Top: Original, Bottom: RAISR super-resolved 2x, Image: Google

Upsampling techniques to create larger versions of low-resolution images have been around for a long time – at least as long as TV detectives have been asking computers to 'enhance' images. Common linear methods fill in new pixels using simple and fixed combinations of nearby existing pixel values, but fail to increase image detail. The engineers at Google's research lab have now created a new way of upsampling images that achieves noticeably better results than the previously existing methods.

RAISR (Rapid and Accurate Image Super-Resolution) uses machine learning to train an algorithm using pairs of images, one low-resolution, the other with a high pixel count. RAISR creates filters that can recreate image detail that is comparable to the original, when applied to each pixel of a low-resolution image. Filters are trained according to edge features that are found in specific small areas of images, including edge direction, edge strength and how directional the egde is. The training process with a database of 10000 image pairs takes approximately an hour. 

Once RAISR has been trained it is capable of selecting the most appropriate filter for each pixel in a given low-resolution image to fill in new pixels in order to create a higher-resolution version. RAISR can also remove aliasing artifacts, such as moiré patterns or jagged lines in the low-resolution images when creating the larger version, something that linear methods are not capable of doing.

Eventually, Google may be able to use the technology to upsample images that are sent at mobile bandwidth-friendly resolutions. More information on RAISR is available on the Google Research Blog.

 Left: Original, Right: RAISR super-resolved 3x, Image. Google