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?
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.