGAN) and talk-to-edit GANs.

Article Topics

ageing  |  biometric identifiers  |  biometrics  |  biometrics research  |  face biometrics  |  spoofing