NVIDIA has learned neural network to remove excess noise and labels with photos

A team from NVIDIA, Massachusetts tehnologitcheskogo Institute and Aalto University have presented a neural network that can edit images and to remove them from the excess noise, as well as various inscriptions. On the work done NVIDIA experts shared in its official blog, but the source code of the algorithm did not spread in open access.

to create a neural network experts used a machine learning method. On the basis of 50 different spoiled pictures, the algorithm has learned to remove from the image noise caused by low light, chrominance noise in the old photographs, as well as get rid of labels, labels, and perhaps even a copyright, available in the images. The developers note that the neural network to work effectively not need to see the same, but a clean image - just to show low-quality images.

If shown in April NVIDIA algorithm demonstrated the ability to restore to clean photographs depicting persons missing parts, and the ability to clean them with unnecessary details, the current algorithm is able to carry out all the work on the basis of "corrupted data".

"You can teach a neural network to restore part of the photo that the algorithm did not even see its original value. If it does not suffer from the performance of the computer, nor does it increase the time to handle, "- say the developers.

The developers note that the method can be used not only in the treatments of artistic photographs, but also for removing noise from the MRI images and other medical imaging, as well as in the preparation of photographs is always a low level of illumination, for example, when the astronomical observations.