Researchers from Duke University have developed an AI tool that can turn unrecognizable, extremely blurry pictures of people’s faces into realistic-looking computer-generated portraits.
The team’s method, called PULSE, can scale an image of a face up to 64 times its original resolution, while other previous methods have only scaled up to 8 times the resolution. Pulse applies features such as eyelashes, fine lines, wrinkles, pores, strands of grey hair, and stubble that weren’t there in the first place and in more exquisite detail than ever before.
Cynthia Rudin, who led the team, said:
Never have super-resolution images been created at this resolution before with this much detail.
The AI tool cannot identify people who were once on the photo. “It won’t turn an out-of-focus, unrecognizable photo from a security camera into a crystal-clear image of a real person. Rather, it is capable of generating new faces that don’t exist, but look plausibly real,” the researchers clarified.
According to co-author Sachit Menon, PULSE could potentially take low-res shots and create sharp, realistic-looking images with applications ranging from microscopy and medicine to satellite imagery and astronomy.
The system uses a tool in machine learning known as GAN (generative adversarial network,) which consists of two neural networks trained on the same data set of photos. One network comes up with AI-made human faces, while the other takes this data and decides whether it’s convincing enough to be mistaken for the real thing. Through experience, the first network gets better until the second network can’t tell the difference.
Within seconds, PULSE can convert a 16×16-pixel image of a face to 1024 x 1024 pixels, adding more than a million pixels, similar to HD resolution. The team asked 40 volunteers to rate 1,440 images generated via PULSE and five other scaling systems on a scale of one to five. PULSE got the highest score, almost as high as high-quality photos of real people.
