Study Proposes Bias-Reduced Noise2Noise Training for HDR Image Denoising

Global: Study Proposes Bias-Reduced Noise2Noise Training for HDR Image Denoising

Background

Researchers have introduced a theoretical framework that identifies a class of nonlinear functions which can be applied to noisy target images in Noise2Noise training with minimal bias. This work targets a well‑known limitation of the Noise2Noise paradigm, where nonlinear processing of noisy targets can skew denoising outcomes.

Challenge with Nonlinearities

Noise2Noise training traditionally avoids the need for clean reference images by using pairs of noisy inputs and targets. However, applying common nonlinear operations—such as tone mapping—has been considered incompatible because the expected value of the noisy targets diverges from that of the clean image.

New Theoretical Insight

The authors demonstrate that certain nonlinear functions, when combined with specific loss functions, introduce only negligible bias. They develop analytical tools to quantify the bias introduced by these nonlinearities, establishing criteria for selecting functions that preserve training fidelity.

Experimental Evaluation

The proposed approach is evaluated on high dynamic range (HDR) images generated via Monte Carlo rendering, a domain where outliers frequently overwhelm the training process. By applying carefully chosen tone‑mapping functions and loss combinations, the method mitigates outlier effects while maintaining denoising quality.

Results

Integration of the technique into an existing machine‑learning Monte Carlo denoiser—originally trained with high‑sample‑count reference images—yields performance that closely matches the baseline, despite relying solely on noisy training data.

Implications

The findings broaden the applicability of Noise2Noise training to scenarios that require nonlinear preprocessing, potentially reducing the cost and complexity of acquiring clean reference data for HDR rendering pipelines.
This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.