A team of specialists from NVIDIA, the Massachusetts Institute of Technology, and the University of Aalto presented a neural network capable of editing images and removing unnecessary noise from them, as well as various inscriptions. NVIDIA specialists shared their work on their official blog, but they did not put the source code of the algorithm in open access.
To use the method of computer-based training, specialists used the machine learning method. Based on 50 different spoiled photos, the algorithm has learned to remove noise from images caused by poor lighting, color noise in old pictures, and to get rid of inscriptions, signatures and, possibly, even copyright, available on images. The developers note that the neural network for effective operation does not need to see the same, but a clean photo – it is enough only to show poor-quality pictures.
If the NVIDIA algorithm shown in April demonstrated the possibility of restoring on clean photos with faces missing parts, as well as the ability to remove unnecessary details from them, then the current algorithm is able to conduct all work based on “damaged data.”
“You can teach neural network restore parts of the photo so that the algorithm does not even see its original value. This does not affect the performance of the computer, as well as the processing time does not increase, “say the developers.
The developers note that the method can be used not only for processing art photographs, but also for removing noise from images of MRI, other medical images, as well as when obtaining photographs with always low illumination, for example, when performing astronomical observations.