Researchers from Duke University created an artificial intelligence (AI) tool that can turn blurry pictures of faces into computer-generated portraits with fine details.
Previous methods to clarify blurry images can only scale up an image of a face up to eight times its resolution, leaving the image blurry and indistinct. However, the new system, called PULSE, can generate realistic faces with up to 64 times resolution, creating images with fine lines, eyelashes and stubble.
Facial features like eyes and lips are barely distinguishable in the blurry photo on the left. Enlarged more than 60 times (right) it is a different story — thanks to artificial intelligence. Source: Rudin lab
According to the team, PULSE cannot be used to identify people. Rather, the system creates an entirely new face from the blurry images, taking lower resolution shots of almost anything and creating sharp and realistic pictures for applications such as medicine, microscopy, astronomy and satellite imagery.
PULSE searches AI-generated examples of high-resolution faces and searches for new ones that look as close to the input image when shrunk to the same size. To achieve this, the team used a machine learning tool called a generative adversarial network (GAN). With this method, two neural networks were trained on the same dataset of photos. One network comes up with AI-created faces that mimic the faces it was trained on, and the other takes the output and decides if it is convincing enough to be mistaken for a real face. The first network gets more accurate with experience until the second network cannot tell the difference between a real photo and an AI-generated photo.
PULSE can create a number of lifelike images from a single blurred image of a face, even if the photo has eyes and mouth that are unrecognizable. It can convert a 16 x 16 pixel image of a face to 1024 x 1024 pixels in seconds, adding more than a million pixels to create HD resolution.
To test the system, researchers asked 40 people to rate 1,440 images that were generated by PULSE and five other scaling methods. The images were rated on a scale from one to five. PULSE performed the best and scored almost as high as high-quality photos of actual people.
The team presented the research at the 2020 Conference on Computer Vision and Pattern Recognition (CVPR). The paper can be accessed here.
