This AI system can fix your grainy, low-light photos
Researchers from Nvidia, MIT, and Aalto University have developed an artificial intelligence (AI) system called Noise2Noise that can reduce noise in photos. The AI can fix a noisy image without ever having seen the noise-free version of that photograph. The researchers will present their paper at the 35th International Conference on Machine Learning in Stockholm, Sweden this week.
"Noise" on images refers to distortion or the grainy element in a photo. Images captured in low-light conditions on phones often develop colored specks or noise on them. Noise is also found in other forms of photos like medical imagery and pictures of space. Images clicked using a digital zoom often contain noise as well.
The AI was trained using 50,000 pictures from the ImageNet dataset, as well as MRI scans and computer-generated images. These images had randomized noise added to them and the AI eventually learned to reconstruct and de-noise them. The researchers claim that the AI could de-noise a photo with remarkable detail and clarity and make it almost like its source image, and thus, usable again.
Researchers hope that Noise2Noise can be used on noise-prone images like astrophotography, pictures clicked during sundown, and MRI or brain scans. It can also be used by present-day noise-reduction software for better and more accurate results that are driven by deep learning.