Despite the breakthroughs in accuracy and speed of single imagesuper-resolution using faster and deeper convolutional neural networks, onecentral problem remains largely unsolved: how do we recover the finer texturedetails when we super-resolve at large upscaling factors? The behavior ofoptimization-based super-resolution methods is principally driven by the choiceof the objective function. Recent work has largely focused on minimizing themean squared reconstruction error. The resulting estimates have high peaksignal-to-noise ratios, but they are often lacking high-frequency details andare perceptually unsatisfying in the sense that they fail to match the fidelityexpected at the higher resolution. In this paper, we present SRGAN, agenerative adversarial network (GAN) for image super-resolution (SR). To ourknowledge, it is the first framework capable of inferring photo-realisticnatural images for 4x upscaling factors. To achieve this, we propose aperceptual loss function which consists of an adversarial loss and a contentloss. The adversarial loss pushes our solution to the natural image manifoldusing a discriminator network that is trained to differentiate between thesuper-resolved images and original photo-realistic images. In addition, we usea content loss motivated by perceptual similarity instead of similarity inpixel space. Our deep residual network is able to recover photo-realistictextures from heavily downsampled images on public benchmarks. An extensivemean-opinion-score (MOS) test shows hugely significant gains in perceptualquality using SRGAN. The MOS scores obtained with SRGAN are closer to those ofthe original high-resolution images than to those obtained with anystate-of-the-art method.
展开▼
