Humans can analyze the relationship of data collected from different domains without any supervision, but automatic learning of discovering data is a challenging task. Image processing is one of the examples where conversion from day to night and night to day is difficult. To avoid drawbacks in image processing, Generative adversarial network (GAN) came into existence, which helps to generate networks without losing properties of attributes such as face identity and orientation. Again, super-resolution generative adversarial network (SRGAN) is a formative task that is capable of producing realistic textures in single image super-resolution. Still, hallucinated descriptions are generally attained with no desirable artifacts. Furthermore, to improve visual quality, we significantly studied the prime components of SRGAN—architecture, perceptual loss, adversarial loss, and training algorithm. Moreover, various potential open challenges, along with the future scope, are also discussed.

Super-resolution-based GAN for image processing: Recent advances and future trends

The Resnet model is similar to the ensemble, and its performance and parameters can be considered according to the modular design. Currently, Resnet is widely used as a backbone network. In particular, the Resnet module that compensates the weight can consider the similarity of pixels. Therefore, in this paper, we propose a method to increase the similarity between pixels by performing the operation of the Resnet module which has an effect similar to the ensemble operation. It give us a better high resolution image.

Wangs-Su Jeon

Papers

Single Image Super Resolution using Residual Learning