acceptodds
Withdrawn from ICLR 2027

Closing Degradation Loop: A Structured Inference-Restoration-Verification Framework for All-in-One Infrared Image Restoration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

All-in-one thermal infrared image restoration recovers clean images from observations with unknown single or compound degradations. Existing methods mainly use predicted degradation cues as one-way prompts or routing signals, without explicitly reasoning about how degradations form the observation or verifying whether restoration is consistent with that process. We propose SIRV, a Structured Inference-Restoration-Verification framework, to close this loop. Specifically, structured degradation inference first estimates a posterior over degradation states, retaining alternatives to avoid premature decisions and provide composition and severity cues. Then, posterior-conditioned restoration integrates cues from the inferred degradation posterior to modulate a shared restorer, adapting restoration to degradation composition, strength, and spatial distribution. Finally, degradation-consistent verification reuses sampled degradation parameters from the posterior in a differentiable redegradation model to reconstruct the input observation, penalizing restorations inconsistent with it. Together, SIRV transforms one-way guidance into a coherent loop in which uncertain degradation explanations guide restoration and the resulting restorations are verified against the input during training, improving robustness to ambiguous compound degradations and preserving structural details. Experiments across infrared benchmarks demonstrate SIRV's state-of-the-art restoration performance and strong generalization. On M\(^3\)FD, SIRV achieves 28.329 dB PSNR and 0.918 SSIM, exceeding the best previous scores by 2.302 dB and 0.038, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
all-in-one image restoration, thermal infrared imaging, structured degradation inference, posterior-conditioned restoration, degradation-consistent verification

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_Dp6WOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SIRV to restore thermal infrared images affected by unknown single or composite degradations. The framework predicts posterior distributions over degradation sequences and parameters to guide a shared restoration network. During training, a differentiable re-degradation process enforces observation consistency. Experiments cover synthetic degradations, cross-dataset transfer, real-world image enhancement, and downstream object detection.

Critical strengths
  1. The framework is clearly designed and well motivated. By combining degradation inference, conditional restoration, and re-degradation consistency, the method provides a practical way to improve infrared image restoration using knowledge of the degradation process.

  2. The overall experimental results are competitive. The method achieves the best average PSNR/SSIM on three paired test datasets. Ablation studies demonstrate the effectiveness of conditional information and re-degradation constraints, while the model contains only about 3.6M parameters.

Critical weaknesses
  1. Insufficient evidence for the probabilistic contribution. The ablations lack a capacity- and supervision-matched deterministic predictor using the same consistency loss, and do not isolate the contribution of coupled sampling. Consequently, the gains cannot be clearly attributed to posterior modeling rather than additional supervision and consistency regularization.

  2. Gap between the variational formulation and implementation. Equation (3) averages over the inferred sequence posterior, whereas training evaluates the parameter, restoration, and consistency terms only on the ground-truth branch and uses the restoration mean. Although these approximations are acknowledged, their effect on the stated variational bound is not characterized.

  3. Limited validation of uncertainty modeling. The factorized Beta posterior uses a fixed concentration, limiting its ability to independently adapt uncertainty to individual samples or represent parameter correlations. Without calibration or coverage analyses, it remains unclear whether the predicted distributions provide reliable uncertainty estimates.

  4. Unclear identifiability of degradation order. Under the specified generator, affine contrast adjustment and normalized Gaussian blur commute before quantization, so distinguishing CB from BC may primarily reflect quantization effects. An analysis of equivalent or nearly equivalent sequences is needed to justify interpreting the predicted order as a physical explanation.

  5. Insufficient validation of the forward-model approximations. The diagonal-Gaussian verifier omits spatial correlations, multiplicative-noise interaction terms, and moment corrections for clipping and quantization. Although these simplifications are disclosed, their impact and robustness to degradation-model mismatch remain untested.

  6. Incomplete evaluation of blind degradation estimation. The parameter diagnostics in Table 19 assume ground-truth sequences and therefore do not assess error propagation during fully blind inference. Kernel-size accuracy is also only 16.7%, compared with 17.6% for the majority-class baseline, motivating further analysis of parameter identifiability and its relationship to restoration quality.

