Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network

Real-world Image Dehazing (RID) aims to alleviate haze-induced degradation in real-world settings. This task remains challenging due to the complexities in accurately modeling real haze distributions and the scarcity of paired real-world data. To address these challenges, we first introduce a cooperative unfolding network that jointly models atmospheric scattering and image scenes, effectively integrating physical knowledge into deep networks to restore haze-contaminated details. Additionally, we propose the first RID-oriented iterative mean-teacher framework, termed the Coherence-based Label Generator, to generate high-quality pseudo labels for network training. Specifically, we provide an optimal label pool to store the best pseudo-labels during network training, leveraging both global and local coherence to select high-quality candidates and assign weights to prioritize haze-free regions. We verify the effectiveness of our method, with experiments demonstrating that it achieves state-of-the-art performance on RID tasks. Code will be available at https://github.com/cnyvfang/CORUN-Colabator.

Paper

Similar papers

Peer review

Reviewer wF188/10 · confidence 5/52024-07-01

Summary

This paper posits that real-world image dehazing is particularly challenging due to the intricacies of accurately modeling haze distributions and the limited availability of paired real-world data. Traditional and deep learning-based methods struggle to address the complexities of real haze, often resulting in color distortion and suboptimal outcomes. To tackle these issues, this paper introduces the CORUN model, which integrates atmospheric scattering and image scenes to incorporate physical information into deep networks. Furthermore, it proposes the Coherence-based Label Generator to produce high-quality pseudo labels for network training, thereby improving the network's generalization in haze removal.

Strengths

1. The proposed method, CORUN (COopeRative Unfolding Network), integrates physical information into deep networks to overcome the limitations of existing deep learning-based dehazing methods. It cooperatively models atmospheric scattering and image scenes by incorporating Transmission and Scene Gradient Descent Modules at each stage, effectively restoring haze-contaminated details. 2. CORUN is constructed on the basis of the atmospheric scattering model using a proximal gradient descent method within a deep unfolding network, which provides strong interpretability to the method. 3. The proposed Colabator framework significantly improves the performance of models pretrained on synthetic datasets in real-world scenarios within a limited number of iterations by fine-tuning with real degraded images. This plug-and-play framework incurs no additional computational cost during deployment. 4. Experiments demonstrate the excellent performance of the proposed CORUN model and Colabator framework in real-world dehazing tasks. Both quantitative and visual results validate the network's capability to model real-world haze and restore real-world scenes effectively.

Weaknesses

1. The paper contains some inconsistencies in the use of symbols. While images are generally denoted by P, in the loss functions (15, 16, 17), the authors use I to represent images. Although the text explains the use of these symbols, it can still impact readability. Additionally, in line 172, the pseudo-label symbol is misspelled and should be corrected to P^{R}\_{Pse\_{i}}. 2. In line 175, the symbols contain unnecessary tildes, which should be removed. The symbols in Figure 2 should also be consistent with those used in the formulas. 3. Equations (5) and (7) do not address the consistency of the number of channels between the Transmission Map and the image. Although this is illustrated in Figure 2, it still needs to be clearly stated in the equations to facilitate readability. 4. This paper adequately demonstrates its method's performance in real-world dehazing tasks through thorough experiments in both visualization and quantitative results. However, the provided ablation studies are insufficient. The authors should conduct more comprehensive ablation studies and experiments on different datasets to further validate the effectiveness of their method and its components.

Questions

1. This paper introduce Colabator, a plug-and-play coherence label generation method designed to fine-tune models pre-trained on synthetic datasets using real degraded images, thereby achieving better real-world image processing results. Can the authors provide additional results of this framework's fine-tuning on various tasks to support this claim? 2. I noticed that the authors assign weights to the local quality of pseudo-labels using a weighted combination of CLIP scores and NR-IQA scores, instead of the more common approach of using their dot product. Can the authors provide experimental results to demonstrate the effectiveness of this choice? 3. Regarding the choice of NR-IQA, why did this paper select MUSIQ over other methods? What are the reasons behind this decision? 4. There is a small portion of similar or identical image content between the RTTS and URHI subsets of the RESIDE dataset. Did the authors notice this issue during their experiments? Have they addressed this issue, or did they directly use the full URHI subset for fine-tuning? If the authors used the entire URHI subset for fine-tuning, have they considered the model's performance on the RTTS subset after removing these similar images?

