Sharp and Real Image Super-Resolution Using Generative Adversarial Network

Recent studies have achieved great progress on accuracy and speed of single image super-resolution (SISR) based on neural networks. Most current SISR methods use mean squared error (MSE) loss as objective function. As a result, they can get high peak signal-to-noise ratios (PSNR) which are however not in full agreement with the visual qualities by experiments, and thus the output from these methods could be prone to blurry and over-smoothed. Especially at large upscaling factors, the output images are perceptually unsatisfactory in general. In this paper, we firstly propose a novel residual network architecture based on generative adversarial network (GAN) for image super-resolution (SR), which is capable of inferring photo-realistic images for 4x upscaling factors. Perceptual loss is applied as the objective function to make output image sharper and more real. In addition, we adopt some tricks to preprocess the input dataset and use improved techniques to train the generator and discriminator separately, which are proved to be effective for the result. We validate our GAN-based approach on CelebA dataset with mean opinion score (MOS) as performance measure. The results demonstrate that the proposed approach performs better than previous methods.