acceptodds
Withdrawn from ICLR 2027

Controlling Re-corruption Bias for Unsupervised Image Denoising under Unknown Noise

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unsupervised image denoising via re-corruption constructs noisy/noisy training pairs whose loss can reproduce supervised denoising under suitable noise assumptions. Existing guarantees, however, require sufficiently specified noise statistics, which are often unavailable in practice. We revisit re-corruption and identify a denoiser-independent criterion that characterizes the bias induced by a given re-corruption. This criterion depends only on a conditional moment of the re-corruption mismatch, yielding a weaker requirement than exact conditional noise matching and an equivalent variational formulation for learning re-corruption from noisy data alone. Based on this characterization, we first learn the re-corruption mechanism and then freeze it to fine-tune an initial denoiser, thereby decoupling re-corruption learning from denoiser optimization. By targeting the bias-relevant mismatch rather than estimating the full noise distribution, our method naturally accommodates unknown real-world noise. Experiments on smartphone and fluorescence-microscopy images show consistent denoising improvements and a clear reduction of the predicted bias.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image denoising, un-supervised learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wQUPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new approach to generate noisy/noisy pairs for training denoising neural networks in a self-supervised fashion. The method is based on the assumption that an initial denoiser f0f_0 is available and improves it with a two-stage procedure: 1) Learn re-corruption, 2) Fine-tune denoiser. The authors show state-of-the-art results on two popular real-world datasets: SIDD and FMD.

Critical strengths

The paper is very well-written and based on a solid mathematical result (Proposition 1). The ultimate goal is highly relevant and well motivated: how to generate noisy/noisy pairs for self-supervised learning. Finally, the experimental results (Table 2 and 3) are impressive.

Critical weaknesses

My main concern is the similarity with DCD-Net [a] and the high dependence on the initial denoiser f0f_0. Instead of generating noisy/noisy pairs in the spirit of Noise2Noise (N2N) as in [a], the authors propose to generate noisy/noisy pairs in the spirit of Recorrupted-to-Recorrupted (R2R). Both approaches rely on an initial denoiser f0f_0. The authors should at least compare with DCD-Net [a] since N2N is theoretically superior to R2R (N2N predicts from yy, while R2R predicts from a noisier version of yy, namely y+zy+z, which degrades performance).

[a] Zou, Y., Yan, C., & Fu, Y. Iterative denoiser and noise estimator for self-supervised image denoising. ICCV, 2023.

My overall rating is weak rejection but I may consider improving my rating based on the responses of the authors. Here are more details about my evaluation:

Major:

  1. The proposed method relies on an initial denoiser f0f_0 which is obtained via the R2R framework. It means that the method is not entirely self-contained. Moreover, R2R is supposed to work under the assumption of additive white Gaussian noise (AWGN). What if this assumption is not fulfilled and this initialization stage (see Fig. 1) fails? According to Table 5, it seems that the success of the method highly depends on f0f_0, and that the gain of the method when f0f_0 is properly chosen is not so important (only +0.23 dB). Conversely, when f0f_0 is not properly chosen, the method fails (35.88 dB vs 38.21 dB, see Table 5). I believe it is essential to include f0f_0 in Tables 2 and 3 and Figure 2 to better assess the gains achieved by the proposed method.

  2. I notice some similarities with the approach proposed in DCD-Net [a] (for which the reference is missing and should be added), inspired by the Noise2Noise framework. In particular, since you assume that you have access to pairs (x^,n^)(\hat{x}, \hat{n}) from noisy images yy thanks to f0f_0, where x^\hat{x} and n^\hat{n} are good approximations of xx and nn, respectively, I wonder if it is not possible to simply learn to generate samples from the same distribution as n^\hat{n} with a diffusion model for example (potentially conditioned on x^\hat{x}), or maybe using the same noise estimator as [a]; and, in a second stage, simply go back to equation (1) where zz is replaced by the generated samples. In this setting, Proposition 1 is not necessary. You could also imagine an iterative scheme to boost performance as in [a]. I therefore strongly recommend comparing your method with DCD-Net [a].

  3. Real-world noise datasets, namely FMD and SIDD, are considered for comparisons, which is great. But I would appreciate if the authors could also validate their method on synthetic noises (for example synthetic Poisson noise and compare with GR2R as upper bound). Also supervised baselines are missing in Table 2 and 3.

