acceptodds
Withdrawn from ICLR 2027

Separate to Restore: Haze-Guided Feature Exchange for Image Dehazing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Single-image dehazing requires preserving scene content while suppressing haze-induced degradation. Modeling these objectives separately can encourage specialized representations, but restoration also requires communication between them. We propose a dual-branch framework that models clean scene content and a signed haze residual through separate pathways coordinated by haze-guided feature exchange. Haar wavelet cues provide complementary frequency biases, guiding the content branch toward structural details and the haze branch toward low-frequency degradation cues without imposing a fixed decomposition. Multiscale gated exchanges enable selective cross-branch communication. At the bottleneck, a routing module combines image-derived haze priors with learned corrections, prior-consistency weighting, and a learnable scaling factor to regulate where and how strongly messages are exchanged. Paired additive and subtractive updates redistribute features between the branches, while clean-image supervision, residual supervision, and input reconstruction consistency jointly encourage task-specific specialization. Under a controlled reimplementation protocol, the framework achieves 27.13 dB PSNR on O-Haze, exceeding the strongest compared baseline by 0.75 dB, with a mean PSNR of 26.98 ± 0.21 dB across five training seeds. Ablations and branch interventions suggest partial functional specialization, while evaluations on three additional real-haze benchmarks reveal remaining limitations under dense and nonuniform haze.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Single-image dehazing, Dual-branch representation learning, Haze-guided feature exchange, Wavelet frequency priors.

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_hMgeOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Separate to Restore (STR), a dual-branch framework for image dehazing that separately models scene content and haze-related residuals. The proposed method incorporates Haar wavelet decomposition for frequency-based guidance and introduces a Haze-guided Feature Exchange Block (HFEB) to facilitate information exchange between the two branches. Additionally, a Haze Evidence Routing Block (HERB) is designed to guide feature interactions using haze-related priors. Experimental results on several dehazing benchmarks demonstrate the effectiveness of the proposed framework, particularly on O-HAZE.

Critical strengths

Clear motivation: This work provides a reasonable motivation for separating scene content and haze-related information into two branches to improve image dehazing.

Well-designed framework: The proposed method integrates frequency-based guidance, feature exchange, and haze-related priors into a unified framework. The overall architecture is clearly presented and easy to understand.

Critical weaknesses
  1. Limited performance across benchmarks: The PSNR results on Dense-HAZE, I-HAZE, and NH-HAZE are lower than those of previous works. Further analysis of these inconsistent dehazing results is needed.

  2. Limited novelty: This work seems to be a combination of existing techniques, including wavelet decomposition, frequency-based guidance, and gated feature fusion. A clearer description of the differences between this work and existing techniques should be added to the Introduction.

Questions for authors

This work assigns low-frequency information to the content path and high-frequency information to the haze path. In my opinion, haze patterns tend to be low-frequency due to their low contrast, while background patterns (content) tend to be high-frequency due to their complex details. Therefore, why does this work feed low-frequency information into the content path and high-frequency information into the haze path?

Reviewer_Go9xOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes STR, a dual-branch image dehazing framework that separately learns clean-image restoration and signed-residual prediction. Haar frequency information provides different initial biases for the two paths. HFEB performs gated feature exchange at multiple scales, while HERB controls information transfer at the bottleneck using image-derived haze cues and learned corrections.

Critical strengths

The clean-image and residual paths each have their own supervision targets, while feature exchange constrains the information flow between them, giving the overall design a clear task-specific rationale. The authors also provide repeated training runs, ablations with retraining, and branch-intervention analyses.

Critical weaknesses
  1. Wavelet-guided low-frequency restoration and detail processing have been studied in related work, so the novelty of this paper needs to be supported primarily by its exchange rule, prior-guided routing, and joint supervision. However, the current evidence does not sufficiently establish the independent advantages of these mechanisms over ordinary dual-branch auxiliary learning, nor does it explain why preserving the sum of features should lead to more reliable content preservation.

  2. The ranking supervision for HERB's haze evidence is not fully defined, and it remains unclear whether reasonable routing could be suppressed when the prior fails. Although the experiments include real haze and cross-domain tests, they lack sufficient comparisons with recent representative methods under matched settings.

  3. The cross-domain results also lack input-image references and matched baselines. Ablations of the key components do not include corresponding restoration results and close-up views, making it difficult to visually assess the practical effects of the different designs.

  4. The efficiency analysis on page 9 refers to Tables 8 and 9 on page 13, and some module names in the method diagram are inconsistent with those in the text, disrupting the reading flow and making the results harder to verify.

Questions for authors
  1. A contribution that warrants further discussion is the haze-evidence-modulated exchange mechanism that preserves the local sum of features. Could the authors compare it with ordinary gated mixing and dual-branch auxiliary learning while matching model capacity, output-head connectivity, and supervision targets, to clarify whether the improvements come from the exchange rule itself or from the network connections and supervision configuration?

