acceptodds
Withdrawn from ICLR 2027

Which Denoiser? Blind Selection from Noise Structure

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Denoising research has produced candidates of increasingly narrow competence, each strong on the noise it was designed or trained for and weak elsewhere. On an arbitrary image, nothing indicates which candidate to run: the scene and the corruption are both unknown, and no clean reference is available. Practitioners therefore apply the denoiser that performs best on average, which in the pools we study falls 1 to 2.5 dB short of the best choice for each image. Closing that gap requires knowing which denoiser suits the image, which is usually approached either by identifying the corruption first or by running every candidate and judging the results. We show that the right denoiser can be found without identifying the corruption exactly, since corruptions that differ often call for the same treatment. In a space of thirteen statistics of the corrupted image, each measuring one physical property of the process that produced the noise, the nearest neighbor of an image has the same best denoiser in 83 to 89% of cases, even though it has the same corruption in only 60%. Matching a new image to its nearest calibration image in this space and running the denoiser that won there recovers 43 to 66% of the available improvement on corruptions withheld from calibration, in one denoising call and with no retraining. Routing falls below the best overall denoiser on noise far from any calibration condition, such as real camera noise, but adding labeled references of that noise restores the gain without retraining. A descriptor space organized by how noise is best treated therefore turns a pool of denoisers with different strengths into a better denoiser than any of its members.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
blind image denoising, denoiser selection, algorithm selection, noise characterization, distribution shift, no-reference image quality assessment

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_q5foOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method to select automatically the best-performing denoiser among a set of off-the-shelf denoisers, given a noisy image with unknown noise type and noise strength. No training is necessary. The method is based on computing basic statistics of the noise directly on the noisy image itself, such as estimations of the variance of the noise, skewness, kurtosis... and then makes a prediction via nearest neighbor selection on a calibration set. The authors claim that their method is better than simply considering the best overall denoiser or the one chosen based on a blind estimation of the PSNR.

Critical strengths

A lot of off-the-shelf denoisers are available, from classical ones to neural networks. Each have their strengths and weaknesses, and some are dedicated to specific noise types. From the practitioner's point of view, it may be difficult to know in advance which denoiser will be the most efficient on their input noisy image. The paper tackles this issue, which is little studied in the literature, to the best of my knowledge. One of the main advantage of the method is that no additional training is required: once chosen, the denoiser can be used directly, without fine-tuning.

Critical weaknesses

Overall, I am not enthusiastic about this paper, which I find rather weak in terms of contributions and novelty. I do not think the paper meets the acceptance bar for ICLR, despite the numerous experiments. Here are the main weaknesses in my opinion:

Major:

  1. The main idea is not very original and not very sophisticated: the authors basically propose to compute some basic statistics of the noise from the input noisy image itself (the descriptor) and then use the denoiser which was identified to perform best for this descriptor (routing procedure), identified via a simple nearest neighbor technique on a calibration set.
  2. For the routing procedure, a calibration set consisting of pairs of noisy/clean images is necessary. Denoising must be run, for each noisy image of the calibration set, for each denoiser in the pool of denoiser and for each condition (i.e. noise type)! If the practitioner wants to denoise a single noisy image, it is not worth it and it would be much faster to test them all on the input noisy image and choose visually which output is best.
  3. The comparisons are not always convincing. For example, it is clear that the best overall denoiser is a poor baseline if the noise types considered are very different. In Figure 2 and 9, the best overall denoiser is TV for example, which is one of the simplest (and least efficient) denoiser, its only quality being that it is versatile. By the way, I am surprised that a denoiser such as SCUNet is not the best overall denoiser. Do you have explanations? I suspect that vertical banding noise and other very specific noises, which may not be present in the training set of SCUNet (but coud be added), weights on the overall performance. Moreover, comparisons could have be made with self-supervised techniques such as blind-spot networks (noise2self, noise2void), neighbor2neighbor, self2self, noise2vst, etc, when the noise is not spatially correlated, and AP-BSN, PUCA, etc, when it is.

