Single Image Super-Resolution Based on Improved WGAN
Lei Yu, Xiang Long, Chao Tong · 2018
RGAN has successfully applied the Generative Adversarial Network to the single image super-resolution reconstruction, which has achieved good results.But the loss function based on feature space in S RGAN objectively sacrifices the pursuit of high peak signal-to-noise-ratio (PS NR), which is the result of a tradeoff.At the same time, Improved Training of Wasserstein GANs makes the training process more stable.We redesign the S RGAN, using VGG16 network for feature extraction, setting discriminator network's working space as feature space, and adding the loss function based on the mean square error of pixel space, then gain more details and high PS NR in the reconstruction at the same time.We use the design of WGAN-GP for reference to make the training more stable.