You’ve probably seen that scene where the policeman zooms in on satellite images to identify a face or license plate. Away from movie theaters, this reality is somewhat frustrating even for those who need to carry out a much simpler task, such as enlarging a family photo. Now Google researchers have unveiled a new technology that promises sharp magnifications using Artificial Intelligence (AI).
Super-Resolution via Repeated Refinements (SR3) is an approach based on probabilistic models that produces sharper results from small, blurry images, using an algorithm to combine complementary information from several different sources.
A TV with “Upscaling” (scaling up, in free translation) can adapt the content you are watching to the resolution of your screen. This technology magnifies the pixels and recognizes which parts of the image need to be optimized to improve the experience when watching movies or series at lower resolutions.
Want to catch up on the best tech news of the day? Access and subscribe to our new youtube channel, Canaltech News. Everyday a summary of the main news from the tech world for you!
The SR3 uses an algorithm with ResNet and PixelCNN neural network systems designed to work together, filling in the missing details in low-resolution images to turn them into high-resolution copies. Rather than “guessing” what’s missing, this new upscaling rebuilds the image, delivering much sharper results.
To improve quality and sharpness, the SR3 uses a trained model to progressively add noise to a high-resolution image. Then, the algorithm “learns” how to reverse this process, removing pure noise until achieving a similar result in low-resolution images.
The system removes Gaussian noise (white noise) to improve the output signal, using a U-Net model — convolutional neural network architecture developed for image segmentation — prepared to reduce noise at various levels of resolution. This ensures high performance in magnifying faces and other types of images.
Super-Resolution achieves amazing results by resizing images at resolutions up to eight times lower, while maintaining quality and sharpness. This model can also be used to increase clarity in cascading images by constantly increasing scaling with resolutions — a 32×32 image, for example, could reach up to 256×256 pixels.
In the future, this lossless image enhancement technology could equip ordinary televisions, telescopes and cell phones with advanced superzoom systems capable of magnifying people and objects miles away without blurring or bizarre distortion. He’s still far from that Hollywood policeman who sees the whites of a criminal’s eyes approaching a photograph taken from a satellite, but he can already dream of a better camera.
Source: Google AI
Did you like this article?
Subscribe your email on Canaltech to receive daily updates with the latest news from the world of technology.