Minor:

  1. Figure 2 is not very clear and coud be improved.
  2. Typo line 228: ImageNette -> ImageNet
Questions for authors
  1. The descriptor D13:Rn→R13D\small 13 : \mathbb{R}^n \to \mathbb{R}^{13} maps the input noisy image to a vector of size 1313 by computing many arbitrary statistics, such as the global variance of the horizontal/vertical differences, or the spatial correlation. Have you thought about learning this descriptor?
  2. I don't understand the difference between specialist and general-purpose denoisers in Table 3. Why is S_SwinIR not considered to be a specialist model since it was trained on Gaussian noise with σ=15\sigma=15. Same remark holds for DRUNet where the value of σ\sigma needs to be provided. Moreover, M_gaussian_25, S DRUNet_s25 and S_SwinIR_s25 are placed in three different pools, it doesn't make sense to me as they are all supposed to be dedicated to handle Gaussian noise with σ=25\sigma=25. By the way, what is the point in distinguishing between Classical/Specialist/General-purpose and Hybrid candidates? A denoiser is a denoiser in the end, isn't it?
  3. What if the noise type of the single-noisy image is different from the ones that are expected, i.e. the ones described in Table 1, and that no clean/noisy pairs are available so that you cannot add references of the new noise? What about Poisson-Gaussian noise for example or compression noise (for example JPEG artifacts)?
3 comments
AuthorsOpenReview ↗

Response part 1

Thank you for your detailed review. We have uploaded a revised PDF in which new text is marked in blue. On why we split the denoisers into separate pools when a denoiser is a denoiser in the end, we fully agree, and the revision now reports results for all 36 routed as a single set (Section 4.2 and Appendix E). With all of them combined, the best overall denoiser is SCUNet (as you expected), and our method still gains 1.20 dB over it with one denoising call per image. On whether the descriptor could be learned, we trained a CNN on the image pixels with the same labels, and it performed about as well as D13 without improving on it, while D13 requires no training (Appendix L).

We also understand why the paper read as a set of basic statistics combined with a nearest-neighbor lookup, since the results were spread across four pools and many sub-cases. However, the paper shows that this simple procedure chooses better than seven methods that run every denoiser and score the outputs, and the new experiments show that a learned router does not improve on it. It works because each calibration image is labeled with the denoiser that serves it best rather than with its corruption, so the denoiser can be chosen without identifying the noise. Since the learned and handcrafted descriptors perform comparably, the handcrafted one has a practical advantage: each statistic measures a known property of the noise, so when selection fails on a new type of noise one can see which property is missing and add a statistic for it, without retraining any denoiser. We also clarify below that calibration is done once for a given set of denoisers rather than once per image. We think the paper is clearer and stronger as a result of your review and, in light of the new results below, we hope you will give it another look.

W1: originality and sophistication. We kept the method simple on purpose, and we think the main result is more notable for it. With only a few basic statistics of the noise, computed on the noisy image, and a single denoiser run, it picked better denoisers than seven blind quality criteria (NIQE, BRISQUE, MUSIQ, CLIP-IQA+, TOPIQ, ARNIQA, and MC-SURE). Each of those has to run all the denoisers on the image and compare the outputs. We also tried learned alternatives trained on the same calibration set, and they did not do better. At no point is any denoiser retrained or fine-tuned.

W2: cost of calibration. You are right that calibration means running every denoiser on the noisy/clean pairs. This happens once for a given set of denoisers, and the result is shared with the code, so a practitioner with a single noisy image only computes its descriptor, which takes about 23 ms on a CPU, and runs one denoiser. The calibration set can also grow over time. Anyone who adds clean/noisy pairs of a new noise type, or a new denoiser, needs to run only the new combinations, and once the addition is in the shared repository, every user benefits from it without retraining anything. On real camera noise, adding such pairs recovered most of the loss without affecting results on the noise types already covered. Testing all the denoisers and choosing visually, by contrast, means running every one of them on every image. The blind quality criteria above are essentially automated versions of that choice, and our method did better than all of them while running only one denoiser.

