ISRGAN: Improved Super-Resolution Using Generative Adversarial Networks

Abstract

In this paper, we propose an approach for single image super-resolution (SISR) using generative adversarial network (GAN). The SISR has been an attractive research topic over the last two decades and it refers to the reconstruction of a high resolution (HR) image from a single low resolution (LR) observation. Recently, SISR using convolutional neural networks (CNNs) obtained remarkable performance in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) metrics. Despite this, these methods suffer with a serious drawback in terms of visualization quality of the SR images; the results look overly-smoothed. This is due to the loss function in those methods has a pixel level difference which increases the values of PSNR and SSIM metrics; however the visualization quality is degraded. The GAN has a capability to generate visually appealable solutions. It can also recover the high-frequency texture details due to the discrimination process involved in GAN. Here, we propose improved single image super-resolution using GAN (ISRGAN) with the concept of densely connected deep convolutional networks for image super-resolution. Our proposed method consists two networks: ISRNet and ISRGAN. The ISRNet is trained using MSE based loss function to achieve higher PSNR and SSIM values and ISRGAN is trained by using a combination of VGG based perceptual loss and adversarial loss in order to improve the perceptual quality of the SR images. This training step forces the SR results more towards the natural image manifold. The efficiency of the proposed method is verified by conducting experiments on the different benchmark testing datasets and it shows that the proposed method of ISRGAN outperforms in terms of perception when compared to the other state-of-the-art GAN based SISR techniques.