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 of style transfer underdrawings and ghost-images in x-rays fine art paintings with a special focus on enhancing their spatial resolution. build upon architecture developed for related synthesizing high-resolution photo-realistic image from semantic label maps. Our achieves high resolution through hierarchy generators discriminator sub-networks, working throughout range resolutions. This coarse-to-fine generator can increase effective by factor eight each direction, or an overall number pixels 64. also show that even just few examples human-generated segmentations greatly improve—qualitatively quantitatively—the generated images. demonstrate our method works such as Leonardo’s Madonna carnation underdrawing his Virgin rocks , which pose several problems transfer, including paucity representative learn information.