W3: the best overall denoiser, SCUNet, and self-supervised methods. We agree that the best overall denoiser is a weak baseline when the noise types are very different. We report it because it is what a practitioner would run without knowing the noise type, and the blind quality criteria above are the stronger comparison. TV comes out as the best overall denoiser only among the classical filters, which is the set shown in Figures 2 and 9 (8 and 9 in the revised version). With all 36 denoisers combined, it is SCUNet (as you suspected), and your explanation holds for impulse noise. SCUNet is about 16 dB behind the best denoiser on impulse noise, and that family alone accounts for half of our 1.20 dB gain over it, with mixtures contributing another 30%. Banding contributes only about 6%, although SCUNet is 8 dB behind there as well. On the calibration images our banding network beats the notch filter, so banding images are sent to it, but it was trained on horizontal stripes at a single period and does poorly on the vertical and longer-period stripes in the test set, where the notch filter does well. A denoiser that is best on the calibration images but fragile on noise that differs from them can therefore reduce the gain, and we will discuss this in the revision.

Self-supervised denoisers would fit in as further candidates, and we will add pretrained blind-spot nets (e.g. AP-BSN) in the final version. Methods that train on the noisy image itself, such as Self2Self, would need a training run for every image, so they fall outside the setting we consider.

AuthorsOpenReview ↗

Response part 2

Q1: learning the descriptor. Yes, and we have now tried it. We trained a CNN on the noisy image pixels, with the same calibration set and labels, to predict the calibration condition, from which the best denoiser is read as in our method. It did about as well as D13 but not better (Table A below and Appendix L). It also has to be retrained whenever a new noise type is added, while D13 only needs some clean/noisy pairs of the new noise.

Table A. Learned alternatives trained on the same data and labels. Routing gain (dB) over the best overall denoiser on the 26 withheld conditions. For each learned alternative, the table gives its best variant in each pool, chosen after the test results were seen, which favors the alternatives. An asterisk marks a significant difference from D13 with nearest-neighbor routing (95% paired interval excluding zero).

MethodClassicalSpecialistPretrainedHybridAll 36
D13, nearest neighbor (ours)+1.36+1.58+0.14+0.99+1.20
Learned router on D13 (best of 8 variants)+1.36+1.56+0.05+0.85+1.10
CNN on image pixels+1.29+1.53−0.11*+0.97+1.04*

Q2: the distinction between pools. We agree that a denoiser is a denoiser, and the revision now reports results for all 36 as a single set (Section 4.2 and Appendix E). We kept the separate pools because they show that selection helps most when the denoisers have different strengths. “General-purpose” was also a poor name for networks like DRUNet and SwinIR, which in our pool are run at a single noise level, and we now call them “pretrained.”

Q3: an unexpected noise type, with no clean/noisy pairs. If there are no clean/noisy pairs for a new noise type, we cannot expect the method to choose well for it, but it can often detect the situation before any denoiser is run. On real camera images from SIDD, the descriptor flagged 90% of them as unlike anything in the calibration set, and those images can simply be sent to the best overall denoiser. A noise-type classifier gives no such warning. The one we reimplemented assigned images of unfamiliar noise types to one of its four classes with a mean confidence of 0.93. For the revision, we tried the two cases you mention (Appendix O and Figure 10). (i) Poisson–Gaussian noise needed no new pairs. The descriptor placed these images next to the Poisson conditions they combine, so only 0.5% of them were flagged. With our specialist networks, routing sent them to the Poisson specialists and gained 3.93 dB, which is 91% of what a perfect per-image choice would achieve. When SCUNet was among the candidates, there was almost nothing to gain over it, and routing lost 0.2 to 0.3 dB. (ii) JPEG is closer to the situation you described. The descriptor flagged 45% of the JPEG images, against 7% of the withheld synthetic images, and without JPEG pairs routing fell 2.04 dB below the best overall denoiser. The flag alone is therefore not enough for JPEG, and labeled examples are needed. We then added labeled JPEG examples at each quality level and one denoiser designed for compression artifacts (FBCNN), without retraining anything. Routing then gained 0.52 dB over the best overall denoiser, 67% of what a perfect per-image choice would achieve, and 0.46 dB at a quality level that had no examples of its own (Appendix O).

