Multi-scale generative adversarial network for image super-resolution

Abstract

In recent years, deep convolutional neural networks (CNNs) have been widely employed in image super-resolution. Thanks to the power of deep CNNs, the reconstruction performance is largely improved. However, the high-frequency information and details in the low-resolution image still can hardly be reconstructed. To deal with the above problems, we propose a multi-scale generative adversarial network in this paper. The multi-scale Pyramid module inside the generator could extract the features containing high-frequency information, and then the high-resolution image with the results of the bicubic interpolations is reconstructed. The discriminator in our model is used to identify the authenticity of the input image after refactoring. Our final loss function includes an adversarial loss and the mean square error (L2) reconstruction loss. In order to further improve the efficiency of training, the generator is pre-trained with the L2 loss, so as to improve the efficiency of the discriminator optimization. Compared with the algorithms based solely on normal plain convolutional networks, the proposed algorithm performs better in two indexes PSNR and SSIM of the super-resolution task.