Recorrupted-to-Recorrupted (R2R) Denoising

• R2R is a self-supervised image restoration method that constructs synthetic noisy pairs to emulate supervised denoising without requiring clean images.
• It generates multiple noisy views from a single observation by adding statistically independent synthetic noise, ensuring consistency with the supervised mean-squared-error objective.
• Extensions like GR2R adapt the framework for non-Gaussian noise and physics-guided inverse problems, achieving competitive performance compared to standard supervised training.

Recorrupted-to-Recorrupted (R2R) is a self-supervised image-restoration methodology in which a model is trained from noisy observations alone by constructing multiple synthetic re-corruptions of a single noisy measurement and enforcing consistency between them. In its canonical form, R2R targets denoising under additive Gaussian noise by generating a pair of noisy surrogates from one observation and optimizing a squared loss that is equivalent in expectation to supervised mean-squared-error training up to an additive constant. Subsequent work has extended the framework beyond Gaussian noise and adapted it to physics-guided inverse problems, including terahertz (THz) denoising and deblurring with explicit point-spread-function modeling and PCA-based spectral decomposition (Monroy et al., 2024, Zhu et al., 17 Jan 2026).

1. Statistical construction and core objective

In the Gaussian formulation summarized in "Generalized Recorrupted-to-Recorrupted: Self-Supervised Learning Beyond Gaussian Noise" (Monroy et al., 2024), the clean image is $x \in \mathbb{R}^n$, the noisy measurement is $y$, and the observation model is

$y \mid x \sim \mathcal{N}(x,\sigma^2 I).$

The objective is to learn a denoiser $f_\theta$ approximating the MMSE estimator by minimizing supervised risk $\|f(y)-x\|_2^2$, but without access to clean targets.

R2R replaces unavailable clean supervision by constructing a recorrupted pair

$$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$

where

$\omega \sim \mathcal{N}(0,\sigma^2 I),$

independent of the original noise in $y$, and $\tau>0$ is a hyper-parameter. Training then uses

$$L_{\mathrm{R2R}}(y;f)=\mathbb{E}_{y_1,y_2\mid y}\|f(y_1)-y_2\|_2^2.$$

The central identity is that

$y$0

so minimizing the self-supervised R2R loss is equivalent, in expectation, to supervised MSE training on $y$1 in the Gaussian case (Monroy et al., 2024). The argument relies on two stated conditions: $y$2 and $y$3 are conditionally independent given $y$4, and $y$5. Conceptually, R2R constructs a pseudo-Noise2Noise pair from a single noisy observation rather than from repeated acquisitions.

2. Generalization beyond Gaussian noise

The principal extension of the framework is GR2R, which broadens R2R to additive non-Gaussian noise and to noise models in the natural exponential family (NEF), including Poisson, Gamma, Binomial, and Gaussian distributions (Monroy et al., 2024). For additive non-Gaussian noise $y$6, the same linear recorruption

$y$7

is retained, but unbiasedness is no longer automatic. The GR2R analysis states that the R2R objective expands into a supervised MSE term plus an error term involving higher-order moments and derivatives of $y$8. For linear $y$9, matching second moments is sufficient; for quadratic behavior, third-moment matching is additionally required.

For NEF noise, GR2R introduces a splitting parameter $y \mid x \sim \mathcal{N}(x,\sigma^2 I).$0 and represents the observation as

$y \mid x \sim \mathcal{N}(x,\sigma^2 I).$1

The reconstruction of $y \mid x \sim \mathcal{N}(x,\sigma^2 I).$2 from $y \mid x \sim \mathcal{N}(x,\sigma^2 I).$3 and $y \mid x \sim \mathcal{N}(x,\sigma^2 I).$4 is

$y \mid x \sim \mathcal{N}(x,\sigma^2 I).$5

Under the stated NEF construction, $y \mid x \sim \mathcal{N}(x,\sigma^2 I).$6 and $y \mid x \sim \mathcal{N}(x,\sigma^2 I).$7 are conditionally independent given $y \mid x \sim \mathcal{N}(x,\sigma^2 I).$8, both have conditional mean $y \mid x \sim \mathcal{N}(x,\sigma^2 I).$9, and the conditional laws $f_\theta$0 and $f_\theta$1 do not depend on $f_\theta$2. The resulting GR2R-MSE loss

$f_\theta$3

is therefore an unbiased estimator of supervised MSE risk, and an NEF negative-log-likelihood analogue is also defined (Monroy et al., 2024).