Minor points. We have moved Figure 2 to Appendix J and are redrawing it with all 36 denoisers, for which the best overall denoiser is SCUNet. ImageNette is the ten-class subset of ImageNet (Howard, 2019), and we will describe it where it first appears.

In summary. You were concerned that the method is too simple, that calibration is too costly to be worth it, and that the best overall denoiser is a weak baseline. On simplicity, a few statistics of the noisy image and a single denoiser run chose better than blind quality criteria that test all the denoisers and compare the outputs, which is the automated version of choosing visually. Calibration is done once and shared with the code, and anyone can extend it with a new noise type or denoiser for all users, so a practitioner pays only for the descriptor and one denoiser per image.

Your questions also led to new results. A learned descriptor did not do better than D13, all 36 denoisers are now also routed as one set with SCUNet as the best overall denoiser, and we tested the JPEG and Poisson-Gaussian cases you raised. We have also rewritten the abstract and introduction so that the main result is stated plainly, and we would be grateful if you could take another look with these changes in mind.

I sincerely thank the authors for their response and appreciate the substantial effort they have made to address my concerns. Nevertheless, despite these efforts, I still do not think that the paper meets the acceptance bar for ICLR. As I noted in my initial review, I am not enthusiastic about the core idea, which I find limited in terms of contribution and novelty. For this reason, I will maintain my original rating.

Reviewer_sES7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores blind denoiser selection. A pool of frozen denoisers is prepared, and noise space is calibrated by which denoiser is preferred to map a new image into this space. No retraining is needed.

Critical strengths
  1. The problem domain is relevant to the community.
Critical weaknesses
  1. Without a clear comparison against established baselines or existing frameworks, it is difficult for me to properly assess the significance and impact of the findings.
Questions for authors
  1. Could the authors briefly clarify the main novel contributions compared to existing work?
  2. What are the primary baseline methods or perspectives that this approach should be directly contrasted with?
2 comments

We have uploaded a revised PDF in which new and changed text is marked in blue.

Thank you for your review. Since your main concern is the comparison with established baselines, we list them here with their location in the paper, as the submitted version may have made them hard to find.

Main contribution. Given a pool of existing denoisers, we choose one for each image from the noisy image alone, before running any of them, and without retraining. Prior methods either run every denoiser and score the outputs without a clean reference, or first classify the noise type and apply the denoiser matched to it. We instead label each calibration image with the denoiser that works best on it, so a new image only needs to be matched to images served well by the same denoiser, and its noise does not need to be identified. With one denoising call per image, this recovers 43 to 66% of the improvement available from a perfect per-image choice (on noise conditions withheld from calibration).

Baselines.

  • The best overall denoiser, chosen on the calibration set and run on every image. All gains in the paper are measured against it.
  • Seven selectors that run every denoiser and keep the output with the best score: NIQE, BRISQUE, MUSIQ, CLIP-IQA+, TOPIQ, ARNIQA, and Monte-Carlo SURE (Section 4.2, Figure 3, Appendix G). Routing outperformed all seven in every pool while running one denoiser instead of all of them.
  • The noise-type classifier of Liu et al. (2020), reimplemented and trained on the same data (Section 4.2, Appendix K). Routing outperformed it, mainly on images containing two kinds of noise.
  • New in the revision: learned routers trained on the same labels, both on our descriptor and as a CNN on the image pixels (Appendix L). Neither improved on our method, which requires no training.
  • New in the revision: MoCE-IR (Zamfir et al., 2025), a mixture-of-experts restoration network, used with its published weights (Appendix L). Applied to every image, it was 2.77 dB below routing over all 36 denoisers.

