acceptodds
Withdrawn from ICLR 2027

PET-DSCFormer: Efficient Degradation-aware Transformer for Real-world Image Denoising

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image denoising is a fundamental task aiming to recover clean images from real-world degradations while preserving structural fidelity and fine-grained details. Recent CNN- and Transformer-based restoration methods have achieved remarkable progress by improving local representation learning and global context modeling. However, existing approaches still face three major limitations: dense Transformer attention is computationally expensive for high-resolution restoration, conventional skip fusion may propagate unreliable degraded features into the decoder, and most methods predict restoration residuals in a deterministic manner without explicitly modeling the uncertainty and complementarity of multi-scale residual estimates. To address these limitations, we propose PET-DSCFormer, a posterior error transport restoration framework for real-world image denoising and general image restoration. Extensive experiments on multiple public real-world denoising datasets, including medical imaging and natural image benchmarks, demonstrate that PET-DSCFormer achieves competitive or superior restoration performance compared with state-of-the-art methods. Moreover, experiments on GoPro deblurring further validate the generalization capability of the proposed framework beyond denoising. The code will be publicly available on GitHub upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Denoising, Image Restoration, Degradation-aware Transformer

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ek8uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PET-DSCFormer. DSCA aggregates dense features into a small set of structural tokens through degradation-weighted soft assignment and performs attention on these tokens. PSF fuses encoder and decoder features using predicted precision, while PET combines multi-scale residuals through a Gaussian product formulation. HDRE and MPC further enhance and calibrate local details. Experiments on medical, natural image, RAW, and low-light datasets validate the method’s effectiveness, demonstrating a favorable balance between restoration accuracy and resource efficiency.

Critical strengths
  1. The paper introduces dedicated modules for global feature aggregation, cross-layer feature fusion, and multi-scale residual restoration. The overall design is coherent, with reliability serving as a unifying theme that connects DSCA’s degradation-aware global modeling, PSF’s precision-weighted feature fusion, and PET’s multi-scale residual fusion. HDRE and MPC further enhance local detail restoration, giving each module a clearly defined role.

  2. The evaluation covers diverse restoration scenarios, including medical images, natural images, RAW images, low-light enhancement, and motion deblurring, helping assess the method’s applicability across different data domains. The method also offers favorable inference time and GPU memory usage, suggesting potential value for practical deployment.

Critical weaknesses
  1. The paper contains some grammatical errors and inaccurate wording that could be further improved. For example, “experiment. we manually inject” -> “experiment. We manually inject.” The title of Table 4 is “Efficiency comparison (PSNR/SSIM),” whereas the table reports parameter count, GFLOPs, memory usage, and runtime. In Section 3.4, μ₁ is described as “the local residual distribution,” but according to the definitions in Eqs. (19)–(23), μ₁ represents the fused residual mean estimate.

  2. DSCA predicts a degradation map d using a lightweight convolutional branch and converts it into confidence weights for weighted coreset token aggregation. However, the paper does not specify whether d receives any explicit supervision or constraints. Although randomizing or shuffling the degradation map leads to performance degradation in the ablation study, this may only indicate that the learned weights exhibit spatial structure, rather than establish that d reflects the true degradation severity.

Questions for authors
  1. The efficiency evaluation in Table 4 requires further clarification. Inference time is reported in seconds, with 60 s listed for the proposed method. Is this a labeling error, or does it represent the total runtime over the entire test set?

  2. In Table 5, replacing the baseline component with DSCA increases GFLOPs from 133.37 to 146.51. Adding the calibration module further increases GFLOPs from 145.16 to 174.79, an increase of approximately 20.4%, while the corresponding PSNR gains on LDCT-Chest and PMRID are only 0.02 and 0.10 dB, respectively. Could the authors further clarify the trade-offs among restoration accuracy, computational cost, memory usage, and runtime?

  3. Figure 7 illustrates differences in expert outputs and their spatial weighting preferences, but it remains unclear whether the routing preferences align with the experts’ local restoration strengths. Could the authors validate this alignment through region-level analysis and statistics across multiple samples?

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

This paper proposes an efficient Transformer for real-world image denoising. Specifically, the authors introduce several modules, including degradation-aware structured Coreset Attention (DSCA), Precision-aware Skip Fusion (PSF), and Posterior Error Transport (PET), to achieve an efficient design. Experiments on several natural images and medical denoising datasets demonstrate competitive performance compared to the existing works.

Critical strengths
  1. The authors propose mutually complementary modules centered around global feature aggregation, encoder-decoder feature fusion, and multi-scale residual prediction in image restoration, resulting in a clear overall framework.
  2. Experiments are conducted on CT, MRI, X-ray, natural image denoising, and GoPro deblurring datasets, reporting competitive PSNR/SSIM results.
Critical weaknesses
  1. The organization of the paper, especially the mathematical formulation, needs substantial improvement. First, there are 38 equations in the main text, making the methodology section really hard to follow. Besides, it contains only one figure, which has a significant layout issue in the MPC module. No other figures like quantitative comparison is provided, which is quite important for a low-level vision paper.
  2. Despite emphasizing degradation-aware denoising, the model introduces no explicit noise distribution modeling or noise-related supervision constraints. The so-called degradation awareness is primarily acquired through internal network learning, making the method essentially closer to a general image restoration architecture rather than a dedicated method designed for denoising tasks. If its core contribution is task-agnostic, why has it not been systematically validated on more image restoration tasks? Currently, only GoPro deblurring experiments are provided, and the PSNR improvement is merely 0.04 dB compared to NAFNet, which hardly suffices to support the generality of the method.
  3. Although the paper highlights efficient attention as one of its main contributions, he inference time of the proposed method is significantly higher than U-Net based method. While the authors demonstrate the theoretical complexity advantage of coreset attention over standard dense spatial attention, this does not imply higher efficiency in practical image restoration networks.
Questions for authors
  1. Lack of validation regarding the rationality of probabilistic modeling and the accuracy of predicted uncertainty. The paper predicts variance through the network, interprets it as the uncertainty of feature or residual estimates, and further utilizes Gaussian posterior for weighted fusion. However, the training process only employs the L1 reconstruction loss on the final image, without any explicit variance supervision, likelihood optimization, or probability calibration constraints. How do the authors ensure that the predicted variance truly reflects the estimation error?
Reviewer_x7mDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PET-DSCFormer, a U-shaped image restoration network that incorporates global attention, skip fusion, and residual prediction, with the core proposed modules named Degradation-aware Structural Coreset Attention (DSCA) and Precision-aware Skip Fusion (PSF). Experiments cover medical and natural image denoising, RAW restoration, and deblurring, demonstrates the performance of PET-DSCFormer across tasks.

Critical strengths

The performance across tasks is overall promising.

Critical weaknesses

If the paper provides more evidence of the motivation about the design of core modules in the model, it would be more convincing instead of purely promising results. I think the modification of modules based on previous methods is not novel. And I would like to see more results of the model dealing with cases in real world, it is more challenging but very meaningfull. The results on benchmarks may be far from real-world evaluations.

Questions for authors

See the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.