A Novel Fast Reconstruction Method for Single Image Super Resolution Task

The in-depth development of generative adversarial networks in the field of specific application tasks, the single image super-resolution problem has been widely studied. We proposed a novel reconstruction model using generative adversarial network for single image super-resolution reconstruction task. Our model directly learns an end-to-end mapping between the low and high-resolution images. The mapping is represented as a generative adversarial network that takes the low-resolution images as the input and outputs the high-resolution one. We further confirmed that the spectral normalization and attention method is exceedingly effective for the training of stably generating adversarial networks. We use a more lightweight external attention method in the network to accelerate the speed of global structure reconstruction. It is different from the previous image super-resolution task. Our model implements multi-scale joint training and optimizes all layers. We explore different networks structures and parameter settings to achieve trade-offs between performance and speed. Quantitative indexes and qualitative results show that our proposed method achieves comparable performance with the state-of-the-art supervised models.