If there is a specific baseline you think is missing, we would be glad to add it.

Thank you for your response. While I appreciate the clarifications, some concerns regarding the novelty of the work remain. Thus, I am inclined to maintain my original rating.

Reviewer_eBVGOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies blind selection of a denoiser from a fixed pool of frozen candidates. It proposes D13, a 13-dimensional handcrafted descriptor of noise properties, and performs nearest-neighbor routing using calibration images labeled by their best-performing denoiser. The central idea is that exact corruption identification may be unnecessary as long as degradations requiring similar treatments are nearby in the descriptor space. Experiments across several denoiser pools and synthetic/real noise settings show that this simple routing strategy can outperform using a single best-overall denoiser and several no-reference selection baselines, although the improvement depends strongly on the candidate pool.

Critical strengths
  1. Clear problem formulation. The paper presents a simple and easy-to-understand formulation of blind denoiser selection from a fixed candidate pool.
  2. Fairly thorough empirical evaluation. The paper evaluates several types of denoiser pools and reports both successful and failure cases rather than only favorable results.
Critical weaknesses
  1. Limited methodological insight. The method mainly collects several plausible noise statistics and combines them with nearest-neighbor routing. It is unclear what new principle is learned about which degradation properties actually determine denoiser choice, beyond showing that a sufficiently informative handcrafted feature set can work.

  2. Unclear generality due to method–benchmark coupling. D13 is designed around the same corruption properties used to construct the benchmark, while the “unseen” tests mostly stay within those known families. This makes it hard to tell whether the paper discovers a general treatment space, or just an engineered representation for an anticipated synthetic corruption space.

Questions for authors
  1. Why these 13 statistics? Which features actually matter for choosing the denoiser? Could a much smaller subset achieve the same result? The paper says these statistics were chosen to cover the corruption properties used in the benchmark, but it is still unclear what the essential treatment-related factors are.

  2. What is the general scientific message behind D13? Beyond collecting many plausible noise statistics, what have we actually learned about which degradation properties determine the best denoiser?

  3. Why not compare against learned routing / recent MoE methods? Since the calibration references are already labeled by the best denoiser, a learned router using exactly the same data and supervision seems like a very natural baseline. This is especially relevant given recent degradation-aware expert-selection and MoE methods in image restoration, such as MEASNet (Yu et al., 2025) and MoCE-IR (Zamfir et al., 2025). Without such comparisons, it is hard to tell whether D13 itself contributes something special beyond the general benefit of routing

  4. What practical problem really needs this routing capability? The current setting mainly selects one denoiser for an entire image from a fixed zoo. If the method already has the ability to decide which denoiser is appropriate, could this be demonstrated in a setting where dynamic decisions are genuinely useful, such as PnP or other adaptive restoration pipelines, rather than only whole-image model selection?

4 comments
AuthorsOpenReview ↗

Response part 1

We have uploaded a revised PDF in which new and changed text is marked in blue.

Thank you for suggesting a learned router trained on exactly the same data and supervision. We had wondered about this comparison ourselves, so we ran it, both on the D13 statistics and as a CNN on the image pixels. Neither did better than nearest-neighbor routing with D13, which requires no training at all. We found the result clarifying, because it answers your question about the general message of the work more directly than the paper did. As your summary put it, exact corruption identification is unnecessary as long as degradations requiring similar treatments are nearby. The gain comes from labeling the calibration images by their best-performing denoiser rather than by their corruption, and once they are labeled this way, a simple handcrafted descriptor is enough. We have also added the ablation you asked for and a test on corruptions the descriptor was not designed for, and we would value your view on both.