  7. Errors and inconsistencies in presentation. Figures 3 and 7 contain clear presentation issues: the bottom-left panel in Fig. 3 shows only a heatmap instead of the grayscale–heatmap composite used elsewhere, while the DRNet noise result in Fig. 7 depicts the wrong scene. Other inconsistencies include labeling observation variance as μ~y\tilde{\mu}_y in Fig. 2(c) and describing the concatenation of two 64-dimensional descriptors as a 144-dimensional input in Appendix B.1.

Questions for authors

The same as the weaknesses.

Reviewer_Q1swOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a Structured Inference-Restoration-Verification (SIRV) framework for all-in-one infrared image restoration. They key idea is to ensure that the restored result from a model is consistent with the degradation process which created it. SIRV first estimates a posterior over degradation states which provides distributional information on the degradations present in the input image. The subsequent posterior conditioned restoration module integrates these cues to modulate the restored result based on the identified degradations. To teach the restoration module to achieve proper usage of the degradation information, the verification module re-degrades the restored result with the identified degradation parameters and supervises it with the original degraded image. The SIRV framework preserves structures in the input and achieves state-of-the-art results in all-in-one infrared image restoration.

Critical strengths
  1. The authors propose a principled solution to make the restoration model aware of the degradations present in the input image, allowing it to produce restored results that are consistent with the input image.

  2. The results demonstrate superior performance to existing approaches.

Critical weaknesses
  1. While the SIRV approach addresses structural distortions, it is not clear why it would favor robustness to ambiguous compound degradations? In-fact, this seems to be apparent from the quantitative results across multiple degradations (single, double and triple in tables 1 and 3). For instance, the gap with prior approaches reduces as the degradations compound, showing that the proposed approach seems to perform well for single degradations and not as well on multiple degradations. Furthermore, the authors choose handcrafted degradation models for representing the input degradations which may not align well with complex real-world artifacts.
Questions for authors
  1. Can the authors explain more about why SIRV improves robustness to complex degradations, given that the PSNR advantage over existing approaches reduces as the degradations increase?

  2. Can the authors provide a breakdown of the results in Table 9 by the number of compound degradations? This may be helpful to address the above question.

  3. Does the reliance on pre-defined degradation operators hurt performance on complex real-world degradations, as such degradations may not be captured by heuristic degradation operators? For instance, whats the performance of the "None" model in Table 9 for the TIIQAD dataset in Table 2?

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

SIRV predicts ordered infrared degradation hypotheses and their parameters, conditions a shared restorer on these estimates, and adds a differentiable redegradation loss during supervised training. Inference averages the two most probable restoration branches; evaluation covers synthetic compound degradations, cross-dataset transfer, authentic infrared images, and object detection.

Critical strengths

Results provide evidence that structured conditioning and observation consistency contribute beyond the shared restorer, with a 0.869 dB gain from verification on HM-TIR. The cross-dataset results and compact 3.6M-parameter model make this a useful empirical contribution.

Critical weaknesses
  1. The practical objective is not the cELBO written in Eq. 3 (Appendices A.2, C.1, and E). It replaces the sequence expectation with the ground-truth branch, substitutes the clean-image mean for its expectation, and normalizes only the image terms by pixel count. These choices are disclosed, but calling the resulting objective an estimator of the bound obscures its bias and the fact that alternative sequence hypotheses are not directly verified during training. This limits the probabilistic interpretation of the central loop, although it does not invalidate the empirical regularizer.

  2. The benefit specifically attributable to coupled posterior sampling remains unclear. Table 9 compares no consistency, L1 consistency, and Gaussian NLL, but does not separate a deterministic parameter-conditioned cycle from the shared-sample construction. Moreover, Appendix B.3 fixes the Beta concentration, and Table 19 mainly establishes prediction accuracy rather than calibrated parameter uncertainty. The evidence therefore supports structured conditioning more directly than the stronger uncertainty-based explanation.

Questions for authors
  1. Can the implemented loss be stated explicitly as a supervised surrogate, with its differences from Eq. 3 and the consequences for alternative branches made clear in the main text?
  2. Is there an existing comparison of deterministic posterior-mode conditioning against coupled sampling with the same restorer and verification loss? This would clarify which part of the probabilistic construction drives the gain.

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

Related papers

Discussion (0)

Sign in to comment.