Researchers from NVIDIA, Aalto University and MIT developed a new deeplearning method with which they can remove a significant amount of noise from photos. To correct the photo, they only need the corrupted version.
The technology, called Noise2Noise, is a neural network which means that it had to be trained and fed with data to perform in unknown situations. Noise2Noise was fed with 50.000 clear and noise-free images from the ImageNet-database which were then covered with noise to train the algorithm. Except for noise, the system can also effectively remove subsequently added text.
In the available paper, researchers describe that the results of their system are comparable with other modern methods that do use a noise-free version of the same photo. The method could also be used to clear up mri-scan images and images in astrophotography.
