We present a novel approach for image completion that results in images that are both locally and globally consistent. In the popular adversarial network scheme, we train an image completion network as well as two---global and local---context discriminator networks. The image completion network is trained to fill-in the missing regions in the input image to fool the discriminators, which are in turn trained to distinguish real images from completed ones.

This presentation is part of Minisymposium “MS26 - New trends in inpainting”
organized by: Yann Gousseau (Telecom ParisTech) , Simon Masnou (Université Lyon 1) .

Authors:

Hiroshi Ishikawa (Waseda University)

Keywords:

computer graphics, computer vision, deep learning, image enhancement, image reconstruction