Discriminative Region Proposal Adversarial Networks for High-Quality Image-to-Image Translation
share
Image-to-image translation has been made much progress with embracing Generative Adversarial Networks (GANs). However, it's still very challenging for translation tasks that require high-quality, especially at high-resolution and photo-reality. In this paper, we present Discriminative Region Proposal Adversarial Networks (DRPANs) with three components: a generator, a discriminator and a reviser, for high-quality image-to-image translation. To reduce the artifacts and blur problems while translation, based on GAN, we explore a patch discriminator to find and extract the most artificial image patch by sliding the output score map with a window, and map the proposed image patch onto the synthesized fake image as our discriminative region. We then mask the corresponding real image using the discriminative region to obtain a fake-mask real image. For providing constructive revisions to generator, we propose a reviser for GANs to distinguish the real from the fake-mask real for producing realistic details and serve as auxiliaries to generate high-quality translation results. Experiments on a variety of image-to-image translation tasks and datasets validate that our method outperforms state-of-the-art translation methods for producing high-quality translation results in terms of both human perceptual studies and automatic quantitative measures.
READ FULL TEXT