Blind2Grad: Blind detail-preserving denoising via zero-shot with gradient regularized loss

Over the past few years, deep learning-based approaches have come to dominate the areas of image processing, with no exception to the field of image denoising. Recently, self-supervised denoising networks have garnered extensive attention in the zero-shot domain. Nevertheless, training the denoising network solely with zero-shot data will lead to significant detail loss, thereby influencing the denoising quality. In this work, we proposed a novel dual mask mechanism based on Euclidean distance selection as a global masker for the blind spot network, thereby enhancing image quality and mitigating image detail loss. Furthermore, we also constructed a recurrent framework, named Blind2Grad, which integrated the gradient-regularized loss and a uniform loss for training the denoiser and enabled the double-mask mechanism to be continuously updated. This framework was capable of directly acquiring information from the original noisy image, giving priority to the key information that needed to be retained without succumbing to equivalent mapping. Moreover, we conducted a thorough theoretical analysis of the convergence properties of our Blind2Grad loss from both probabilistic and game-theoretic viewpoints, demonstrating its consistency with supervised methods. Our Blind2Grad not only demonstrated outstanding performance on both synthetic and real-world datasets, but also exhibited significant efficacy in processing images with high noise levels.