Improved Generative Adversarial Network for Generating High-Resolution Images from Low-Resolution Images

Abstract

The procedure of generating a high-resolution (HR) image from a low-resolution (LR) image is known as super-resolution (SR). In this paper, we try to perform SR using Deep Learning techniques. For better performance in medicinal imaging, forensics, pattern recognition, satellite imaging, surveillance, etc., zooming of a particular area of attention in the image is required, making high resolution necessary. We present ISRGAN (Improved Super Resolution Generative Adversarial Network), an improved version of SRGAN (Super Resolution Generative Adversarial Network) for image SR. For training the network, images from the DIV2K dataset have been used. The dataset consists of 800 different images which have been resized to 32x32 and 128x128 pixels to form LR and HR images respectively. A Generative Adversarial Network is based on the idea of Adversarial training, and comprises of 2 parts, a discriminator network, and a generator network. The generator produces a HR image from the LR image, whereas the discriminator network classifies the generated image as fake or real. The presented model produces decent results and our final super-resolved results show that the presented ISRGAN model produces images with enhanced features.