Rating

8

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Regarding limitations, the proposed method partially addresses previous issues such as the lack of paired real-world data and the difficulty in modeling complex real-world haze distributions. Section A.1 "Limitations and Future Work" in the appendix discusses the current work's limitations and potential future research directions to address these unresolved issues. As for broader impacts, Section A.2 in the appendix provides a thorough discussion. As stated in the paper, this work has no foreseeable negative impacts on the field or society and offers substantial positive contributions.

Reviewer h5fd7/10 · confidence 4/52024-07-10

Summary

This paper focuses on challenging real-world image dehazing problem. The authors develop their network based on unfolding network, while leveraging cooperative proximal mapping modules to facilitate the estimation of transmission map and image content. In addition, the authors propose an interesting teacher network, named as Colabator, with newly developed quality assessment metrics to better provide the high-quality pseudo-labels for dehazing. The subjective metrics on different benchmark, and the user study with downstream tasks evaluation demonstrate the effectiveness of this work.

Strengths

1. The Colabator devised by authors sound reasonable, it also provides some insights on training scheme design for restoration problems where paired-images data are hard to collect. 2. The CPMM modules further improve the representative abilities of unfolding networks to model the gradient descent process of image dehazing problems. 3. The experimental results give better subjective qualities of the proposed network compared to other methods. Moreover, the results of user study and downstream tasks confirm this point. Judging by myself, I also agree that the authors’ method generates more natural restored images with minimal color shift and the best quality.

Weaknesses

1. The derivation of equation (7) and (10) are unclear. The authors may need to provide some basic theories before stating these equations. 2. The details of using DA-CLIP as haze density evaluator is unclear, the statement in line 164 indicates that the authors use a fixed text feature, which is also missing in the article. Moreover, the reason behind the patch partition in line 165 is not clearly stated, the information about patch size N cannot be found in the article. 3. The article lacks ablation studies about CPMM modules. I wonder how network will performance if removing all CPMM modules.

Questions

First of all, please address my concerns listed in the weakness part. I have some additional questions: 1. About density loss: similar as the problems in the second point of weakness, how does this loss function formulate? 2. In the fine-tuning phase, as the student network is optimized using strong augmented data, is the performance gain coming from this augmentation strategy? 3. Why partition images before quality assessment? How about using full-resolution images as input? I hope to see the answers and will adjust my rating accordingly.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have stated the limitations of their article in the supplementary materials.

Reviewer ivCJ5/10 · confidence 3/52024-07-12

Summary

The paper aims for real-world image dehazing, where the paper tries to handle the difficulties of modeling real haze distributions and the scarcity of real data. For modeling haze, the paper proposes to jointly model atmospheric scattering and image scenes by a cooperative unfolding network. To handle the scarcity of data, the paper proposes to generate pseudo labels using an iterative mean-teacher network, which the paper terms Coherence-based label Generator (Colabator).

Strengths

- The paper is well-written and easy to follow. - The proposed method is effective in that object detection results are improved on images dehazed by the proposed method in comparison to previous works. - The iterative mean-teacher framework for generating pseudo labels seems to be effective in image dehazing.

Weaknesses

- The major concern I have for this paper is that the proposed method seems to be mostly combination of existing works. - The joint estimation of atmospheric scattering/transmission (or atmospheric light) and clean image is not new: - Mondal et al., Image Dehazing by Joint Estimation of Transmittance and Airlight using Bi-Directional Consistency Loss Minimized FCN - Im et al., Deep Variational Bayesian Modeling of Haze Degradation Process - Zhang et al., Joint Transmission Map Estimation and Dehazing using Deep Networks - Ren et al., Single image dehazing via multi-scale convolutional neural network - Deep Unfolding Networks (DUNs) has been utilized for dehazing by Yang and Sun [50], mentioned in the paper - The use of PGD for DUNs in the context of image restoration has been done by DGUN [33], as mentioned in the paper. - Also, using mean teacher for creating pseudo labels is commonly performed in deep learning. - Considering all of the above, the proposed method seems to be the combination of the existing methods. - What is a major difference and technical novelty in comparison to the combination of existing methods above? - Since the proposed method is a combination of many components, a detailed ablation study would be needed to justify the effectiveness of each component.

Questions

My major concern and questions have been written in the weakness section.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have mentioned the limitation of their work in the supplementary document.

Reviewer ivCJ2024-08-13

I appreciate the authors' rebuttal, which has addressed most of my concerns. One additional concern I have after reading the rebuttal is the authors have made comparisons against other methods on I-HAZE, O-HAZE, but not on RTTS, a benchmark the authors have used in the main paper. I believe this has to do with the other reviewer's recommendation. It would be great if the authors could make comparisons on RTTS as well in the final version. Also, the authors have presented a lot of new experimental results during the rebuttal, requiring a major change to the paper. I'm ok with the acceptance and hence raise the rating, given that the authors make these major changes to the final version of the paper, along with new comparisons on RTTS as well.

