Abstract
Following the performance breakthrough of denoising networks, improvementshave come chiefly through novel architecture designs and increased depth. Whilenovel denoising networks were designed for real images coming from differentdistributions, or for specific applications, comparatively small improvementwas achieved on Gaussian denoising. The denoising solutions suffer fromepistemic uncertainty that can limit further advancements. This uncertainty istraditionally mitigated through different ensemble approaches. However, suchensembles are prohibitively costly with deep networks, which are already largein size. Our work focuses on pushing the performance limits of state-of-the-artmethods on Gaussian denoising. We propose a model-agnostic approach forreducing epistemic uncertainty while using only a single pretrained network. Weachieve this by tapping into the epistemic uncertainty through augmented andfrequency-manipulated images to obtain denoised images with varying error. Wepropose an ensemble method with two decoupled attention paths, over the pixeldomain and over that of our different manipulations, to learn the final fusion.Our results significantly improve over the state-of-the-art baselines andacross varying noise levels.