Nvidia's new AI takes a one-stop approach to fixing grainy photos
If you've ever taken a photo in low light you've probably encountered the grainy effect that can dilute the finished product. A new AI tool could prove an incredibly easy way to remove this so-called noise, with the ability to automatically produce a clean image after analyzing only the corrupted version.
The AI was built by researchers from Nvidia, MIT and Finland's Aalto University, who set out to improve on recent work in the field of machine learning and image processing. This has led to neural networks that can scan a noisy image along with a clean version, and then use machine learning to bridge the gap.
Their solution, dubbed Noise2Noise, is capable of producing a noise-free image without ever seeing an original clean version, removing artifacts, noise and grain automatically. The team developed the tool by training it on a catalog of 50,000 images, and tested it out on three different datasets comprising images of different kinds.
"It is possible to learn to restore signals without ever observing clean ones, at performance sometimes exceeding training using clean exemplars," the researchers stated in their paper. "The neural network is on par with state-of-the-art methods that make use of clean examples – using precisely the same training methodology, and often without appreciable drawbacks in training time or performance."
While this kind of tool could one day make Instagram feeds a lot prettier, in the realm of science it could also have quite a meaningful impact. Medical images taken from by MRI and space images taken with scientific instruments are two typically grainy examples that could benefit from this type of technology.
"There are several real-world situations where obtaining clean training data is difficult: low-light photography (e.g. astronomical imaging), physically-based rendering, and magnetic resonance imaging," the team says. "Our proof-of-concept demonstrations point the way to significant potential benefits in these applications by removing the need for potentially strenuous collection of clean data. Of course, there is no free lunch – we cannot learn to pick up features that are not there in the input data – but this applies equally to training with clean targets."
A research paper describing the technology is available online, with the work to be presented at the International Conference on Machine Learning in Stockholm, Sweden, this week. The video below provides a few more examples of it in action.