Through deep learning, researchers at Duke University have developed an algorithm called Pulse. He manages to create realistic images of faces from a simple low-resolution miniature.
Pixelating faces in a photo may soon no longer be enough to protect people’s identities. Researchers at Duke University in the United States have just developed an artificial intelligence capable of creating high-definition images from a miniature. The tool, dubbed “Pulse,” manages to multiply the definition by 64, from a 16 x 16 pixel miniature to a high-definition image in 1,024 x 1,024 pixels. The usual techniques try to add high-definition elements to a low-definition image, which often creates a blurry result. This new technique generates successive high-definition images that are compared to the miniature to a plausible result.
The project is based on deep learningand generative antagonistic networks (GAN). In practical terms, a neural network generates plausible faces, after being drawn onto a large collection of real photos. A second neural network tests each generated image and decides whether, when reduced, it resembles the miniature. As time went by, the first network refined its creations, until the miniature of the generated image was identical to the original one.
A realistic but approximate reconstruction
This technique does not magically reconstruct the person’s face in the photo original photo. It is simply a plausible possibility. Pulse is able to generate many variants with subtle differences, all of which will give the same miniature.
The researchers indicated that the technique is not limited to faces. The same system should be able to be applied to improve the definition of images in many areas, such as medicine, astronomy, microscopy and satellite imaging.