Denoising Discrepancy Metrics

• Denoising discrepancy metrics are label-free evaluation measures that assess the quality of denoised signals using only noisy observations and robust statistical frameworks.
• They employ diverse methodologies including unsupervised estimators, variational measures, and uncertainty quantification to provide unbiased and consistent performance estimates.
• These metrics are vital in applications such as event cameras, diffusion models, and medical imaging, ensuring optimal trade-offs between noise removal and signal preservation.

A denoising discrepancy metric quantitatively evaluates the fidelity or quality of a denoised signal when no access to the clean ground-truth is possible or practical. In modern imaging and signal processing, robust evaluation of denoising performance—using only observed, potentially noisy data—is essential in contexts including event cameras, unsupervised learning algorithms, diffusion-based generative models, and medical imaging. Recent research has yielded a variety of such metrics, ranging from domain-specific measures grounded in physical signal properties, to generic statistical estimators derived from optimal transport and uncertainty principles. This article surveys the mathematical formalisms, theoretical guarantees, algorithmic implementations, and comparative characteristics of denoising discrepancy metrics, with a particular focus on metrics that remain label-free, non-monotonic, or model-agnostic.

1. Mathematical Foundations of Denoising Discrepancy Metrics

A denoising discrepancy metric is designed to assess the difference between a denoised output and an underlying reference, which may be a noisy observation, a distributional posterior, or an abstract property inferred from observed data streams. Several mathematical frameworks underlie such metrics:

• Statistical estimators: Unsupervised variants of MSE and PSNR, such as uMSE and uPSNR, leverage multiple independent noise realizations to asymptotically estimate the supervised error (Marcos-Morales et al., 2022).
• Variational measures: Optimal transport distances and the Kantorovich–Rubinstein (KR) norm are used as discrepancy terms in variational denoising models. The KR norm interpolates between ℓ¹-fidelity and Wasserstein distances and is formalized through dual-Lipschitz or primal transport plans (Lellmann et al., 2014).
• Contrast-curve analysis: In event-data, contrast metrics exploit the temporal evolution of spatial event frame statistics, leading to metrics like the Area of the Continuous Contrast Curve (AOCC), which directly encode spatiotemporal edge preservation (Shi et al., 2024).
• Uncertainty quantification: Bayesian posterior sampling and maximum mean discrepancy (MMD) are employed to aggregate uncertainty in the denoised signal, particularly in high-dimensional data such as diffusion MRI (Fadnavis et al., 2022).
• Flow divergence: In diffusion generative models, divergence between learned and ideal denoising flows is quantified via total variation rates (Schedule Deviation), providing a diagnostic for model consistency (Pfrommer et al., 21 Dec 2025).

2. Properties and Theoretical Guarantees

Denoising discrepancy metrics differ critically in terms of label dependence, monotonicity, sensitivity, and statistical robustness.

• Label-free nature: Metrics like AOCC, uMSE/uPSNR, NUQ, and several no-reference quality regressors require only observed or denoised data, never the clean ground truth (Shi et al., 2024, Marcos-Morales et al., 2022, Fadnavis et al., 2022, Lu, 2018).
• Non-monotonicity and sensitivity: Label-free AOCC is explicitly non-monotonic with respect to denoiser aggressiveness, peaking only at the optimal trade-off between signal preservation and noise removal. In contrast, monotonic metrics like NeRr, VeRr, or SNR can encourage over-denoising (Shi et al., 2024).
• Unbiasedness and consistency: Unsupervised MSE estimators are proven to be unbiased and asymptotically consistent under independent noise models, with variance decaying as $\mathcal{O}(1/n)$ in the number of samples (Marcos-Morales et al., 2022).
• Optimal transport and mass preservation: The Kantorovich–Rubinstein discrepancy enforces mass-conservation properties and interpolates between ℓ¹ and Wasserstein distances by tuning its amplitude and transport-scale parameters (Lellmann et al., 2014).

3. Key Domain-Specific Metrics

A selection of domain-specific denoising discrepancy metrics is outlined, each grounded in mathematically precise definitions:

Metric	Domain	Core Principle
AOCC	Event cameras	Area under non-monotonic contrast vs interval curve (Shi et al., 2024)
uMSE, uPSNR	General imaging	Unbiased error via triple-noisy reference (Marcos-Morales et al., 2022)
Schedule Deviation	Conditional diffusion	TV-rate of deviation from ideal denoising flow (Pfrommer et al., 21 Dec 2025)
KR–TV Discrepancy	Variational imaging	Optimal-transport fidelity interpolating between norms (Lellmann et al., 2014)
NUQ	Diffusion MRI	Posterior uncertainty via kernel MMD (Fadnavis et al., 2022)
No-Reference Quality	Generic images	Feature regression surrogate for PSNR/SSIM (Lu, 2018)

AOCC: Area of the Continuous Contrast Curve

Given an event stream $E = \{\mathbf{e}_k = (x_k, y_k, p_k, t_k)\}$, AOCC bins events over variable intervals $\Delta t$ and computes the spatial contrast of resulting event frames. The AOCC is the integral over the curve $C_\text{avg}(\Delta t)$, with maxima indicating optimal denoising (Shi et al., 2024).

uMSE and uPSNR

With triple-independent noisy references $a, b, c$, the unsupervised MSE is $$uMSE(f) = \frac{1}{n}\sum_{i=1}^n (a_i - f(y)_i)^2 - \frac{1}{2n}\sum_{i=1}^n (b_i - c_i)^2$$. The correction term removes noise variance, yielding an unbiased estimator of true MSE in expectation (Marcos-Morales et al., 2022).

