Tarsier: Evolving Noise Injection in Super-Resolution GANs
This work addresses enhancing image super-resolution quality for applications like media and computer vision, but it is incremental as it builds directly on an existing method.
The paper tackles the problem of improving super-resolution GANs by optimizing noise injection at inference time using evolutionary methods, resulting in outperforming the state-of-the-art NESRGAN+ on standard datasets as validated by perceptual scores and human studies.
Super-resolution aims at increasing the resolution and level of detail within an image. The current state of the art in general single-image super-resolution is held by NESRGAN+, which injects a Gaussian noise after each residual layer at training time. In this paper, we harness evolutionary methods to improve NESRGAN+ by optimizing the noise injection at inference time. More precisely, we use Diagonal CMA to optimize the injected noise according to a novel criterion combining quality assessment and realism. Our results are validated by the PIRM perceptual score and a human study. Our method outperforms NESRGAN+ on several standard super-resolution datasets. More generally, our approach can be used to optimize any method based on noise injection.