Image Super-resolution Reconstruction Based on an Improved Generative Adversarial Network

Abstract

To solve the problem that images reconstructed by traditional super-resolution reconstruction (SR) techniques are smooth and lack good details, in this paper, we have presented an improved generative adversarial network for image super-resolution. The improved method was based on deep neural networks whose generative model contained a multi-layer convolution module and multi-layer deconvolution module, in which a layer hopping connection and a loss function was added to the perceptual loss. The discriminant model was made up of a multi-layer neural network whose loss function was based on the discriminant model loss function that was generated from the generative adversarial network. Finally, we selected PSNR and SSIM as the indicator in the experiments. In the experiments, the PSNR value of 2x, 3x and 4x magnification factor are improved on average by 1.125, 2.175 and 2.075 respectively and the SSIM value of 2x, 3x and 4x magnification factor are basically improved. Compared with the existing image super-resolution reconstruction methods, the effectiveness of the method that we proposed in image super-resolution reconstruction was proven.