SRWGANTV: Image Super-Resolution Through Wasserstein Generative Adversarial Networks with Total Variational Regularization

Abstract

The study of generative adversarial networks (GAN) has enormously promoted the research work on single image super-resolution (SISR) problem. SRGAN firstly apply GAN to SISR reconstruction, which has achieved good results. However, SRGAN sacrifices the fidelity. At the same time, it is well known that the GANs are difficult to train and the improper training fails the SISR results easily. Recently, Wasserstein Generative Adversarial Network with gradient penalty (WGAN-GP) has been proposed to alleviate these issues at the expense of performance of the model with a relatively simple training process. However, we find that applying WGAN-GP to SISR still suffers from training instability, leading to failure to obtain a good SR result. To address this problem, we present an image super resolution framework base on enhanced WGAN (SRWGAN-TV). We introduce the total variational (TV) regularization term into the loss function of WGAN. The total variational (TV) regularization term can stabilize the network training and improve the quality of generated images. Experimental results on public datasets show that the proposed method achieves superior performance in both quantitative and qualitative measurements.