Authorsrebuttal2024-08-13

Thanks for recognizing the value of our work!

We wish to express our sincere appreciation to the reviewer for recognizing the substantial significance of our contribution, specifically the Colabator framework and the CORUN method, within the realm of real-world image dehazing! First, we will incorporate the experiments and major changes conducted during the rebuttal phase into our revised manuscript. Our initial plan is to include critical experiments, such as detailed comparisons with existing methods and essential ablation studies, directly in the main text, while other experiments will be placed in the supplementary material due to space constraints. To make room for these important experiments, we will move some less critical visual results to the appendix. Additionally, following the approach in **W1**, we have added comparisons with existing methods on the RTTS dataset. As shown in the table, we observed significant performance improvements when integrating either our joint estimation component or our unfolding framework proposed in CORUN. This experiment will also be included in our revised version. Please feel free to let us know if you have any further concerns. We would be happy to address them. | Metrics | Ren[66] | Ren+CORUN | Mondal[64] | Mondal+CORUN | Im[65] | Im+CORUN | PDN[50] | PDN+CORUN | DGUN[33] | DGUN+CORUN’s model | DGUN+CORUN | CORUN (Ours) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | NIMA↑ | 3.173 | 3.420 | 3.372 | 3.518 | 3.304 | 3.414 | 2.925 | 3.271 | 4.272 | 4.419 | 4.583 | 4.823 | | BRISQUE↓ | 31.335 | 28.441 | 29.951 | 27.773 | 26.906 | 25.103 | 33.186 | 29.342 | 29.753 | 27.728 | 25.308 | 20.944 | | FADE ↓ | 2.065 | 1.878 | 1.732 | 1.626 | 1.719 | 1.633 | 2.159 | 2.006 | 1.663 | 1.535 | 1.376 | 1.051 |

Reviewer ivCJ2024-08-14

Thanks for the update. I have no further concerns at the moment.

Authorsrebuttal2024-08-14

Thanks for recognizing our work!

Your acknowledgment holds great importance to us and serves as a meaningful validation of our dedicated efforts to advance this critical area of research. In this work, we are the first to propose a plug-and-play iterative mean-teacher framework (Colabator) for real-world image dehazing, along with a robust dehazing algorithm (CORUN) with theoretical guarantees. It’s important to emphasize that the Colabator framework is not limited to dehazing; it can also be applied to other significant low-level vision tasks, such as image desmoking and underwater image enhancement. In this case, **we believe this work has the potential to inspire advancements across the entire field of low-level vision for real-world image processing.** We would greatly appreciate any further discussion, as well as a reconsideration of the rating.

Reviewer ivCJ2024-08-14

I would like to note that I have updated the rating, as noted in the previous comment, considering that authors would make the changes to the final version

Reviewer Udmv5/10 · confidence 5/52024-07-13

Summary

This paper focuses on Real Image dehazing. They propose a cooperative unfolding network (CORUN) to integrate physical knowledge for image dehazing. The proposed CORUN exploits the complementary information between components in Atmospheric Scattering Model. Besides, due to the lack of real paired data, this paper proposes a Coherence-based Label Generator to generate pseudo labels for real hazy inputs. The experimental results demonstrate that the proposed method achieves good dehazing performance.

Strengths

1. The research direction of this paper is meaningful. 2. The quantitative results show that the proposed CORUN+ achieves state-of-the-art performance in terms of NRIQA metrics. 3. The visual results shown in Fig.5 outperform other competing methods.

Weaknesses

1. There is still room for improvement in the writing of this paper. a) Typo in line 167 b) How to simplify A in line 111 is unclear. According to my understanding, after simplification, P in Eq. 2 has different physical meaning as defined in line 110. c) In lin3 172, the same symbols are used to indicate different images. d) Line 187 presents two kinds of reconstruction loss, which is inconsistent with the perceptual loss in line 186. 2. The experiment is a bit unfair. This paper utilizes the data generation pipeline proposed in RIDCP, which has been proved that can improve the real image dehazing performance of dehazing networks. However, the quantitative results of PDN and MBDN reported in Table 1 seem to be calculated by using their original models, rather than re-training on the same training data used in this paper. 3. The experiment is a bit insufficient. a) The author only validates the effectiveness of Colabator on DGUN, neglecting other widely-used dehazing networks. b) The full-reference evaluation on real dehazing benchmarks (e.g., O-HAZE, I-HAZE, etc.) can also be an important evidence to support the effectiveness of the proposed method.

