Zero-shot blind image denoising via implicit neural representations

Abstract

Recent literature focuses on image denoising without the prerequisite of external data. It is considered a more interesting phenomenon to excavate underlying natural features under noisy signals given only a single image. Nevertheless, existing zero-shot methods rely on the blind spot strategy, weakening low- noise and real noise denoising performance. Aiming to develop a more powerful and effective denoiser in this setting, we propose to leverage the inductive bias of Implicit Neural Representations(INR). We observed the highest frequency bias within INR when comparing the output precision and learning curve to other architectures. Furthermore, we examined a hierarchical structure of earlier layers learning the global features and deeper layers predicting the fine local details. In order to penalize the latter layers from identical fitting of noise, we applied a simple regularization to boost the denoising performance. Extensive experiments show that the implicit neural representations outperform existing zero-shot based methods, with the most stable and consistent denoising performance.