Deep Image Prior for Super Resolution of Noisy Image

Single image super-resolution task aims to reconstruct a high-resolution from low-resolution image. Recently, it has been shown that by using deep prior (DIP), single neural network is sufficient capture low-level statistics only without data-driven training such can be used for various restoration problems. However, tasks are difficult perform with DIP when the target noisy. The super-resolved becomes noisy because reconstruction loss of does not consider noise in Furthermore, contains noise, optimization process unstable and sensitive noise. In this paper, we propose noise-robust stable framework based on DIP. To end, noise-estimation method generative adversarial (GAN) self-supervision (SSL). We show generator learn distribution proposed framework. Moreover, argue stabilized incorporated. experiments quantitatively qualitatively outperforms existing methods images.