If you’ve ever taken photos after sundown, you’re probably familiar with the ‘noise’ that distorts your pictures with colored specks. It’s possible to fix your images to some degree with software – but an AI developed by Nvidia, MIT and Aalto University can go much further – without ever having seen noise-free photographs.
The Noise2Noise AI was trained using 50,000 pictures – as well as MRI scans and computer-generated images -that had randomized noise added to them. In their paper, the researchers show that their AI can successfully remove enough noise to make the pictures usable again, with detail and clarity that’s remarkably close to the source images.
The AI could likely find use in software for cleaning up noisy photos captured in low-light conditions on phones and cameras, as well as improving astronomical imagery and MRI scanning.
That’s similar to this neural network developed by researchers from Intel and the University of Illinois at Urbana-Champaign, which brings details back to extremely dark images. However, it may be better suited to real-time image processing for night-vision.
The Noise2Noise paper will be presented at the Thirty-fifth International Conference on Machine Learning in Sweden this week; you can find out more about the event here, and read the paper here (PDF).
If you’re into this, you’ll want to check out our other stories about cool AI for imaging, like this one that fixes low-res photos, this one from Google that edits Street View images like a pro photographer, this one that can unblur pictures, and this one that can reconstruct partially erased pictures.