It is not uncommon to see that Google is strongly committed to the development of artificial intelligence in the world of images. The Californian has a strong development in neural processes that increasingly seek to break the limits of technology. The world of computational imaging is one of the fields of greatest interest. Therefore, this time the company shows us progress on its super image enlargement system.
Single image magnification
As you can see on the company’s blog, Google’s new neural system is capable of resizing images at approximately 16 times their original resolution. Unlike other models where there may be very noticeable artifacts or an image almost without texture is produced, andGoogle’s model shows great results when scaling the image. This result is obtained with a new variant of analysis based on image diffusion models.
Image diffusion models work using a destructive and reconstructive method. The image is gradually destroyed with Gaussian noise, removing most of the detail. An algorithm then reconstructs that information through a regenerative process that uses destructive data, explains DPReview. Through this analysis of destruction and reconstruction, the model can identify and predict information variations, resulting in a better obtaining of final detail.
To achieve the scaled image, Google’s process starts with a dual system using Repetitive Image Refinement (SR3) and a Cascade Diffusion Model (CMD). While the SR3 model achieves very comprehensive ‘super-resolution’ results, when mixed with the data processed with CDM to create a network of high fidelity images the results obtain a percentage of score much higher than any other model.
Among the comparend models we can see that of Pulse, which we presented in June of last year. Although PULSE is one of the most complete, the PULSE reconstruction process is not accurate enough in the final result of the image reinterpretation. However, where a perfect algorithm would receive 50% confusion rate value (how the effectiveness of algorithms is measured), the PULSE model reaches only 33.7% while the Google model reaches a rate of 47.4%.
Where will we see it applied?
As is to be expected from these developments, seeing them applied in software can be a matter of days, months or years. With the Google Pixel 6 and the new Tensor processor, the Californian will apply more elements based on artificial intelligence for the optimal development of images and mobile processes. It is very likely that this image enhancement technology be implemented commercially to help limit the resolution and detail that mobiles can process natively.
Google has shown us that it is working on tools for color enhancement, scene backlighting, and we even know how the smart bracketing works enabled by HDR + technology that can currently be used in GCam. So it wouldn’t be uncommon to see the company focus on maintaining lower MP cameras but with better resolution upscaling capabilities using algorithms like the ones we see today.