Multi-loss Super-Resolution Generative Adversarial Network
Jian Xu, Chunmeng He · 2023
Single-image super-resolution (SISR) based on deep neural networks has achieved excellent performance in recent years. However, how to recover texture details is still a challenging problem in the field of super-resolution. In this paper, in order to obtain reconstructed images with natural and realistic textures, we propose a network structure and a super-resolution reconstruction algorithm combining variational autoencoder (VAE) with generative adversarial network (GAN). This paper proposes optimized variational autoencoder (OVAE) as a discriminator to predict relative realness. The discriminator could project the input images into a latent space. And the latent space in OVAE could learn the latent distribution of high-resolution images. By learning the latent distribution of the input images, the discriminator can obtain high-frequency features and contextual information. In addition, this paper uses Kullback-Leibler (KL) divergence to optimize the discriminator and constrain the data distribution. Finally, to achieve stable training, we improve GAN by using Wasserstein divergence for GANs (WGAN-div). The network structure acquires a sophisticated progressive growing training scheme and consistently achieves better visual quality. Experimental results have proven that the method in this paper outperforms several state-of-the-art methods in both objective and subjective evaluation, and the proposed algorithm can efficiently reconstruct high-resolution images with natural and realistic textures.