Research and Improvement of Single Image Super-Resolution Based on Generative Adversarial Network

Abstract

This paper proposes a new image super-resolution method based on Generative Adversarial Network (GAN). Firstly, the algorithm model includes generating model and discriminant model, generating model to generate high-resolution image, discriminating model to distinguish the image true or false, the original image is true, and the generated image is false. Using alternate training method, the generated model and discriminant model achieve Nash equilibrium, and finally generate high-quality image. Compared with previous super-resolution method based on generative adversarial network (SRGAN), the following changes have been made: modifying the network structure, removing the unnecessary batch normalization layer in the standard residual block, deepening the network layer number and improving the loss function. The experimental results show that compared with the traditional bicubic interpolation method and compared with SRGAN, the proposed algorithm improves the actual image effect, peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) in varying degrees.