Researchers from Duke University in North Carolina (USA) managed to do what was previously possible only in films. Namely – to turn photos with insufficient resolution into more megapixel images. The PULSE neural network is capable of transforming a low-quality pixelated image of a person into a portrait with high definition and detail.
“Never before has it been possible to reproduce an ultra-high resolution image with so many fine details,” said co-author Cynthia Rudin.
Low resolution images are images with insufficient data. Therefore, the algorithm essentially “thinks out” and draws small features and details of the face, mimic wrinkles, eyelashes and the like. A portrait that matches the original cannot be reconstructed from a pixelated image in principle, the researchers say.
Therefore, the images are very realistic, but at the same time unreal – such a face may not exist.
Existing algorithms for increasing image resolution learn from high-resolution images and their reduced copies. But this method does not allow you to select high-quality textures for hair, skin and other complex elements. Because of this, the final image completed by the neural network still looks fuzzy.
Researchers at Duke University used a different principle, using a neural network that independently generates an image. The system generates a high definition image from its data set. The resulting image is reduced by a traditional algorithm and compared with a given image. The process is repeated until the generated thumbnail is 100% the same as the original image.
Experts agree that the research findings represent a breakthrough in machine learning. But we are not talking about a breakthrough in practical use. The authors of the development also insist on this.
Although the neural network is not capable of reconstructing the original portrait of a person from a low-resolution photograph, it is suitable for the opposite: reshaping people’s faces. The PULSE system generates a clear portrait, but it is impossible to recognize a real person in it.
For example, the system will be useful for police officers who need to hide the face of informants or key witnesses, and journalists. Previously, the face of such people was blurred in a photo editor, and there were concerns that this process could be reversible.
Also, this technology can make a breakthrough in the field of media and cinema, where a particular person is not particularly important, but the image plays a role. The neural network will allow you to get ultra-high definition of old films, archival newsreels, cartoons. The characters will not be 100% true to the original, but the end result will look attractive.