Your concern about method–benchmark coupling is fair, and we address it more directly in the revision. The paper already had one test outside the designed families. Real camera noise played no part in the design of D13, and routing on SIDD recovered most of its loss once a few hundred labeled examples of that noise were added, without retraining any denoiser. We have now also tested JPEG compression and Poisson–Gaussian noise, neither of which was used to design D13. For JPEG, labeled examples together with one new denoiser restored the gain without retraining, and the descriptor placed Poisson–Gaussian next to the Poisson noise it resembles. The descriptor does not need to anticipate every corruption, as long as a new one can be covered by adding labeled examples. We also agree that the paper was harder to read than it should have been. We have therefore rewritten the abstract and introduction around the main claim, and added results for all 36 denoisers routed as a single pool (Section 4.2 and Appendix E).

W1 and Q2: what we learn about which degradation properties determine denoiser choice. The main thing we learned is that choosing a denoiser is a coarser problem than identifying the corruption. In D13, an image’s nearest neighbor shares its best denoiser 83 to 89% of the time, well above the 65 to 69% expected by chance, but shares its corruption only 60% of the time. Which properties matter then depends on the pool. A property determines the choice only when the candidates treat it differently. Banding orientation is a good example. D13 does not record which way the stripes run, and this costs nothing with classical filters, because the notch filter removes stripes in any direction, but it costs most of the available gain with our specialist networks, since the only banding network was trained on horizontal stripes. Because each statistic measures a known property, a failure of this kind can be traced to the missing property and fixed by adding a statistic, without retraining any denoiser. So the essential treatment-related factors are a property of the pool as much as of the noise, and the learned routers above suggest that the descriptor does not need to be learned once the labels refer to the best denoiser.

AuthorsOpenReview ↗

Response part 2

W2: method–benchmark coupling. We agree that the withheld synthetic conditions stay within the ten families, and the paper says so (Appendix F). That is why the real-noise experiment matters to us. On SIDD, with only synthetic calibration images, routing fell below the best overall denoiser, as one would expect. Before any denoiser was run, the descriptor flagged 90% of the SIDD images as unlike anything in the calibration set, against 7% of the withheld synthetic images. The method can therefore tell when its own choice should not be trusted, which a noise-type classifier cannot do. The one we reimplemented assigned images of unfamiliar noise types to one of its four classes with a mean confidence of 0.93. Adding 320 labeled real images then recovered most of the loss, while changing the synthetic results by at most 0.04 dB. We have now also tested JPEG compression and Poisson-Gaussian noise, neither of which was used to design the descriptor (Appendix O). The two behaved quite differently in the D13 space. The descriptor flagged 45% of the JPEG images as unfamiliar, against 7% of the withheld synthetic images, and without JPEG examples routing fell 2.04 dB below the best overall denoiser. After we added labeled JPEG examples and one new candidate designed for compression artifacts (FBCNN), without retraining anything, routing gained 0.52 dB over that denoiser, 67% of what a perfect per-image choice would achieve. The added examples lowered the gain on the withheld synthetic conditions by 0.26 dB. Poisson–Gaussian noise, on the other hand, landed on the Poisson calibration conditions it combines, and only 0.5% of these images were flagged. With our specialist networks, routing sent them to the Poisson specialists and gained 3.93 dB, 91% of the available gain. When SCUNet was the best overall denoiser, a perfect per-image choice would have gained only about 0.1 dB, and routing lost 0.2 to 0.3 dB. The new Figure 10 in Appendix O shows where both fall in the D13 space, together with the SIDD images.

Q1: why these 13 statistics, and could fewer suffice. The statistics measure four properties of the noise: its amplitude, the shape of its distribution, how it depends on the signal, and its spatial structure. We removed each group in turn and repeated the routing experiment, splitting spatial structure into the four statistics that describe the shape of the spectrum and the five that describe correlation (the lag-one autocorrelation and four radial bands).

