Nvidia uses artificial intelligence to remove noise from photos

Researchers from Nvidia, Finland’s Aalto University and MIT demonstrate a deep learning method that can remove significant amounts of digital noise from photos. The AI ​​only needs the photos with noise for that.

The applied system is called Noise2Noise and can eliminate an extreme amount of noise from a photo, without the used photo ever being shown. The system is a neural network and that means that it had to be trained a lot first. To this end, Noise2Noise was presented with 50,000 clear, noise-free images from the ImageNet database. Then all these photos were digitally noised and used to train the algorithm. Nvidia Tesla P100 GPUs were used in combination with TensorFlow. The system is not only able to effectively remove noise, but also text pasted in a photo.

The researchers describe in their paper that the results of their system are comparable to those of modern methods that also use a noise-free variant of the same photo. The research from Nvidia and the two universities will be presented this week in Stockholm at the International Conference on Machine Learning. According to the researchers, the method can be used for cleaning up MRI scan images and for photography in dark conditions, such as astrophotography.