Abstract
Self-supervised single-image denoising is a critical task in computational photography, particularly when ground-truth clean images are unavailable. Traditional methods often rely on operations like downsampling or masking, which create an "information-lossy barrier" resulting in aliasing or blurring. We present Positive2Negative (P2N), a novel paradigm that leverages noise symmetry to construct "Positive" and "Negative" renoised image pairs. By preserving full spatial information, P2N achieves state-of-the-art performance on benchmarks such as SIDD and FMDD, while demonstrating significantly faster convergence than existing approaches.
The Information-Lossy Barrier
Real-world image denoising is often hampered by the impossibility of obtaining clean ground-truth images for training. This has led to the development of self-supervised paradigms, which fall into two main categories:
- Noise2Noise-based paradigms: These typically require downsampling the image to create training pairs, which causes aliasing and a loss of spatial resolution.
- Noise2Void-based paradigms: These use blind-spot networks or masking to hide the central pixel's information, resulting in blurry outputs and residual noise.
The authors argue that these operations—downsampling, masking, and adding extra noise—inherently limit the quality of the restored image by discarding critical spatial data.
The Positive2Negative (P2N) Paradigm
P2N overcomes these limitations by preserving all original spatial information through two key stages:
1. Renoised Data Construction (RDC)
Leveraging the observation that noise is typically zero-mean and symmetric, the network predicts a denoised version $\hat{x}$ and the noise $\hat{n}$. It then generates "Positive" and "Negative" noisy versions by adding scaled versions of the predicted noise ($\sigma_p \hat{n}$ and $-\sigma_n \hat{n}$) back to the predicted clean image. This creates multiple renoised versions of the same scene without losing spatial resolution.
2. Denoised Consistency Supervision (DCS)
The network is trained to ensure that the denoised outputs for both the "Positive" and "Negative" noisy images are identical. A gradually varying norm ($\Vert\cdot\Vert_{2 \to 1.5}$) is employed for the loss function, allowing the model to handle diverse noise distributions robustly without manual parameter tuning for different datasets.
Experimental Results
P2N was evaluated on several standard benchmarks, including SIDD (Smartphone Image Denoising Dataset), CC, PolyU, and FMDD (Fluorescence Microscopy Denoising Dataset).
- State-of-the-Art Performance: On the SIDD benchmark, P2N achieved a PSNR of 35.80 dB, outperforming previous self-supervised methods by over 0.4 dB.
- Visual Fidelity: Qualitative analysis shows that P2N preserves fine textures and high-frequency details, such as text on battery labels, that are often blurred or distorted by competing methods.
- Extreme Efficiency: P2N demonstrates remarkable efficiency. A "fixed" version converges in fewer than 20 iterations, while the standard version converges in under 100 iterations—significantly faster than methods requiring thousands of iterations per image.
Conclusion
The Positive2Negative paradigm represents a significant shift in self-supervised denoising. By replacing information-lossy operations with a symmetry-based renoising strategy, P2N sets a new standard for both accuracy and efficiency in single-image restoration. Future work may extend this logic to broader datasets and multi-view scenarios.
Summary updated