Cycle-Iterative Image Dehazing Based on Noise Evolution

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes an image dehazing algorithm based on an iterative noise evolution approach. It combines three modules: a dehazing module based on the Atmospheric Scattering Model (ASM), a random haze addition module to enrich the data, and an inverse enhancement module based on the Retinex model. Each component leverages lightweight U-Net architectures, ensuring low computational complexity. The proposed approach aims to improve generalization on real-world images while providing real-time performance. Experiments show that the algorithm achieves good results in terms of restoration quality and efficiency.

• This study presents an interesting approach, but several aspects need to be addressed to improve the quality of the manuscript:
• The introduction mentions the effects of haze on images and vision task performance, but does not sufficiently elaborate on:
  • The limitations of traditional methods,
  • The specific motivations for using noise evolution,
  • The rationale behind combining ASM and Retinex.
• The stated contributions are vague and lack precision. The authors should reformulate their contributions more explicitly, highlighting the truly novel aspects of their approach.
• A paragraph presenting the structure of the paper should be added immediately after the list of contributions to better guide the reader.
• I encourage the authors to include a diagram (synoptic schema) illustrating the overall methodology. Such a visual representation would help clarify the sequence of modules, their interactions, and the overall processing flow within the proposed model.
• The DNMGDT model achieves quantitatively better results than the proposed method, with a PSNR of 24.55 dB and SSIM of 0.873 indoors, and a PSNR of 21.12 dB and SSIM of 0.757 outdoors. In this context, it would be helpful for the authors to explain the reasons for this performance gap and to clearly state the specific advantages or added value of their approach compared to DNMGDT.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper addresses the key problem of improving generalization and efficiency in image dehazing algorithms in real-world scenarios where current deep learning models often fail.

The topic is relevant and original, as it introduces a novel perspective by integrating a cycle-iterative process inspired by noise evolution -combining dehazing, controlled haze reintroduction, and Retinex-based inverse enhancement.

Methodologically, the approach is well-motivated and tested extensively.

The conclusions are generally well-supported by quantitative (PSNR, SSIM) and qualitative results, including ablation and object detection performance analyses.

The references are appropriate and up-to-date.

Tables and figures are highly informative and clearly illustrate performance gains across scenarios, particularly in challenging real-world settings.

Furthermore, the mathematical formulations are precise, and align well with the theoretical foundations.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper proposes a cyclic iterative image dehazing algorithm based on noise evolution, aiming to solve the problems of incomplete data set, insufficient prior extraction and large network size in existing deep learning dehazing methods. Experimental results show that this method has certain advantages in dehazing effect, computational efficiency and generalization ability. Overall, the topic of the article is of practical significance, the technical route is clear, and the experimental design is basically reasonable, but there is still room for improvement in innovative explanation, method details and experimental demonstration.

1. It claims to be "the first time to build a dehazing framework from the perspective of noise evolution", but it has not conducted an in-depth comparison with existing iterative dehazing methods or data enhancement strategies, making it difficult to identify the technical breakthroughs and its innovation is questionable.
2. The choice of the gamma value range [1.1,1.5] in the random fog module lacks experimental verification, and there is no explanation as to why this range can "balance the high and low transmittance areas", nor is there any comparison of the impact of different ranges on performance.
3. Only the tractable RESIDE-SOTS synthetic data is used, and the authoritative real fog image benchmarksare not tested. More experiments are needed.
4. Table 1 shows that the PSNR of this paper (23.83dB) is lower than DCMPNet (24.53dB) and DNMGDT (24.55dB), but it claims to be "better than SOTA".

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

My concerns have been addressed by the authors, and I recommend this article for publication.

Author Response

Thanks for your positive comments. Many thanks!

Reviewer 3 Report

Comments and Suggestions for Authors

the authors have modified the manuscript according to the comments one by one.

Author Response

Thanks for your efforts! Many thanks.