Google's New Compression Tool Uses 75% Less Bandwidth Without Sacrificing Image Quality (thenextweb.com)

An anonymous reader quotes a report from The Next Web: Google just released an image compression technology called RAISR (Rapid and Accurate Super Image Resolution) designed to save your precious data without sacrificing photo quality. Claiming to use up to 75 percent less bandwidth, RAISR analyzes both low and high-quality versions of the same image. Once analyzed, it learns what makes the larger version superior and simulates the differences on the smaller version. In essence, it's using machine learning to create an Instagram-like filter to trick your eye into believing the lower-quality image is on par with its full-sized variant. Unfortunately for the majority of smartphone users, the tech only works on Google+ where Google claims to be upscaling over a billion images a week. If you don't want to use Google+, you'll just have to wait a little longer. Google plans to expand RAISR to more apps over the coming months. Hopefully that means Google Photos.

Comments

This is awesome

[Esteanil]

I can't wait until you get technology like this combined with eye tracking to decide on-the-fly what parts of your VR experience are the most visually important and can optimize rendering accordingly.

On of my main pet peeves with current VR is that I can't see why you'd need to render at full resolution outside of the eye's focus area, which should make it possible to massively reduce the rendering required to get amazing quality.

If you can also optimize by using machine learning to decide which areas are perceptually important that should make it possible to focus your processing resources even better on the parts that matter for the visual experience.

Example of the JBIG2 compression fiasco

[Guppy]

Let's hope Google has had the forethought to have the image recognition algorithm pre-screen for images containing numbers, letters, and diagrams. Pattern-matching compression can be pretty scary when it decides two patterns are close enough:

http://www.dkriesel.com/en/blo...