“Enhance!” is the often heard vernacular of crime shows such as CSI.
Lab technicians in these shows were seemingly forever prompted to turn a blurry, pixelated image of a face or a crime-scene into a crisp, high resolution version with a click of a mouse. Of course, software like that did and still does exist, but not to the extent that a heavily pixelated image (usually through extensive zooming in) could be rendered magically into anything but an approximation at best. You can't turn 256 pixels into anything recognisable unless you really squint and use your imagination.
Until now that is.
The AI called PULSE is taking 16x16 pixel images and converting them into a high resolution 1024x1024 pixel images. Pulse stands for Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models (if we once again use our imaginations a little), and is the world-leader by a long shot in generating what is termed super-resolution images from such limited resolution input. The project, based at Duke University, almost mirrors the accuracy of TV and Movie Magic that seem to pull information about an image out of thin air.
“Never have super-resolution images been created at this resolution before with this much detail.” Cynthia Rudin, lead researcher for the PULSE project and computer scientist at Duke University.
The Duke University researchers are hoping that their algorithm could be implemented across many fields, including medicine, satellite imagery and microscopy.
The Pulse AI uses machine learning in a technique called GAN (Generative Adversarial Network) in order to feed two neural networks a large amount of photos. One neural network then generates images whilst the other analyses the results to make sure the images pass strict parameters.
The results are spectacular, and once again prove the amazing diversity of application of AI and neural networks.