Nonlinear Noise2Noise for HDR Image Denoising
Analysis
This paper addresses a key limitation of the Noise2Noise method, which is the bias introduced by nonlinear functions applied to noisy targets. It proposes a theoretical framework and identifies a class of nonlinear functions that can be used with minimal bias, enabling more flexible preprocessing. The application to HDR image denoising, a challenging area for Noise2Noise, demonstrates the practical impact of the method by achieving results comparable to those trained with clean data, but using only noisy data.
Key Takeaways
- •Addresses the bias problem in Noise2Noise caused by nonlinearities.
- •Provides a theoretical framework for analyzing the effects of nonlinear functions.
- •Identifies a class of nonlinear functions with minimal bias.
- •Applies the method to HDR image denoising, a challenging application.
- •Achieves results comparable to those trained with clean data, but using only noisy data.
Reference
“The paper demonstrates that certain combinations of loss functions and tone mapping functions can reduce the effect of outliers while introducing minimal bias.”