  4. L2R is also a work close to this one. However, as you mention, line 850, "[L2R] results should be interpreted as a transfer study rather than a reproduction of the original L2R experiments". I think that comparing with L2R trained from scratch on SIDD is necessary rather than a simple transfer study. However, when I see Table 6, I am confused. Did you train from scratch a network with the L2R loss?

Minor:

  1. Line 073, you state that the gap from supervised training is determined by CθC_\theta but you forget to mention that the denoiser has input y+zy+z (instead of yy for supervised learning), which could also introduce a non-negligible gap.
  2. Notations: Line 065, you introduce the notations xx, yy, zz, nn, without specifying what they represent or which spaces they belong to. Even though those notations are quite common in the image denoising literature, you should define them properly when using them for the first time. By the way, for the sake of clarity, it is common to write vectors in bold (for example, x\boldsymbol{x}, y\boldsymbol{y}, z\boldsymbol{z}, n\boldsymbol{n}), to better distinguish them from scalar quantities such as λμ\lambda_\mu or αft\alpha_{\text{ft}}. In Equation (3), you do not define what is gg. You should mention that g:RN×RN→RNg : \mathbb{R}^N \times \mathbb{R}^N \to \mathbb{R}^N.
  3. For the sake of clarity, you could mention that (1) is obtained by developing the squared ℓ2\ell_2 norm.
  4. Typos: line 229 and 235, equation 11 -> Equation 11 and line 231, equation 12 -> Equation 12
Questions for authors
  1. In Equation 1, you consider the symmetric re-corruption pair (y+z,y−z)(y+z, y-z) but in the original R2R paper, the pair (y+αz,y−z/α)(y+\alpha z, y-z / \alpha) is considered instead with α>0\alpha>0 for extra flexibility. Why don't you consider this parameter α\alpha as well from the very beginning? You call this hyperparameter αft\alpha_{\text{ft}} afterwards in equation (26).

  2. I didn't get why you need a stop-gradient in your implementation? Can you clarify please.

  3. Instead of considering (21) as learning objective, the generator objective is instead (25). Can you explain why simply considering (21) does not work? As you say, "Overall, the two stages are deliberately decoupled", while theory suggests that they should be coupled according to (21). Can you discuss this point?

  4. What happens when you iterate the process in the spirit of DCD-Net, that is replacing f0f_0 by the refined denoiser you have just learned and so on?

  5. Why not considering a blind-spot network for f0f_0 instead, as in [a], potentially more robust, the latter being based on less restrictive assumptions?

  6. Is the f0f_0 with white Gaussian (34.96 dB) in Table 5, the same as R2R in Table 2 (35.04 dB)?

  7. What do you mean by "R2R plain" in Figure 3? Does it correspond to (f0f_0 / White Gaussian) in Table 5? If so, could you also provide a visual comparison with (f0f_0 / NLF and ISP)?

Reviewer_MV5FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on re-corruption-based unsupervised denoising under unknown noise assumption. It estimate the loss bias through a denoiser independent criterion. A practical surrogate replace clean images and noise with estimates from an initial denoiser. A re-corruption generator is then learned against a regression critic, and used to finetune the denoiser. Experiments on SIDD and FMD show improvements over the initial model, and further analysis explores how different re-corruption strategies affect training bias

Critical strengths
  • The paper provide a clear statistical insight, and is easy to follow and understand.
  • The experiments cover both camera and microscopy images and compare the method with recent sota baselines, while showing promising performances
Critical weaknesses
  • The best result relies on noise calibration, which doesn't hold under the assumption of unknown noise assumption. Appendix B.1 uses raw noise-level function and ISP for initialization. This makes the result less convincing for the unknown-noise setting
  • The practical surrogate may not track the true bias. It depends on an initial denoiser whose residual can contain missing image details. The paper acknowledges that the surrogate is neither guaranteed to be unbiased nor an upper bound, but provides limited evidence about when it remains reliable
  • The generator uses an approximate stop-gradient update. It also learns symmetric re-corruption, while the strongest SIDD result uses (αft=5)(\alpha_{\mathrm{ft}}=5). The bias measurements do not cover this main configuration.
  • The unique contribution needs to be explained more explicitly. GR2R already connects re-corruption bias to moment conditions, and L2R already learns re-corruption. The paper should explain more clearly what its unique criterion and separate critic beyond these ideas
Questions for authors
  • How should the unknown-noise claim be interpreted given the calibrated initialization? Do existing controls show that the gain over the matched starting model comes from learned re-corruption rather than additional fine-tuning?
  • What evidence support that the practical updates reduce actual bias in the main SIDD configuration?
Reviewer_HK1WOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies re-corruption-based unsupervised denoising in the R2R family. The training loss differs from the supervised risk by a cross term, which the authors bound with a denoiser-independent quantity QQ; Proposition 1 shows that QQ is the energy of the conditional mean E[z−n∣x,n+z]\mathbb{E}[z-n\mid x,n+z]. In practice, xx and nn are replaced by the output and residual of an initial R2R denoiser f0f_0, and a stochastic re-corruption generator is trained against an MSE-regression critic, with mean and energy constraints. The generator is then frozen and f0f_0 is fine-tuned with reciprocal R2R pairs. Experiments cover SIDD and FMD, a diagnostic study relating the measured cross term, the surrogate Q^\hat Q and PSNR, and appendix ablations.

