AI has been revolutionizing various different fields of technology lately, and it seems that Google is at the forefront of most of them. Image processing is one of these, and Google has once again wowed photography enthusiasts with its newest AI tool, a fully functioning “Zoom and Enhance” engine.
This new artificial intelligence system by Google does exactly what it sounds like; it zooms in on an image and enlarges it without ruining the picture quality, just like zooming in on criminals’ faces from CCTV cameras in the movies. This technique works by adding more details into the original picture which were originally not captured by the camera using similar-looking images or just some really good guesswork. The technique, called natural image synthesis by Google allows you to turn a blocky, pixelated image into a sharp and natural-looking photo.
Google has taken help from two AI tools to perfect this technique. The first is the SR3 or Super-Resolution via Repeated Refinement which works by adding noise to an image and then reversing it by taking it away using a neural network. The second tool is CDM or Cascaded Diffusion Models, which are like pipelines through which diffusion models are directed for high-quality image resolution which enhances the images and makes them larger. If you want to read more in detail about how the model works, check out Google’s published paper on it.
The new AI engine was tested on ImageNet, which is a giant database comprising of training images. The results of using SR3 and CDM on these images were quite impressive as about 50% of images generated by SR3 were mistaken for real human faces which means that the AI is doing something right! While the enhanced images aren’t exact matches for the originals, they’re as good as technology allows for now.