The shape of the distribution matters most. Without skewness and kurtosis, routing loses 0.8 to 1.5 dB depending on the pool, mostly on impulse noise, mixtures, and additive noise. Amplitude and signal dependence each contribute between 0.02 and 0.4 dB. The spectral-shape statistics contribute only 0.13 to 0.24 dB on average, but that average conceals a large effect on a few conditions. Without them, routing loses 0.8 to 2.0 dB on annular noise in every pool and about 2 dB on banding in the classical and hybrid pools. Annular noise is only 2 of the 26 test conditions, so a 2 dB loss there moves the average by about 0.15 dB.

On whether fewer would suffice, the answer on this benchmark is partly yes. Removing the five correlation statistics changes the gain by between −0.07 and +0.14 dB, and no family loses more than 0.3 dB. They are computed on a residual that attenuates the low frequencies, where correlated noise has most of its energy, which may explain why they contribute little. We keep them, because this variant was identified on the test conditions and its effect changes sign between pools. A much smaller descriptor does lose accuracy. Keeping only the four statistics of amplitude, distribution shape, and signal dependence loses 0.17 to 0.40 dB in every pool. Tables A and B below summarize the results, and Appendix M gives the full analysis.

AuthorsOpenReview ↗

Response part 3

Table A. Effect of removing each group of statistics. Change in routing gain (dB) on the 26 withheld conditions, relative to the full descriptor with 95% intervals. The first row gives the gain of the full descriptor over the best overall denoiser. An asterisk marks an interval that excludes zero.

VariantClassicalSpecialistPretrainedHybridAll 36
Full D13 (gain)+1.36+1.58+0.14+0.99+1.20
Remove marginal law (2)−0.77* [−0.92, −0.65]−1.11* [−1.20, −1.01]−0.95* [−1.05, −0.85]−1.35* [−1.47, −1.23]−1.53* [−1.66, −1.39]
Remove amplitude (1)−0.06 [−0.13, 0.01]−0.39* [−0.48, −0.31]−0.23* [−0.34, −0.12]−0.32* [−0.44, −0.21]−0.32* [−0.45, −0.19]
Remove spectral structure (4)−0.15* [−0.22, −0.07]−0.13* [−0.23, −0.04]−0.22* [−0.33, −0.10]−0.24* [−0.37, −0.06]−0.18* [−0.30, −0.05]
Remove signal dependence (1)−0.09* [−0.14, −0.04]−0.17* [−0.21, −0.13]−0.02 [−0.07, 0.03]−0.02 [−0.07, 0.04]−0.02 [−0.08, 0.04]
Remove spatial correlation (5)+0.03 [−0.01, 0.08]−0.07* [−0.13, −0.02]+0.06 [−0.00, 0.12]+0.14* [0.06, 0.21]−0.03 [−0.10, 0.03]
Keep amplitude, marginal law, and signal dependence only (4)−0.23* [−0.31, −0.14]−0.40* [−0.53, −0.28]−0.19* [−0.27, −0.10]−0.17* [−0.29, −0.02]−0.35* [−0.48, −0.23]
Keep spectral structure and spatial correlation only (9)−0.95* [−1.07, −0.84]−1.68* [−1.79, −1.57]−1.16* [−1.29, −1.04]−1.66* [−1.80, −1.52]−1.75* [−1.90, −1.60]

The number after each variant is the number of statistics removed or kept. “Pretrained” was the “general-purpose” pool.

Table B. Effect of removing each group, by noise family. Change in the routing gain of each family (dB) when the group is removed. Intervals are omitted because some families contain a single condition.