The same paper further states that Stein’s unbiased risk estimator can be viewed as a limiting case of GR2R as $f_\theta$4. In the Gaussian setting, this limiting form recovers classical SURE exactly. This unifies R2R and SURE as finite- and infinitesimal-splitting views of the same risk-estimation principle.

3. Position relative to other self-supervised restoration methods

The THz-SSDD paper explicitly situates R2R against Noise2Noise, blind-spot methods such as Noise2Void, and related recorrup­tion frameworks (Zhu et al., 17 Jan 2026). Relative to Noise2Noise, the decisive distinction is that R2R requires only one measured noisy image per sample and synthesizes the second noisy view through re-corruption. Relative to Noise2Void and Noise2Self, R2R does not depend on masking-based blind-spot architectures and does not require the denoiser to ignore the center pixel.

The GR2R paper adds a second axis of comparison (Monroy et al., 2024). It states that blind-spot methods impose architectural constraints that prevent learning the full MMSE estimator, whereas GR2R directly trains a denoiser against pseudo-targets generated from the assumed noise model. It also contrasts GR2R with SURE-, GSURE-, and PURE-type estimators, emphasizing that GR2R avoids explicit divergence computation by working with finite re-corruptions rather than differential identities.

The common premise across these formulations is that one can generate statistically valid synthetic noisy views from a single observation. In the Gaussian case, this yields exact equivalence to supervised MSE. In additive non-Gaussian settings, the GR2R analysis indicates that exactness depends on moment conditions. A plausible implication is that the usefulness of R2R is strongest when the corruption model is either analytically tractable or well enough specified to support faithful re-corruption.

4. Physics-guided THz formulation and PCA integration

"Principal Component Analysis-Based Terahertz Self-Supervised Denoising and Deblurring Deep Neural Networks" adapts the general R2R idea to THz imaging, where degradation combines low-frequency blurring and high-frequency noise, and embeds the strategy inside a physics-aware inverse-problem pipeline (Zhu et al., 17 Jan 2026). The observed THz image at each frequency is modeled as

$f_\theta$5

with empirical frequency-dependent variance

$f_\theta$6

The THz recorruption step is

$f_\theta$7

where $f_\theta$8, $f_\theta$9, $\|f(y)-x\|_2^2$0 is a $\|f(y)-x\|_2^2$1 average filter, and $\|f(y)-x\|_2^2$2 is an invertible identity-like matrix. Training uses a composite loss

$\|f(y)-x\|_2^2$3

with

$\|f(y)-x\|_2^2$4

The first term provides pixel-wise self-supervision for denoising, the second enforces consistency with the THz forward model after PSF convolution and down-sampling, and the third regularizes against spurious high-frequency artifacts.

The same system couples R2R with PCA on the THz spectral cube. The cube $\|f(y)-x\|_2^2$5 is reshaped into $\|f(y)-x\|_2^2$6, mean-centered, and decomposed through the covariance matrix and eigenproblem

$\|f(y)-x\|_2^2$7

Retaining the first $\|f(y)-x\|_2^2$8 principal components produces

$\|f(y)-x\|_2^2$9

and restored components are reconstructed by

$$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$0

The authors state that lower-index PCs capture most of the variance and large-scale, low-frequency content, whereas higher-index PCs capture small-scale details and high-frequency noise, and they recommend retaining PCs explaining $$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$1–$$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$2 of the variance (Zhu et al., 17 Jan 2026). PCA is therefore not part of R2R itself but a complementary decomposition that makes the R2R-trained network effective across frequencies.

