D ECAY 2D ISTILL :L EVERAGING SPATIAL PERTURBATION AND REGULARIZATION FOR SELF - SUPERVISED IMAGE DENOISING Manisha Das Chaity Rochester Institute of Technology NY, USA manisha.kuet@gmail.com Masud An Nur Islam Fahim Chosun University Gwangju, South Korea mostofafahim21@gmail.com February 11, 2023 ABSTRACT Unpaired image denoising has achieved promising development over the last few years. Regardless of the performance, methods tend to heavily rely on underlying noise properties or any assumption which is not always practical. Alternatively, if we can ground the problem from a structural perspective rather than noise statistics, we can achieve a more robust solution. with such motivation, we propose a self-supervised denoising scheme that is unpaired and relies on spatial degradation followed by a regularized refinement. Our method shows considerable improvement over previous methods and exhibited consistent performance over different data domains. 1 Introduction Image denoising is an ever-going issue that confirms its presence in most vision problems. Anyhow, recovery of noisy observation has attracted considerable interest from the researcher and consequently, we have observed a bulk of studies, mostly dominated by the supervised approaches. Since the existence of in-domain clear observation is mostly impractical/costly, the current trend in signal recovery is dominated by self-supervised approaches. Typically we minimize ‖ f (x) − y ‖ 2 in a supervised setup, where y is the available ground truth image, f is the given estimator and x is the noisy observation. In the self-supervised setup, we minimize ‖ f (x) − x ‖ 2 with a respective novel empirical risk minimization process. The whole development roughly started right after this ’weakly/noisy’ supervised study Noise2Noise (N2N) [1], where authors obtained supervised baseline results without using a clean target. Later, Noisier2Noise (Nr2N) [2] achieves good performance just by doubling the input noise, although it’s a clever modification of the N2N. Later, Noise2Self (N2S) [3] introduces a masking strategy to address the unpaired denoising problem, followed by similar clever alternatives Noise2Void (N2V) [4], Noise2Same [5] (N2Sa), Blind2Unblind (B2Ub) [6]. Nevertheless, overall unpaired or self-paired denoising studies follow two sects in terms of algorithmic strategy [7] : (a) signal augmentation arXiv:2208.01948v2 [cs.CV] 4 Aug 2022