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

Tong Li, Lizhi Wang, Zhiyuan Xu, Lin Zhu, Wanxuan Lu, Hua Huang

Abstract

Self-supervised single-image denoising is critical for real-world applications where clean ground-truth images are unavailable. However, existing methods often encounter an "information-lossy barrier" due to destructive operations like masking or downsampling. We introduce the Positive2Negative (P2N) paradigm, which leverages the statistical symmetry of noise to construct synthetic "positive" and "negative" noisy image pairs from a single observation. By enforcing consistency between the denoised outputs of these pairs, P2N preserves 100% of the original image information. Experimental results on benchmarks like SIDD and PolyU demonstrate that P2N achieves state-of-the-art performance with superior visual fidelity and remarkably fast convergence.

Introduction

Image denoising is a fundamental task in computational photography. While supervised deep learning has achieved impressive results, it requires pairs of noisy and clean images, which are difficult to obtain for many real-world scenarios. Self-supervised learning offers a solution by training models on noisy data alone.

Current self-supervised paradigms generally fall into two categories: Noise2Noise-based and Noise2Void-based. Noise2Noise methods often rely on downsampling to create training pairs, which reduces sampling density and introduces aliasing. Noise2Void methods use "blind-spot" masking to prevent the network from learning the identity mapping, but this hides crucial local information, leading to blurred textures and artifacts. P2N aims to break this barrier by utilizing all available pixels without destructive preprocessing.

Methodology: The Positive2Negative Paradigm

The P2N paradigm avoids information loss by exploiting the inherent symmetry of noise distributions. It consists of two primary stages:

1. Renoised Data Construction (RDC)

RDC utilizes the current state of the denoising network to estimate the noise component $n$ from the input image. Assuming real-world noise is roughly symmetrical and zero-centered, if $n$ is a valid noise sample, then $-n$ is also statistically plausible. P2N constructs a "Positive" noisy image by adding noise and a "Negative" noisy image by subtracting the estimated noise. Unlike masking or downsampling, these synthetic images maintain the original resolution and pixel density.

2. Denoised Consistency Supervision (DCS)

The model is trained using a consistency loss that forces the denoised versions of the Positive and Negative images to be identical. This supervision encourages the network to extract the underlying clean signal while ignoring the symmetric noise components. To enhance robustness across various noise distributions, a gradually varying norm ($\|\cdot\|_{2\rightarrow 1.5}$) is employed in the loss function.

Comparison of self-supervised denoising paradigms, highlighting the non-destructive nature of the Positive2Negative approach compared to Noise2Noise and Noise2Void methods.

Experimental Results

The authors evaluated P2N on several real-world denoising benchmarks, including the Smartphone Image Denoising Dataset (SIDD), Cross-Channel (CC), PolyU, and the Fluorescence Microscopy Denoising Dataset (FMDD).

Quantitative Performance

P2N consistently outperformed existing self-supervised single-image methods. On the SIDD benchmark, it achieved a PSNR improvement of over 0.4 dB compared to previous state-of-the-art techniques. Furthermore, P2N demonstrates exceptional training efficiency; while competing methods may require thousands of iterations, P2N can converge in fewer than 100 iterations.

Qualitative Analysis

Visually, P2N preserves fine details and sharp edges that are often lost by other methods. As shown in the visual comparisons, P2N avoids the blurring seen in downsampling-based methods (like MASH) and the artifacts associated with masking-based approaches (like ZS-N2N). It is particularly effective at reconstructing complex textures and legible text in low-light conditions.

Qualitative comparison of denoising results on real-world images. P2N (Ours) preserves sharper text and finer textures compared to ZS-N2N, MASH, and ATBSN.

Conclusion and Future Work

The Positive2Negative paradigm represents a significant step forward in self-supervised denoising by bypassing the information-lossy barrier. By using noise symmetry as a prior, it enables high-fidelity restoration without sacrificing spatial resolution. Currently, the method requires a pre-trained Gaussian model for initialization to avoid trivial solutions. Future research will focus on removing this dependency and extending the P2N logic to other image restoration tasks.

Summary updated

BibTeX
@inproceedings{li2025positive,
  author = {Tong Li and Lizhi Wang and Zhiyuan Xu and Lin Zhu and Wanxuan Lu and Hua Huang},
  title = {{Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.01670},
}
Figure