An Improved Super-Resolution Generative Adversarial Network

Abstract

An improved SRGAN image generation model proposed to address the increasing demand for high-resolution images and the problem of model gradient disappearance due to the excessive amount of parameters and deeper convolution layers in SRGAN for reconstructing high-resolution images. First, used a null convolution residual block model in the SRGAN discriminator to alleviate the gradient disappearance; second, the CBAM attention module adds to the discriminator as a way to extract image features further; finally, adaptive averaging pooling is added to the discriminator to reduce the number of model parameters. The experimental results show that the improved SRGAN reconstructed images are evaluated in the standard datasets AID and RSOD, reaching 29.58 and 27.37 in peak signal-to-noise ratio (PSNR), respectively, and 0.86 and 0.84 in structural similarity (SSIM), respectively.