A neural network for digital cleaning of an oil painting

Abstract

We demonstrate that a neural network can be trained to learn the transformation from dirty to clean segments of a painting. The inputs to the network are pixels belonging to dirty paint segments and the desired outputs are pixels from clean segments. We find that the transformation from dirty to clean portions is nonlinear which is contrary to the assumption of most of the previous works on digital cleaning. Finally, we use the neural network to virtually clean a digital image of Fernando Amorsolo’s Malacañang by the River.