Group removedFamilyClassicalSpecialistPretrainedHybridAll 36
Marginal lawImpulse−3.26−7.50−2.62−7.88−9.38
Marginal lawMixture−1.03−1.06−1.13−1.32−1.53
Marginal lawAdditive−0.38−1.73−2.38−3.37−2.81
Spectral structureBanding−1.97+0.13−0.19−2.24+0.13
Spectral structureAnnular−0.77−1.98−1.36−1.30−2.00
Spectral structureAnisotropic−0.17+0.38−1.02−1.01−0.36
Spectral structureCorrelated−0.31−0.94−0.47−0.44−0.93
Signal dependenceSignal-dependent+0.12−0.40−0.68−0.71−0.83
Spatial correlationCorrelated+0.29+0.83+0.40+0.39+0.92
Spatial correlationBanding+0.98+0.03+0.06+1.11+0.03
Spatial correlationAnnular−0.10−0.30+0.05+0.05−0.22

Q3: learned routing and MoE methods. Our method can be read as a mixture of experts whose router is a nearest-neighbor lookup, so the learned routers above are also learned gating functions (Table C below and Appendix L). MoCE-IR, run with its published weights (trained on Gaussian noise, rain, and haze), reaches 24.91 dB on the 26 withheld conditions, 2.77 dB below routing over all 36 denoisers (95% interval 2.61 to 2.93) and 1.57 dB below the best overall denoiser. Adding it as a 37th candidate requires only running it once on the calibration images. Routing then gains 1.26 dB instead of 1.20 (difference 0.06 dB, interval 0.02 to 0.11), and MoCE-IR is selected for 34% of images. MEASNet has no public code or weights. Both methods train their experts jointly with the router on a fixed set of degradations, so adapting them to our noise types would mean retraining their experts, whereas we route among frozen denoisers developed independently.

Table C. Learned alternatives trained on the same data and labels. Routing gain (dB) over the best overall denoiser on the 26 withheld conditions. For each learned alternative, the table gives its best variant in each pool, chosen after the test results were seen, which favors the alternatives. An asterisk marks a significant difference from D13 with nearest-neighbor routing (95% paired interval excluding zero).

MethodClassicalSpecialistPretrainedHybridAll 36
D13, nearest neighbor (ours)+1.36+1.58+0.14+0.99+1.20
Learned router on D13 (best of 8 variants)+1.36+1.56+0.05+0.85+1.10
CNN on image pixels+1.29+1.53−0.11*+0.97+1.04*
AuthorsOpenReview ↗

Response part 4

Q4: settings where dynamic decisions are useful. The setting we address is common in practice: an image of unknown origin, several denoisers each suited to a narrow range of noise, and no clean reference. The mixture results suggest the most useful extension. An image containing two kinds of noise would be better served by routing each region separately, and the method applies to regions with little change. Choosing a prior at each iteration of a plug-and-play method is another natural use, which we will mention as future work.

In summary. You asked what principle the paper offers beyond a collection of handcrafted statistics, whether its generality is limited by the coupling between descriptor and benchmark, and whether a learned router would do just as well. The principle is the one in your own summary of the paper: denoiser choice depends on which degradations need similar treatment, which is coarser than which corruption is present. The learned routers you proposed confirm this, since with the same supervision they do no better than nearest-neighbor routing on D13. On generality, the descriptor flags real camera noise as unfamiliar before any denoiser is run, routing on JPEG artifacts is restored by adding labeled examples and one new denoiser without retraining, and Poisson–Gaussian noise lands next to the Poisson conditions it combines, although D13 was designed for none of them. The ablation shows which properties determine the choice of denoiser, and for which corruptions. The shape of the noise distribution decides impulse noise and mixtures, and the spectral statistics decide annular noise in every pool. Because each property is tied to specific corruptions, a routing failure can be traced to a missing property and repaired by adding a statistic, without retraining any denoiser. We have also rewritten the abstract and introduction so that this principle is stated plainly, and we would be grateful if you could reassess the manuscript in light of the new results.

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

Related papers

Discussion (0)

Sign in to comment.