Preeti Sharma¹, Manoj Kumar² and Hitesh Kumar Sharma¹, ¹University of Petroleum and Energy Studies (UPES), India, ²University of Wollongong in Dubai, UAE

To produce super-resolution images, it is essential to eliminate the noise elements and givea clear noise-free output. To achieve this purpose multiscale image representation is found to be effective in many ways for its accuracy of correct feature extraction capacity. This denoising approach is integrated as a chosen enhancement tool in the form of an ensemble GAN model, and accordingly, the generator-discriminator training concept is transformed to adopt the approach as per the desired demands. In this research, a multiscale image approach is implemented using an ensemble GAN model with hybrid discriminator architecture. No one form of noise is "ideal" to eliminate while denoising with the proposed model. Instead, based on the properties of the data and the noise inherent in it, the proposed ensemble GAN can handle various sorts of noise. The technique optimises training through simultaneous generator and discriminator model updates, improving output quality, by using the least loss value for discriminator selection. Inception Score (IS) and Frechet Inception Distance (FID) evaluations show that it outperforms pixel-based denoising, with an amazingaccuracy of 99.91%

GAN, multiscale image representation, ensemble GAN, pixel based denoising, Multiscale denoising.