In this paper, against generation network (GAN) on the super-resolution image, and uses the new loss function to obtain a result having a realistic visual effects.
Restore detail of the image in the process of super-resolution is a serious problem, many use MSE as a function of the job losses have a high PSNR, but they recovered the result is a lack of high-frequency information and the visual effect is not very satisfaction.
In order to have high resolution fidelity fit, proposed SRGAN generates realistic visual effects HR picture, in SRGAN, the network uses the generated SRResNet, and uses the new loss function, i.e., loss of visual function (Perceptual loss). Vision loss function consists of two parts, against loss (adversarial loss) and the loss of content (content loss). Against loss is a super-resolution image and the difference image generated by the original, but the content is lost rather than by the visual similarity of similarity defined in the pixel domain.
Content-based MSE loss function considering two kinds of losses and network-based VGG in this article. Using the MSE will be high PSNR image, but lacks high frequency content and the texture smooth transition; networks VGG employed PSNR loss although not as MSE obtained, but the visual effect is very good.
In this paper, respectively, for the three cases did a test, SRGAN-MSE is MSE as a content loss function, SRGAN-VGG22 low-level features are defined in the content network VGG loss function, and SRGAN-VGG54 is defined in the deep VGG content loss function on the advanced features of the network, it has a greater potential to focus on the content of the image. The final results show that the effect of SRGAN-VGG54 is the best.
In general, this method of generating confrontation by the network, using the new visual loss function, so that the final picture generated by HR have better texture detail and more comfortable visual effects, while avoiding lead to the loss as a function of MSE high-frequency image information deletion, and over-smoothing.