# Zero-Shot Noise2Noise: Efficient Image Denoising without any Data

**Authors:** Youssef Mansour, Reinhard Heckel  
**arXiv:** [2303.11253](https://arxiv.org/abs/2303.11253)  
**Date:** 2023-03-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.11253v1/

## Key points

- Self-supervised neural networks have shown promising results in image denoising
- Existing dataset-free methods have limitations such as high computational cost, noise model requirement, and inadequate image quality
- The authors propose a simple 2-layer network called ZS-N2N that achieves high-quality image denoising without training data or knowledge of the noise distribution
- ZS-N2N is inspired by Noise2Noise and Neighbor2Neighbor techniques and works well for pixel-wise independent noise
- Experiments show that ZS-N2N outperforms other dataset-free methods while maintaining low computational cost
- Traditional dataset-based methods face challenges with small datasets or different noise levels in test data, highlighting the need for dataset-free denoising algorithms
- Methods relying on explicit noise models or tailored to specific distributions may struggle to generalize or require accurate estimation of noise levels
- BM3D performs well on Gaussian noise but not on Poisson and real-world noise, emphasizing the advantage of using model-free techniques like ZS-N2N.

## Summary

Recently, self-supervised neural networks have shown promising results in image denoising. However, existing dataset-free methods suffer from limitations such as high computational cost, the requirement of a noise model, or inadequate image quality. In this study, the authors propose a simple 2-layer network called ZS-N2N (Zero Shot Noise2Noise) that can achieve high-quality image denoising without any training data or knowledge of the noise distribution. The approach is inspired by Noise2Noise and Neighbor2Neighbor techniques and is particularly effective for denoising pixel-wise independent noise. The experiments conducted on artificial, real-world camera, and microscope noise demonstrate that ZS-N2N often outperforms other dataset-free methods while maintaining low computational cost. This makes it suitable for scenarios with limited data availability and compute resources. It is worth noting that traditional dataset-based methods excel in achieving state-of-the-art results but face challenges when trained on small datasets or when the test data differs from the training data in terms of noise levels. This highlights the importance of developing dataset-free denoising algorithms. Methods that rely on an explicit model of the noise distribution or those tailored to specific distributions may struggle to generalize well to unknown distributions or require accurate estimation of noise levels. For example, BM3D performs well on Gaussian noise but fails to keep up with other baselines on Poisson and real-world noise. These findings emphasize the advantage of using noise model-free techniques like ZS-N2N.