Schedule Deviation in Diffusion Models

Schedule Deviation quantifies the instantaneous total-variation rate between the learned denoising velocity field $\hat v_s(x,z)$ and the ideal model-consistent flow $v^{\text{ideal}}_s(x,z)$ over the trajectory of the generative process (Pfrommer et al., 21 Dec 2025).

Kantorovich–Rubinstein Norm

For signals $u, u^0$, the KR-norm discrepancy is $$\|u-u^0\|_{\mathrm{KR},(\lambda_1,\lambda_2)} = \sup\{\int f\,(u-u^0)\;dx: \|f\|_\infty \leq \lambda_1, \|f\|_{\text{Lip}} \leq \lambda_2\}$$, interpolating between $\ell^1$ and Wasserstein-$1$ distances and supporting mass-conservation (Lellmann et al., 2014).

NUQ: Noise Uncertainty Quantification

NUQ uses posterior samples of microstructural measures, split into two ensembles, and computes their MMD$^2$ as a discrepancy score, insensitive to mean-shifting but sensitive to remaining uncertainty (Fadnavis et al., 2022).

No-reference Quality Metrics

Aggregates hand-engineered perceptual and structural features into a regression model trained to estimate PSNR/SSIM using only noisy and denoised image pairs (Lu, 2018).

4. Practical Algorithms and Implementation Considerations

Implementation details vary by metric. For example, AOCC computation involves binning and gradient evaluation over multiple time intervals, requiring attention to the choice of $\Delta t$ grid, gradient operator (Sobel, Prewitt), and efficient integration (trapezoidal rule) (Shi et al., 2024). uMSE/uPSNR require three independent noisy inputs, with estimators implemented via vectorized operations and bootstrap sampling for confidence intervals (Marcos-Morales et al., 2022). Schedule Deviation necessitates estimation of high-dimensional score functions and Jacobian traces, which can be accelerated via random projection techniques (Pfrommer et al., 21 Dec 2025). KR–TV discrepancies are computed via primal–dual solvers with convex–concave splitting, balancing numerical efficiency and convergence (Lellmann et al., 2014).

5. Comparative Evaluation and Experimental Results

Metrics are validated using both synthetic and real-world datasets.

• AOCC robustly identifies the optimal denoising threshold in event camera data, exhibiting a unique maximum at the point of best edge-contour preservation, unlike strictly monotonic classical event metrics (Shi et al., 2024).
• uMSE/uPSNR closely match supervised MSE/PSNR rankings when independent references are available ($|uPSNR - PSNR| \lesssim 0.1$ dB); performance degrades gracefully under spatial subsampling in natural images (Marcos-Morales et al., 2022).
• Schedule Deviation correlates strongly ($r > 0.9$) with the observed disagreement between sampler outputs (e.g., DDPM vs DDIM), revealing “drift” due to inductive biases in the conditional flow (Pfrommer et al., 21 Dec 2025).
• In diffusion MRI, NUQ discriminates denoisers and reveals subtle spatial patterns of residual uncertainty, directly impacting group-analysis sensitivity by modulating Bayesian inference weights (Fadnavis et al., 2022).
• No-reference image quality regressors significantly outperform previous methods (Kendall’s $\tau = 0.854$ for PSNR ranking on seen noise levels), supporting robust ranking and parameter selection in the absence of clean images (Lu, 2018).

6. Limitations and Open Problems

Several limitations and domain challenges remain:

• Sampling and computational complexity: Metrics involving integration over large grids (AOCC), kernel-based pairwise statistics (NUQ), or high-dimensional Jacobians (Schedule Deviation) can be compute-intensive; optimizations include coarser grids, sub-sampling, or GPU acceleration (Shi et al., 2024, Fadnavis et al., 2022, Pfrommer et al., 21 Dec 2025).
• Domain sensitivity: Some metrics (e.g., AOCC) may be insensitive to image color/texture if contrast is preserved; others (e.g., uMSE via subsampling) may bias results in non-flat image regions (Marcos-Morales et al., 2022).
• Model assumptions: Metrics such as uMSE require independence and zero-mean of noise, while the KR-norm presupposes measure properties that may not be met in all data (Lellmann et al., 2014).
• Generality vs. specificity: Highly domain-specific metrics (e.g., AOCC or NUQ) leverage particular signal properties but may be less interpretable out-of-domain; generic methods can lack sensitivity to application-critical structure.

Future work may extend metrics to be polarity-aware (AOCC), fully spatiotemporal (edge-contrast in 3D), or act as differentiable losses for self-supervised denoiser training. Handling correlated noise or adapting to more general signal models are other open challenges.

7. Impact and Future Directions

The development of denoising discrepancy metrics has enabled rigorous benchmarking, parameter selection, and algorithm validation in settings characterized by few or no ground-truth clean examples. Their direct integration as label-free objectives supports auto-tuning, unsupervised evaluation, and deployment in real-time or mission-critical applications (e.g., event vision, medical imaging). Future trends may involve:

• Incorporating discrepancy metrics into self-supervised learning frameworks (e.g., differentiable AOCC losses) (Shi et al., 2024).
• Extending uncertainty-based approaches to more general imaging modalities (e.g., extensions of NUQ to non-linear or non-Gaussian posteriors) (Fadnavis et al., 2022).
• Generalizing discrepancy metrics to cover a wider array of restoration tasks (e.g., super-resolution, inpainting) via reference-based strategies (Marcos-Morales et al., 2022).
• Bridging the gap between statistical consistency and perceptual quality through hybrid metrics combining feature-driven, task-dependent, and distributional perspectives (Lu, 2018).

Denoising discrepancy metrics, through rigorous mathematical formulation and empirical effectiveness, constitute a critical component of modern signal and image processing pipelines, fundamentally shaping the theory and practice of unsupervised quality assessment.