Blurred Faces Are Now 60 Times Sharper With AI
Previous methods for improving the clarity of a picture can scale an image of a face up to eight times its original resolution but this team came up with a way to create realistic-looking faces with up to 64 times its original resolution, having features such as eyelashes or wrinkles that weren’t seen.
Traditionally, low-resolution images are used and ‘guess’ the extra pixels needed to get images to match, on average, with corresponding pixels in high-resolution images the computer has seen before. As a result of this, textured areas such as hair and skin may not line up perfectly from one pixel to the other and end up looking hazy.
The low-resolution images here are indistinct and hazy as the previous method had to ‘guess’ the extra pixels needed in order to form a complete picture. Image Credit: Daily Mail.
A new approach called PULSE searches through AI-generated examples of high-resolution faces to match ones similar to the input image when it is compressed to the same size. It is believed that this system cannot be used to identify people as the researchers state that it won’t turn an out-of-focus, unrecognisable photo from a security camera into a clear image of a real person. Rather, it’s capable of fabricating new faces that don’t exist but look moderately genuine.
This picture is from the method PULSE, which gives a much clearer picture of the person displaying the human’s pores, wrinkles and more. Image Credit: Daily Mail.
The team used a machine learning tool called a ‘general adversarial network,’ or GAN, which are two neural networks – one develops AI-created human faces which imitates the ones it was trained on, while the other takes the output and decides whether it is convincing enough to be mistaken for a real photo. It is believed that the first network gets better and better with experience until the second network is unable to tell the difference.
PULSE can create realistic images from poor-quality input that other methods cannot achieve. From a single blurred image of a face, it can generate a number of uncannily lifelike possibilities, each of which looks subtly like a different human.
This system is capable of converting a 16x16 pixel image of a face to 1024x1024 pixels in just a few seconds, which is 64 times the resolution, akin to HD resolution. Details such as wisps of hair, pores and stubble are unnoticeable in the low-res photos but become crystal clear in the computer-generated version.
40 people were asked by researchers to rate 1,440 images generated via PULSE and five other scaling methods on a scale of one to five, and results showed that PULSE did the best, scoring almost as high as high-quality photos of actual people.
Article Thumbnail Credit: https://techxplore.com/news/2020-06-artificial-intelligence-blurry-sharper.html