Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising

TL;DR

This work tackles self-supervised denoising from a single noisy image by addressing the information-lossy barrier of prior paradigms. It introduces Positive2Negative, a two-step framework comprising Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS), which together create multi-scale, information-preserving noisy data and enforce consistent denoising across these observations. The authors provide theoretical and empirical support showing that learning robust denoising is possible under zero-mean, approximately symmetric noise, achieving state-of-the-art performance among self-supervised single-image methods and notable efficiency gains. The approach generalizes across standard benchmarks (SIDD, CC, PolyU, FMDD) and offers a public implementation, facilitating practical adoption for image restoration tasks where only a single noisy image is available.

Abstract

Image denoising enhances image quality, serving as a foundational technique across various computational photography applications. The obstacle to clean image acquisition in real scenarios necessitates the development of self-supervised image denoising methods only depending on noisy images, especially a single noisy image. Existing self-supervised image denoising paradigms (Noise2Noise and Noise2Void) rely heavily on information-lossy operations, such as downsampling and masking, culminating in low quality denoising performance. In this paper, we propose a novel self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier. Our paradigm involves two key steps: Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS). RDC renoises the predicted denoised image by the predicted noise to construct multiple noisy images, preserving all the information of the original image. DCS ensures consistency across the multiple denoised images, supervising the network to learn robust denoising. Our Positive2Negative paradigm achieves state-of-the-art performance in self-supervised single image denoising with significant speed improvements. The code is released to the public at https://github.com/Li-Tong-621/P2N.

Figures (8)

• Figure 1: Information-lossy operations in current paradigms. The noise addition operation reduces the signal-to-noise ratio, the downsampling operation diminishes the sampling density, and the masking operation disregards crucial central pixels. Each of these operations introduces an information-lossy barrier ATBSNNyquistShannon, culminating in imprecise denoising.
• Figure 2: The information-lossy barrier significantly compromises the image quality, resulting in residual noise, aliasing effects or texture loss.
• Figure 3: Noise distribution is zero-mean and approximately symmetrical. (a) shows a noisy image. (b) and (c) show the noise distributions, which are calculated at the center pixels marked by the green circle and the red circle in the noisy image, respectively. It is evident that the noise distribution is zero-mean and approximately symmetrical.
• Figure 4: The opposite noise $-n$ is similar to the original noise $n$. The opposite noise $-n$ approximately follows the original noise's distribution, which is zero-mean and approximately symmetrical.
• Figure 5: Overview of the proposed self-supervised single image denoising paradigm. For training, the Positive2Negative paradigm consists of 2 steps. The first step is Renoised Data Construction (RDC): multiply the predicted noise by positive and negative coefficients to construct multi-scale Positive noise and Negative noise, respectively. The second step is Denoised Consistency Supervision (DCS): train the network through supervising the consistency of the denoised images. For inference, denoising can be achieved in just one pass through the network. (The networks shown in the figure are the same one. In addition, for better visualization, the images within the red boxes on the bottom left have been zoomed in and are displayed in the red boxes on the bottom right.)