Image Super-Resolution using a Improved Generative Adversarial Network

Abstract

In recent years, there have been a variety of learning methods applied to single image super-resolution problems (SISR). Generative adversarial network (GAN) for image super-resolution which can infer photo-realistic natural images for 4× upscaling factors has been proposed. Images generated from SRGAN have sharper details, but some texture will be distorted and deformed. In this situation, the image looks unsatisfactory. In order to generate clearer, more eye-catching picture, we improved the SRGAN network. We proposed a encoder block in the generator to extract more crucial features. In this condition, we can generate more clear and natural image.