Resolution enhancement in the recovery of underdrawings via style transfer by generative adversarial deep neural networks

We apply generative adversarial convolutional neural networks to the problem\nof style transfer to underdrawings and ghost-images in x-rays of fine art\npaintings with a special focus on enhancing their spatial resolution. We build\nupon a neural architecture developed for the related problem of synthesizing\nhigh-resolution photo-realistic image from semantic label maps. Our neural\narchitecture achieves high resolution through a hierarchy of generators and\ndiscriminator sub-networks, working throughout a range of spatial resolutions.\nThis coarse-to-fine generator architecture can increase the effective\nresolution by a factor of eight in each spatial direction, or an overall\nincrease in number of pixels by a factor of 64. We also show that even just a\nfew examples of human-generated image segmentations can greatly improve --\nqualitatively and quantitatively -- the generated images. We demonstrate our\nmethod on works such as Leonardo's Madonna of the carnation and the\nunderdrawing in his Virgin of the rocks, which pose several special problems in\nstyle transfer, including the paucity of representative works from which to\nlearn and transfer style information.\n