Questions

Please see the weakness for more details. Besides, I was wondering: 1. What are the essential advantages of CORUN compared to state-of-the-art end-to-end dehazing models? Are there any experiments? 2. Whether the proposed method can remove medium or dense haze correctly. Since the dehazing results shown in Fig. 4 still have haze residues even if the hazy input is not very dense. 3. Why use both perceptual loss and contrastive loss simultaneously? There are overlapping parts between them. 4. How robust is the estimation of haze density by DA-CLIP?

Rating

5

Confidence

5

Soundness

2

Presentation

2

Contribution

2

Limitations

The authors have discussed the limitations.

Area Chair 97Af2024-08-09

Read the rebuttal and discuss with the authors

Hi Reviewers, The authors have submitted their rebuttal responses addressing the concerns and feedback you provided. We kindly request that you review their responses and assess whether the issues you raised have been satisfactorily addressed. If there are any areas where you believe additional clarification is needed, please do not hesitate to engage in further discussion with the authors. Your insights are invaluable to ensuring a thorough and constructive review process. Best AC

Reviewer wF182024-08-09

After reviewing all the materials and discussions on this page, I believe the authors have made significant efforts to address my main concerns. The proposed framework Colabator has strong generalization performance, including dehazing, desmoking and underwater image enhancement tasks. I believe that this quality-evaluation-based pseudo-label selection framework provides a new direction for real-world image restoration. I am now happy to update my rating to strong acceptance.

Authorsrebuttal2024-08-09

Thanks for recognizing the value of our work!

We wish to express our sincere appreciation to the reviewer for recognizing the substantial significance of our contribution, specifically the Colabator framework and the CORUN method, within the realm of real-world image dehazing, along with image desmoking and underwater image enhancement. Your acknowledgement holds great importance to us and serves as a meaningful validation of our dedicated efforts to advance this critical area of research.

Reviewer h5fd2024-08-11

We thank the authors for providing a detailed rebuttal in such a short time duration. After reviewing the rebuttal, I believe the authors have solved all my concerns about their paper. I will raise my rating to accept, and I support accepting this paper. However, PLEASE make sure to update missing information about DA-CLIP and equation parts in the final revision to make the paper more comprehensive.

Authorsrebuttal2024-08-11

Thanks for recognizing the value of our work!

We wish to express our sincere appreciation to the reviewer for recognizing the substantial significance of our contribution, specifically the Colabator framework and the CORUN method, within the realm of real-world image dehazing, along with image desmoking and underwater image enhancement. Your acknowledgment holds great importance to us and serves as a meaningful validation of our dedicated efforts to advance this critical area of research. We will update the missing information about DA-CLIP in our final version.

Reviewer Udmv2024-08-13

Thank you for your detailed reponse

The detailed rebuttal from the authors addresses most of my concerns. I want to thank the authors for their work. At this stage, I am considering the overall rating of this work mainly based on its novelty, contributions, and insights.

Authorsrebuttal2024-08-13

Thanks for your encouraging reply!

We would like to sincerely thank the reviewer for the encouraging feedback. In our **seven** response to reviewer Udmv, we have added **nine** tables and **two** figures to thoroughly demonstrate the superiority of our CORUN dehazing model and the generalizability of our Colabator framework. Additionally, we encourage reviewer Udmv to briefly review the content in the global rebuttal as well as our responses to the other reviewers. These sections further validate the generalization of our Colabator framework across other critical low-level tasks and provide a detailed comparison between our method and existing dehazing techniques, supported by extensive experimental evidence. We believe this will help the reviewer fully understand the novelty, contributions, and insights of our paper. Specifically, **we are the first to propose a plug-and-play iterative mean-teacher framework (Colabator) for real-world image dehazing, along with a robust dehazing algorithm (CORUN) with theoretical guarantees.**

Reviewer Udmv2024-08-13

final score

I was convinced by the statement from the authors. This work is somewhat acceptable. I decide to raise my score to 5.

Authorsrebuttal2024-08-13

Thanks for recognizing the value of our work!

We wish to express our sincere appreciation to the reviewer for recognizing the substantial significance of our contribution, specifically the Colabator framework and the CORUN method, within the realm of real-world image dehazing! Your acknowledgment holds great importance to us and serves as a meaningful validation of our dedicated efforts to advance this critical area of research.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC