Answers are grounded in this paper and cite their sources.
Multiplicative noise removal presents significant challenges in image processing due to its inherent nonlinearity. In this work, we address this problem in a single-image unsupervised setting, where a model is trained only from the noisy image to be restored. By eliminating the need for external training data, this setting enables image-specific adaptation and reduces potential mismatch between training and test data. Specifically, we derive unbiased risk estimators of the mean squared error (MSE) for two predominant types of multiplicative noise models, characterized by Gamma and normal distributions. We further extend the noise model to mixed multiplicative-additive noise and derive the associated risk estimators. Although the proposed statistical risk estimators are theoretically unbiased, their covariance and divergence terms are not directly tractable for neural-network training. To enable feasible training, we derive practical, computable estimators to approximate the covariance-related terms and implement a Monte-Carlo approximation to compute the derivatives in the divergence term. We theoretically prove that the proposed exact statistical risk estimators are unbiased, while the corresponding practical estimators are exact for linear denoisers and empirically accurate for nonlinear denoisers. Experiments show that the proposed estimators in the single-image setting achieve competitive or improved denoising performance compared with model-based and single-image unsupervised learning methods. Code is available at https://github.com/TISGroup/URE-Multi-Noise .
No takes yet. Share an insight, caveat, or question.
Yan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: