Real-World Underwater Image Enhancement Based on Attention U-Net

Round 1

Reviewer 1 Report (Previous Reviewer 4)

Comments and Suggestions for Authors

This paper can be accepted now.

 

Author Response

Dear Reviewer:

Thank you very much for taking the time out of your busy schedule to read and revise my articles. Hope that my paper can be published in your journal.

Sincerely,

Pengfei Tang

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

This study proposes a new deep learning based method for underwater image enhancement. Following is my comments:

1- Abstract part should be revised and It should also include the best results handled.

2-Only 1 underwater image dataset is used, at least one more dataset should be included and an ablation study should be employed between the datasets in order to see the generalization of the network.

3-There should be more state of the art deep learning based methods to be benchmarked. The dehazing methods can also be tried for underwater image enhancement problem.

4-Language should be improved.

5-Training and validation loss and regression data should be visualized to evaluate the training phase in terms of overfitting issues.

6- Conclusion part should be revised. It should emphasize the most important contributions and the most important results handled.

Author Response

Dear Editor,

Thank you for your valuable suggestions. Please refer to the attachment for a comprehensive account of the response details.

Author Response File: Author Response.pdf

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

Underwater image enhancement technology has become important for performing relevant underwater tasks such as object detection and sample capture. Due to the degradation of underwater light, directly captured visual information is usually mixed with a lot of noise, making it difficult to perform these tasks. To solve these problems, a generative adversarial network based on the attention gate mechanism has been proposed in this paper. This network can be trained end-to-end using real underwater image datasets, and the attention gate mechanism highlights the salient features that are passed through the skipped connections. In addition, a new objective function was proposed and the model trained on a real underwater image dataset. This research paper well written and the results of the experiments varied and well presented; nevertheless, some critical issues need to be addressed:

 - Overall, the work developed in this paper provides a good overview of the challenges of underwater image processing and the importance of image enhancement in addressing these challenges.

 - The abstract section could benefit from more specific information about the experimental results, such as the specific quantitative measures used to evaluate performance and how the proposed model outperformed other state-of-the-art models.

 - The proposed approach, which uses a conditional generative adversarial network model based on U-Net attention, which includes an attention gate mechanism, is a promising solution.

 - In the first step of the method, the network restores the degraded image using a U-Net architecture with attention mechanisms. How does this mechanism improve the quality of the restored image? What is the irrelevant or noisy information ?

 - The conclusion section seems to provide a clear summary of the proposed model and its achievements, as well as a direction for future work. Nevertheless, it would be useful to provide more specific information about the quantitative results obtained, such as the parameters used to evaluate the performance of the model and how it compares to other existing methods. In addition, the authors could provide more details on how the proposed method is "universal" and able to adapt to different and complex underwater environments.

 - Experimental results show that the proposed method outperforms several state-of-the-art methods in terms of objective and subjective assessment. Can this approach be used for other underwater imaging applications?

 - Review references of the manuscript.

Author Response

Dear Editor,

Thank you for your valuable suggestions. Please refer to the attachment for a comprehensive account of the response details.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

1- The results on new dataset should be given for all benchmarked methods.

2-I couldn't see the numeric and visual results for DEA-Net in the revised manuscript.

3-Validation loss is still not in the manuscript.

Author Response

Dear Editor,

Thank you for your suggestion, we have revised for your suggestion, see the attachment for details.

Author Response File: Author Response.pdf

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Authors present a method for enhancing real-world underwater images that, to my knowledge, is novel and it is an important topic in image analysis.

References are ok, but there are only one reference from 2021 and one reference from 2022. Since the subject is current, I think the authors should revise the state of the art and look for some other missing recent references.

My main concern is regarding the experiments. Authors explained that they took 90 images (roughly 10% of the dataset) for testing their method, but it is not detailed how they choose these 90 images (randomly maybe?), or if they repeat the same experiment several times taking different 90 images at a time.

I could suggest to run the experiments using at least 20% of images for testing, and repeat the experiment using a different test set of images each time.

On the other hand, it is not clear if, for comparison purpose, the state of the art methods were run using the same 90 images.

 

Another minor remarks:

Authors should revise the whole document. Below, some typos are shown:

- Section 3.1: Generator with skinp

- "Fuison" in tables 1,2 and 3 seems to be a typo for Fusion.

- Section 4.3.3, line 272: However, fter enhanced. An "a" is missing.

Author Response

Q1: References are ok, but there are only one reference from 2021 and one reference from 2022. Since the subject is current, I think the authors should revise the state of the art and look for some other missing recent references.

A1: Due to the long period of research and experiments in this paper, some reference materials may not be novel enough. In the future, I will continue to conduct research in this area and keep track of the latest research results and progress in the field of improvement.

Q2: My main concern is regarding the experiments. Authors explained that they took 90 images (roughly 10% of the dataset) for testing their method, but it is not detailed how they choose these 90 images (randomly maybe?), or if they repeat the same experiment several times taking different 90 images at a time.

A2: The division of training set and test set is completely random, and all methods participating in the comparison use the same test set and training set. My expression is not fresh enough, sorry, I have added a modification to the article and added a corresponding description. Added a corresponding description at lines 205-207

Q3: I could suggest to run the experiments using at least 20% of images for testing, and repeat the experiment using a different test set of images each time.

A3: Since the data set used in this article has only 890 available pictures, in order to obtain a better training effect, a higher proportion is divided into the training set.

Q4: On the other hand, it is not clear if, for comparison purpose, the state of the art methods were run using the same 90 images.

A4: The methods involved in the comparison in this article all use the same training set for training, and the test also uses the same 90 pictures for testing

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposed a conditional generative adversarial network model based on attention U-Net and formulate an objective function by three different loss functions to evaluate image quality from global content, color, and structural information. It is a good solution for real-world underwater image enhancement. However, there are some confusing problems in this article.

(1) In lines 62-65, the problem summarized by the authors does not correspond well to the solution proposed below, and the authors are advised to revise the problem to be solved.

(2) In lines 180-181, the author mentions the fusion weights of the three types of losses, which should be explained in the experimental section to show how they were determined.

(3) In the section "4.3.3 Application on object detection", it is suggested that the numerical results of specific evaluation metrics like mAP be used to illustrate the validity, as the visualization of a single image is not sufficient to illustrate the effectiveness.

(4) Lines 26-28, 48-50, 95, 199, and 272 have grammatical errors, and the authors are advised to check and revise them.

Author Response

Q1: In lines 62-65, the problem summarized by the authors does not correspond well to the solution proposed below, and the authors are advised to revise the problem to be solved.

A1: Thanks, the description has been modified in the corresponding place of the article, so that the description of the proposed problem and solution better matches

Q2: In lines 180-181, the author mentions the fusion weights of the three types of losses, which should be explained in the experimental section to show how they were determined.

A2: There are corresponding descriptions in lines 182-189 of the article. It may be that my description and typesetting are not clear enough, so you did not see it.

Q3: In the section "4.3.3 Application on object detection", it is suggested that the numerical results of specific evaluation metrics like mAP be used to illustrate the validity, as the visualization of a single image is not sufficient to illustrate the effectiveness.

A3: The object detection results are only for subjective reference. Due to the tight revision time, it is temporarily impossible to add detailed experiments and evaluation indicators.

Q4: Lines 26-28, 48-50, 95, 199, and 272 have grammatical errors, and the authors are advised to check and revise them.

A4: Thanks for the reminder, it has been revised in the corresponding part of the article

Reviewer 3 Report

Comments and Suggestions for Authors

The abstract can be further improved.

Author Response

Q1: The abstract can be further improved.

A1: Many thanks, the abstract section of the article has been supplemented and revised

Reviewer 4 Report

Comments and Suggestions for Authors

Real-world Underwater Image Enhancement Based on Attention U-Net

In this paper, the author proposed a conditional generative adversarial network model based on attention U-Net which contains an attention gate mechanism, it could filter invalid feature information and capture contour, local texture, and style information effectively.

Comment:

- Please revised your abstract, 146 words can not represent the whole manuscript. Abstract at least contain 250 words.

- In academic work, comparing the obtained results to some related/recently published works under the same conditions (i.e., databases + protocols of evaluation) is necessary. The objective is to show the superiority of the presented work against the existing ones. Please explain more about the previous research result in this field.

- This paper does not answer the main purposes of this paper that explain in the introductions.

- Need to discuss more the proposed method's benefits and limitations.

- The author uses the Underwater Image Enhancement Benchmark (UIEB), can you do testing in another dataset?

- Revised Figure 4 with a good resolution. Image blur and the text in the red box cannot be read.

- Need to add more discussion about the experiment and results.

- This paper needs to rewrite again. 

Author Response

Q1: Please revised your abstract, 146 words can not represent the whole manuscript. Abstract at least contain 250 words.

A1: Many thanks, the abstract section of the article has been supplemented and revised

Q2: In academic work, comparing the obtained results to some related/recently published works under the same conditions (i.e., databases + protocols of evaluation) is necessary. The objective is to show the superiority of the presented work against the existing ones. Please explain more about the previous research result in this field.

A2: This article has exemplified some of the latest research results and progress in this field. At the same time, several representative methods are selected for detailed discussion and comparative experiments in this paper. There may be problems with typography or some descriptions, which have been corrected in the text, thank you again for your suggestions

Q3: This paper does not answer the main purposes of this paper that explain in the introductions.

A3: Sorry, there may be some problems with the description. This paper aims to propose a neural network-based underwater image enhancement network that facilitates underwater vision tasks.

Q4: Need to discuss more the proposed method's benefits and limitations.

A4: Thank you for your suggestion. The discussion of the main research content and the result elaboration have been carried out in this article, and some descriptions may not be clear enough. Corrected in the paper.

Q5: The author uses the Underwater Image Enhancement Benchmark (UIEB), can you do testing in another dataset?

A5: the main reason for choosing the UIEB dataset in this article is that this dataset is composed of completely real underwater images, and using this dataset for training and testing can effectively restore the real and complex underwater environment

Q6: Revised Figure 4 with a good resolution. Image blur and the text in the red box cannot be read.

A6: The resolution of the original image is 256 x 256, so the labeling of the detection results will be blurred to a certain extent

Q7: Need to add more discussion about the experiment and results.

A7: Thanks for your advice. Due to the tight time for modifying the article, it is impossible to add corresponding experiments in a short time

Q8: This paper needs to rewrite again.

A8: Thank you for your suggestion, some descriptions of the article have been revised to make the expression more clear。

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Authors present a method for enhancing real-world underwater images that, to my knowledge, is novel and it is an important topic in image analysis.

After a second round for carefully reviewing the manuscript, authors have not addressed my comments. They only take into account the minor remarks I highlighted. However, the experimental design is not adequately described even when this situation was observed in the first round of reviews.

For that reason, in my opinion, the paper is rejected.

Reviewer 4 Report

Comments and Suggestions for Authors

Dear author please answer my questions correctly and clearly, and show your revision in the page number or which line number. Dont just reply in general or revised in the paper. It is not clear which one you revised.

Regarding to my comment:

Q7: Need to add more discussion about the experiment and results.

A7: Thanks for your advice. Due to the tight time for modifying the article, it is impossible to add corresponding experiments in a short time.

Yes, in my opinion, the author needs more time to conduct the experiment and rewrite the manuscript.