Noise2Noise denoising performance driven by training data distribution, not loss choice

A new paper revisits the Noise2Noise (N2N) self-supervised denoising technique, challenging common assumptions about why L1 loss outperforms L2 loss. The research suggests that the training pair distribution, rather than the choice of loss function, is the primary factor influencing N2N performance. Experiments on synthetic and real-world noise datasets demonstrate that N2N models trained on specific noise distributions achieve significant gains, even surpassing traditional methods like BM3D when trained on relevant noisy pairs without clean references. AI

IMPACT This research clarifies fundamental aspects of self-supervised learning, potentially guiding future development in image denoising and related fields where clean data is scarce.

RANK_REASON The item is an academic paper detailing a new finding in self-supervised learning for image denoising. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Block-matching and 3D filtering
• Kodak24
• L1 Loss
• L2-loss Large-scale Linear Nonparallel Support Vector Ordinal Regression
• Lasso regularization for left-censored Gaussian outcome and high-dimensional predictors
• Noise2Noise
• The Buddha

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →