Google prototype uses machine learning to zoom and enhance photos, CSI-style

[sincity]

Low-resolution photos are fine for social media, but snaps captured with your phone are likely to look terrible if they’re blown up for printing.

Rapid and Accurate Image Super-Resolution (RAISR) is a new software prototype from Google that applies smart sharpening filters based on the photo’s characteristics to create larger, crisper images without unwanted artefacts.

RAISR can’t create data that isn’t in the original picture, so you’re unlikely to be able to

, but the results are still very impressive.

Sharper vision

Upscaling involves inserting extra pixels into a small image to create a larger one. Photo editors like Photoshop use various methods resampling methods to determine the color of these pixels (usually bicubic, bilinear or Lanczos interpolation). Whichever method you choose, the same filter is applied evenly to the entire image. The results often look blurry, or show unwanted artefacts.

RAISR works by sampling small parts of an image and choosing a filter that works on each pixel selectively to deliver the best possible results. These filters don’t use much processing power; the tricky part is knowing which one to use.

RAISR upscales images without creating moire patterns, jaggies or other artefacts

Google ‘trained’ RAISR using pairs of images – one low-resolution and one high-resolution. The software creates a new filter that most accurately recreates the high-res picture from the data in its low-res counterpart (a process that takes about an hour).

When presented with a new image, RAISR analyses its edge features, then selects and applies and most relevant filter from its collection.

RAISR isn’t available to download yet, but Google says its system requirements are modest enough for real-time photo enhancing on mobile devices.