5. Empirical behavior, assumptions, and limitations

The GR2R study reports experiments with Gaussian, Poisson, Gamma, and log-Rayleigh noise, using PSNR and SSIM as evaluation metrics (Monroy et al., 2024). For additive log-Rayleigh noise, the paper reports $$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$3 dB for second-moment-only R2R, $$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$4 dB for third-moment-matched GR2R, and $$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$5 dB for supervised training. For Poisson, Gamma, and Gaussian denoising, the reported GR2R variants are stated to be essentially on par with supervised training in the tested regimes.

The THz-SSDD work evaluates the R2R-based method on four sample types and states that training requires only a small set of unlabeled noisy images (Zhu et al., 17 Jan 2026). Training uses noisy GFRP calibration data, Adam, Kaiming initialization, an initial learning rate of $$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$6, and takes approximately $$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$7 hours on an NVIDIA 4060-class GPU. During training on GFRP, overall loss decreases monotonically with epochs, and visual PSNR in the training sample shown in Fig. 4(b) improves from approximately $$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$8 dB to approximately $$y_1 = y + \tau \omega, \qquad y_2 = y - \frac{1}{\tau}\omega,$$9 dB. Across pyrolyzed woods, HDPE dog-bone specimens, and hybrid composites, the method is reported to reduce low-frequency blur and high-frequency noise and to preserve spectral structure, with relative spectral error values mostly around $\omega \sim \mathcal{N}(0,\sigma^2 I),$0.

These results rest on explicit assumptions. In the Gaussian derivation, conditional independence and correct centering of the pseudo-target are essential (Monroy et al., 2024). In THz-SSDD, the added synthetic noise must be independent of the existing noise, zero-mean, and statistically consistent with the estimated local variance, while the PSF and down-sampling operator must correctly represent the imaging system (Zhu et al., 17 Jan 2026). The GR2R paper also notes that independence across pixels is assumed for its NEF constructions, and that additive non-Gaussian unbiasedness is only approximate unless sufficient higher-order moments are matched. This suggests that R2R is not a generic corruption-agnostic recipe; it is a model-based self-supervised estimator whose reliability is tied to the fidelity of the recorruption process.

6. Terminological ambiguity of “R2R”

The acronym “R2R” is not unique across the literature. In image restoration it denotes Recorrupted-to-Recorrupted, but several unrelated uses coexist.

Usage	Domain	Meaning
R2R	Self-supervised restoration	Recorrupted-to-Recorrupted
R2R	Vision-and-language navigation	Room-to-Room
R2R	Function-preserving transforms	Residual-compatible widening/deepening transforms
R2R	ROS 2 software	Async Rust ROS 2 bindings

In vision-and-language navigation, both "General Scene Adaptation for Vision-and-Language Navigation" and "Sub-Instruction Aware Vision-and-Language Navigation" state that R2R refers to the classic Room-to-Room benchmark, not Recorrupted-to-Recorrupted (Hong et al., 29 Jan 2025, Hong et al., 2020). In "Towards a More Complete Theory of Function Preserving Transforms", R2R denotes a pair of residual-compatible function-preserving transforms, R2WiderR and R2DeeperR; the paper explicitly says the name is a playful reference to symmetric cancellation rather than to stochastic corruption (Painter, 2024). In "A first look at ROS 2 applications written in asynchronous Rust", R2R refers to a community Rust client library for ROS 2 with a fully asynchronous API (Škoudlil et al., 27 May 2025).

Because of this ambiguity, the term “Recorrupted-to-Recorrupted” is most precise when used in the context of self-supervised denoising and inverse problems. Within that literature, it denotes a specific strategy: generate statistically valid synthetic re-corruptions of noisy data and train against another corrupted view so that, in expectation, the optimization reproduces the supervised objective (Monroy et al., 2024).