Semi-Supervised Image Super-Resolution Based on Domain Transformation

For now, most Single Image Super-Resolution(SISR) tasks are based on fully supervised image pairs of high and low pairs (obtained from fixed downsampling on high-resolution images). However, the fixed degradation process can not simulate the real image degradation well, making the model ineffective in reconstructing real images. In this paper, a semi-supervised reconstruction network based on domain transformation is proposed, the network consists of a domain transformation network and a reconstruction network, the real low-resolution image first passes through the domain transformation network for domain distribution adjustment, and then the reconstruction network generates the matching high-resolution image. In the network, we introduce an attention module, use perceptual loss, and use a 1*1 parallel convolution in the reconstruction part to increase the network nonlinearity, which in turn improves the image reconstruction quality. Experiments demonstrate the effectiveness of our method on single real image reconstruction both qualitatively and quantitatively.