Critical strengths
  1. A clean, denoiser-independent handle on re-corruption bias that leads directly to a trainable objective. Proposition 1 turns the cross term into a conditional-mean regression problem, so the critic is a plain MSE regressor rather than a real/fake discriminator, and only the denoiser is needed at test time. Under a fixed f0f_0, the measured ∣2Cθ∣|2C_\theta| drops from 17.75 to 1.78 (10−410^{-4} per pixel), Q^\hat Q follows the same ordering, and PSNR improves (Table 4). The gain over the matched initialization is +1.01 dB from a white-Gaussian f0f_0 and positive for all three initializations (Table 5).
  2. Transparent reporting. The appendix discloses the f0f_0 numbers, the effect of the shared IDR round on FMD, the exact-energy-matching coupling and the caveats of the L2R transfer. Several weaknesses below rely on these disclosures.
Critical weaknesses
  1. For the generator class used, the relaxation over distribution matching is not substantiated, and the practical objective is only loosely tied to the oracle criterion.

    • In the penalty variant, zϕ=hϕ(x^,ϵ)z_\phi=h_\phi(\hat x,\epsilon) is conditionally independent of n^\hat n given x^\hat x. Under additional regularity assumptions (e.g. non-vanishing or analytic conditional characteristic functions), exact Q^ϕ=0\hat Q_\phi=0 then coincides with conditional distribution matching to the residual of f0f_0. This concerns the surrogate, not the oracle conditioned on xx, and does not cover the exact-energy-matching variant (B.4); the method also still avoids explicit density modeling. So the general statement may hold, but the "weaker target" claim is not substantiated as instantiated. Proposition 1 itself is the standard L2L^2-projection characterisation of the conditional mean, so the theoretical novelty is modest.
    • The optimized Q^\hat Q comes from a finite critic and is, by the authors' own statement, neither an unbiased estimator nor an upper bound of QQ; the stop-gradient update is a disclosed approximation (App. A.2).
    • Headline SIDD uses αft=5\alpha_{ft}=5, where the relevant bias is Eq. 27, not the symmetric criterion used for training. No bias diagnostic is given for this setting.
    • The link between Q^\hat Q and the true bias rests on three re-corruptions and one initialization, without estimator uncertainty or a critic-capacity study.
  2. The evidence does not support positioning the headline result as an unknown-noise, state-of-the-art result.

    • The Table 2 result starts from an f0f_0 trained with the SIDD noise-level function and ISP rendering, i.e. with camera noise metadata. This f0f_0 already reaches 37.98 / 38.26 dB but is not listed; the method adds about 0.2 dB on top of it.
    • Without noise metadata (B.2), the method gains +1.01 dB over its f0f_0 but reaches only 35.97 dB on validation, below most blind-spot methods in Table 2, and this setting is absent from the main table. (Minor: the Sec. 4.3 gains are computed against the αft=1\alpha_{ft}=1 white-Gaussian row; against the αft=0.5\alpha_{ft}=0.5 row they are 0.96 / 0.72 dB.)
    • On FMD, the gain over matched R2R is 0.10 dB after the shared IDR round, and the 0.01 dB margin over FM2S comes from a different protocol.
    • Checkpoints and αft\alpha_{ft} are selected on SIDD validation with ground truth, which is also a reported column, and no run-to-run variability is given for 0.1–0.2 dB effects.
    • There is no synthetic experiment with a known noise model comparing oracle QQ with Q^\hat Q, which is also L2R's setting. The warm-started L2R comparison (Table 6) is, as the authors note, not a reproduction of L2R's own setting.
Questions for authors

see the weakness part.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.