We know that GAN easier to get on the line with better visual image when the image restoration, this article introduced today - ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks, which was published in Workshops ECCV 2018, the author of the SRGAN based on the improved, including the improvement of the structure of the network, in the form of decision of the decision unit, and replace the pre-trained network for calculating a perceptual domain loss.
Super-resolution generated against the network (SRGAN) is a pioneering work, can generate realistic textures in a single image super-resolution. This work was published in CVPR 2017.
However, details of the amplified usually accompanied by unpleasant artifacts. To further enhance the visual quality, the authors examine the three key parts SRGAN: 1) network structure; 2) antagonism loss; 3) loss of perception domain. And every improvement, get ESRGAN.
Specifically, the article proposes a Residual-in-Residual Dense Block (RRDB) network elements, in this unit, removing the BN (Batch Norm) layer. In addition, the authors draw Relativistic GAN idea, let discriminator prediction image of authenticity rather than "whether the image is a fake."
Finally, to improve the perception of domain loss, before using the feature activation, such as brightness consistency and texture can provide stronger oversight recovery. Under these improvements help, ESRGAN get a better visual quality and more realistic and natural textures.
FIG improved effect (four-times enlarged):