Researchers at Duke University have developed an artificial intelligence tool that can turn blurry, unrecognizable photos of faces into new computer-generated photos that can be more than 60 times more sharp. Cynthia Rudin of Duke University says the new system we developed is not intended for face recognition, but rather produces a more realistic, high-definition image. Previous methods generally only increased the resolution of images to eight times the original, but our new method could make them up to 64 times higher.
The artificial intelligence tool is called the Build Anti-Network (GAN), two neural networks trained on the same set of photo data. Instead of using low-resolution images and slowly adding new details, the tool searches for high-resolution face samples generated by artificial intelligence, zooming in to the same size to make it look as like an input image as possible.
The system converts a 16×16 pixel face image to 1024×1024 pixels in seconds. In low-resolution photos, details such as pores, wrinkles, and a wisp of hair can be made clearer in computer-generated images.