Single Image Super Resolution Using AI Generated Images

Abstract

Image super-resolution has become increasingly important in various applications because of their demand for producing high output images from the low input images. Earlier for the image enhancements techniques like deblurring were performed to get the quality image. With the advancements in the Generative Adversarial Networks (GAN), the generating of high-quality image from the low-quality image has been outstanding. The models like SRGAN, ESRGAN [12]are the competitive models which make the Image-Resolution look good because of their performance on the images. But the architecture of the SRGAN which is a state-of-art model is complex and ESRGAN is built on the SRGAN, but by observing the results of the SRGAN the image quality looks good. We try to build a Super-Image Resolution by having the less complex architecture which is faster than SRGAN and the results aren’t compromising even after reducing the architecture complexity. We have built our base model based on the SRGAN by reducing the complexity in the architecture. In our final model we added another discriminator layer which enhances the sub parts of the images to improve the image quality. Our aim is to build an efficient model where the architecture of our model is less complex than SRGAN [14]and give as competitive results as SRGAN. Our results for the final model compared to our base model shows that there were significant improvements in the image quality. The code link for our project is here:https://github.com/faisalkhansk3283/ Computer_Vision_Extended_SRGAN