Dual Discriminator Generative Adversarial Network for Single Image Super-Resolution

Abstract

Recently, several algorithms have been proposed to achieve the single image super-resolution by using deep convolutional neural networks. In this study, we present a dual discrimination generative adversarial network (D2GAN) for single image super-resolution (SISR). The proposed model has better stability to complete the reconstruction of super-resolution images for ×4 scale factor. The improved residual network and perceptual loss function are applied in the proposed algorithm which demonstrates a superior performance over state-of-the-art restoration quality. Meanwhile, the proposed reconstruction network has a faster training and convergence speed compared with other super-resolution methods. The proposed approach is evaluated on standard datasets and gets improved performance than previous works that based on deep convolutional neural networks.