Super-Resolution Reconstruction Algorithm Based on Improved Generative Adversarial Network

Abstract

The current super-resolution algorithm for generative adversarial networks (SRGAN) has problems such as an unstable model training process and excessive smoothing of reconstructed images, which can affect the quality of generated images to a large extent. In this paper, based on SRGAN, all BN layers in the generative network are removed, and WGAN is used instead of JS scatter to optimize the discriminate network, This efficiently prevents the phenomenon of gradient disappearance and resolves the issue of unstable training of generative adversarial networks. The SA module is added to the vgg19 feature extraction network to obtain better feature information and improve the quality of the generated images. The experiments show that the proposed method has better stability in the training process compared with the traditional SRGAN on the DIV2K datasets, improvements are made to the reconstructed images' peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and visual effect performance.