Google's RAISR uses machine learning to create high-quality upscaled versions of low-res images

Google's RAISR uses machine learning to create high-quality upscaled versions of low-res images
Google has just unveiled a new machine-based image upscaling technology. The technique is called RAISR and stands for Rapid and Accurate Image Super-Resolution. In a nutshell, RAISR can create high-quality upscaled versions of low-resolution images.

In the blog post than unveiled RAISR, Google Research Scientsit Peyman Milanfar says that RAISR obtains results that are comparable or better than the current super-resolution upscaling methods. While the quality is often comparable, Google claims that RAISR works 10 to 100 times faster than current alternatives. Another of RAISR's strengths is its ability to avoid recreating any aliasing artifacts contained in the original low-resolution image.

Standard upscaling methods create new pixels by applying simple combinations to the existent pixels in the low-resolution images. Such linear filters are fast but their results are not always satisfying.

Instead of following the traditional route, Google's RAISR applies filters selectively to each pixel in the image. RAISR uses machine learning to find out which filter works best for each pixel. It does this by analyzing the differences between upscaled versions of low-resolution images and their direct high-resolution pairs. As the system goes through more image pairs, it is better able to predict which filter will work best on each pixel. Obviously, this is just a simplified explanation of how the system works. If you're interested in learning more about RAISR, head on over to the Google Research blog post through the source link below.

Now that you've been introduced to how RAISR works, here are a few sample photos:


The potential for a fast image upscaling technology is immense. Aside from enabling the refresh of any photo taken at low resolution by modern standards, we can imagine that RAISR could be integrated with Google's Camera app, Google Photos, and a range of other apps.

source: Google

FEATURED VIDEO

9 Comments

1. FlySheikh

Posts: 443; Member since: Oct 02, 2015

Impressive!

9. Jango

Posts: 372; Member since: Oct 24, 2014

Imagine if they incorporate this in the already amazing HDR+ they have on the Nexus 6P and the Pixel! Also, the wonders it can do if it were incorporated in the Google photos app, to upscale all those free unlimited images uploaded at not true resolution.

2. Macready

Posts: 1813; Member since: Dec 08, 2014

Without knowing the scale of enlargement and the technique used in the "oldschool" example, there's no telling how effective the machine learning method is. But judging by the lack of anti-aliasing in the former, they used a very crude way of enlarging. I bet using Lanczos filters, the difference would be minimal at best.

4. xondk

Posts: 1904; Member since: Mar 25, 2014

Generally a lot of manual correction after using filters and such. It doesn't seem crude, what you are seeing in the pictures are jpg artifacts from original image which it has also done, it is quite well done actually.

7. Macready

Posts: 1813; Member since: Dec 08, 2014

No, I was talking about the stairstepping in the "regular" enlargement. Using a good resize/resampling algorithm, none of that would have been present. In fact, If I downsize the machine learning example by a factor 4 and then resize it back to the size we see here, it looks almost identical to the machine learning example, much better than the "regular". Meaning, I'm not impressed by the example. Which is not to say that there are no benefits. It's just that they need to give us better examples and an explanation of the methods used, preferably providing the original file.

8. xondk

Posts: 1904; Member since: Mar 25, 2014

I think that's meant to show the low res, simply upsampled without anything, but yeah you are right, there's a lot of upsampling that works quite well, as long as source material is good enough quality.

3. Boybawang

Posts: 204; Member since: Jun 02, 2013

Does all the processing work in the clouds?

5. Zomer

Posts: 361; Member since: May 31, 2013

No, on servers.

6. Captain_Doug

Posts: 1037; Member since: Feb 10, 2012

ENHANCE

Latest Stories

This copy is for your personal, non-commercial use only. You can order presentation-ready copies for distribution to your colleagues, clients or customers at https://www.parsintl.com/phonearena or use the Reprints & Permissions tool that appears at the bottom of each web page. Visit https://www.parsintl.com/ for samples and additional information.