  2. The authors need to fully define the ranking supervision used to train the haze evidence and explain its relationship to the learned corrections and prior-consistency constraint. When the image prior is incorrect in regions such as the sky, white objects, or areas of strong illumination, could routing that departs from the prior but is actually reasonable be suppressed?

  3. Could the authors add comparisons with recent representative methods under a unified training and model-selection protocol, and include input-image references and matched baselines in the existing cross-domain tests, to more fully validate competitiveness and generalization in real-world hazy scenes?

  4. Could the authors provide corresponding restoration results and close-up views for the ablations of key components, visually showing how the exchange mechanism, prior-guided routing, and joint supervision each affect low-frequency restoration, detail preservation, and artifact suppression?

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

This paper proposes STR (Separate to Restore), a dual-branch network for single-image dehazing. The method uses separate content and haze branches to learn scene content and haze-related features, respectively. It employs Haar wavelet decomposition to provide different frequency information and introduces the Haze Feature Exchange Block (HFEB) for cross-branch feature interaction. In addition, the authors propose the Haze Evidence Routing Block (HERB), which incorporates physical priors from existing dehazing models to control information exchange between the two branches. The authors validate their conclusions on several datasets.

Critical strengths

The authors provide relatively detailed ablation studies, including comparisons involving HERB and sum-preserving exchange, which provide preliminary evidence for the effectiveness of these modules.

Critical weaknesses

Lack of sufficient novelty and design motivation in the core method. The authors mainly build their network using existing techniques, such as dual-branch architectures, frequency decomposition, and cross-branch feature interaction. However, these ideas have already been widely explored in image dehazing, such as DW-GAN [1], Frequency and Content Dual Stream Network [2], and WDMamba [3]. Using existing techniques is not itself a problem, but the authors do not adequately explain what new insight their design offers compared with existing methods. In particular, the sum-preserving exchange in HFEB and the prior-consistency constraint in HERB lack clear motivation specific to the dehazing task. The paper mainly describes how the modules are implemented, without sufficiently explaining why dehazing requires these designs or what specific dehazing problems these minor differences from prior work can address. In my view, the current paper does not meet ICLR's bar.

Insufficient explanation and justification for some methodological designs. In Lines 46–51, the authors propose separately predicting a clean image (J) and a residual (R), with the constraint (J+R\approx I). However, considering that (R^=I-J^) is directly determined by the clean image and the input, the authors do not adequately explain why an additional residual prediction branch is needed or what advantages this design offers over directly predicting (J). Although Section 3.7 introduces residual supervision and reconstruction consistency losses, the paper lacks corresponding ablation experiments to demonstrate the necessity of these designs.

In addition, the combination weights of the three haze priors in Eq. (13), as well as the fixed parameters in Eqs. (15), (16), and (18), lack clear justification. In particular, HERB uses the difference between the learned haze evidence and the original prior as an indicator of reliability (Lines 257–262, Eqs. (16)–(17)), but the authors do not adequately explain why a larger difference implies lower reliability. These settings are important components of the method and require more convincing analysis.

Clear inconsistencies in the figures make the method difficult to understand. Figure 1 contains module names such as DFFB and HEGR, while the main text introduces HFEB and HERB. The overall architecture diagram also fails to clearly show the locations and connections of these two core modules. For a paper whose main contribution lies in network architecture design, these issues make the method more difficult to understand and indicate that the presentation still needs improvement. [1] Fu et al., "DW-GAN: A Discrete Wavelet Transform GAN for NonHomogeneous Dehazing," CVPR Workshops, 2021.

[2] Wang et al., "Frequency and Content Dual Stream Network for Image Dehazing," Image and Vision Computing, 2023.

[3] Sun et al., "WDMamba: When Wavelet Degradation Prior Meets Vision Mamba for Image Dehazing," 2025.

Questions for authors

Please further clarify the core novelty and design motivation of the proposed method. What are the main differences between STR and existing dual-branch, frequency decomposition, and feature interaction methods? In particular, why does image dehazing require sum-preserving exchange? What advantages does this constraint offer over conventional residual feature fusion? Please explain the motivation behind the prior construction and gating mechanism in HERB. How were the weights in Eq. (13) and the fixed parameters in Eqs. (15), (16), and (18) determined? Why should a larger difference between the learned haze evidence and the original prior lead to a lower feature exchange weight? I suggest further analyzing whether this design assumption is consistent with the actual dehazing process.

I suggest providing additional analysis of the residual branch and feature exchange mechanism. Why does the network need to predict (J) and (R) separately, rather than predicting only (J) and obtaining the residual through (R=I-J)? What benefits do the additional residual prediction and reconstruction consistency constraint bring to dehazing performance?

Please unify and correct the module names in the architecture diagrams. The names DFFB and HEGR in Figure 1 are inconsistent with those in the main text. I suggest revisiting Figures 1–3, clearly indicating the actual locations and connections of HFEB and HERB, and ensuring consistency between the diagrams and the method descriptions.

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

Related papers

Discussion (0)

Sign in to comment.