Deep Learning Based Filtering Algorithm for Noise Removal in Underwater Images

Round 1

Reviewer 1 Report

The paper presents a sound and innovative approach for noise removal in underwater images. The article made an interesting read.

However, there are quite a number of concerns, some of which are minor:

1. Your result seemed to be better just by the MSE test. However, you did claim in line 638 that you have a "higher PSNR value compared to state of art methods." Why is this so?
2. This sentence in lines 41-42 is not clear - 

   "As a result, visibility is reduced beyond the 20m range, 41 and visibility is reduced to less than 5m in turbid coastal waters."

3. When you start a new paragraph with "As a result" it creates a disconnection between the preceding paragraph and the new one.
4. There are a number of places you missed to add a full stop, close brackets (e.g. 352), and you need to make your paragraph indentation consistent.
5. Your images and tables need to be properly aligned.
6. Some of the figures need huge improvement. It will be a shame for you to present such interesting research of making images better and then the figures you presented do not look good. Figures 2, 4, and 5 are either blurred or have unclear texts. Note there is no Fig. 3, why?
7. I'm not sure about this referencing style. You might need to check whether to have only (David et al. 2003) or [34], having both doesn't seem right.
8. It appears you were editing this document in word and didn't finalise it before exporting it to PDF, see lines 137 - 159; there are other places as well.
9. You need to thoroughly proofread this work. There are quite a number of typos and a lack of consistency in the use of terms (e.g. Backward scattering vs Back scattering vs Backscattering; also O(n) vs O(N)). Check lines 125-126; and end of 357.
10. Merge the paragraph on 201 to the one ending in 200.
11. In 216, you have a) but no b), it seems to be a mistake.
12. These papers will give you insight into the mechanism of deep learning applications. Cite them https://doi.org/10.1007/s00530-020-00701-5 
    
    https://dl.acm.org/doi/10.1145/3448614

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors spent huge effort on explaining the physics of various underwater distortions in the introduction section, but then proposed such a simple approach for image denoising and histogram enhancement, making none of the explained physics relevant to the proposed methods.

The different blocks of the proposed processing chain are made by pre-existing algorithms. The authors spent huge effort to re-describe others’ methods within the “method” section, which should actually be the place to write their own contributions.

There is missing comparison and justifications to demonstrate why the different processing blocks are needed. For example, the authors applied a CNN network for image denoising, but actually a simple CNN network should be capable of doing everything in one go - all the small processing blocks seem redundant if CNN is introduced.

The results are not convincing. With a few lines of OpenCV code or using online video editing tools, it is very likely to be able to achieve the similar result. If there is photoshop or Final cut, the result could be even better.

The abstract and introduction section are well written. But starting from the method section, the texts are more like a student’s note that are full of typos, redundant and unnecessary information, formatting issues of figures and references.

The presented content could form a good project report considering the different experiments that are made, but is not appropriate for publication as a research article for aforementioned reasons. Before adding any new content to the existing draft, I suggest the authors to first cut down all the unnecessary and redundant text (e.g., the text that describes a pre-existing method), and then rewrite what is actually new.

Detailed comments

1. All references should to be cited in the order of their appearance and some of the references need to be reformatted (e.g., line 120, 127, 131).
2. The word “additionally” has been frequently used throughout the text, but consider rephrasing in a more elegant way of writing, e.g., line 73-75
3. Line 44-68 & Line 85-91: please add a figure to show examples of the various issues that are described here.
4. Line 149-159: instead of outlining your research objective, it is more appropriate in research articles to outline your key contributions.
5. Line 162: … have been developed. Time?
6. Line 314: aris4d.com this site does not exist, please correct.
7. Line 344: ANN typo.
8. Figure 2: this should be replaced by a clear figure showing the architecture of your own CNN network. If you are reusing pre-existing networks, the reference should be provided in caption.
9. Line 347 – 363 & Figure 2 : it sounds really odd to review the basics of CNN in the section of “methods” which should be the place to describe your own methods.
10. Line 357: The out B JKL;put function -> typos?
11. Figure 4: your output of the image enhancement module is a denoised image – this is not image enhancement but denoising. It is hard to see any differences between the input and output after denoising, and input and output after the bilateral filtering block. It
12. Where is Figure 3??
13. Line 385-387: what is your contribution here? I believe you have made a comprehensive review on previous work and should not continue reviewing others’ methods in the “method” section.
14. Line 388: rephrase “This enables to avoid the averaging of …”
15. Line 455-495: these paragraphs look like a student’s notes, please only describe your new contributions to the existing CLAHE methods and why your implementation is optimal to the existing one. It looks really odd to describe existing methods in such great details in your own “method” section.
16. Figure 6: this is not convincing, a few lines of opencv code on bilateral filtering and contrast adjustment would provide the same results.
17. Figure 7: adding additive noise to the input image doesn’t make any sense on demonstrating your methods for denoising and enhancing underwater images.
18. Table 2: this should either be a figure with measurements superimposed or a plain table without the pictures.
19. Line 584: majorly contributed -> mainly caused
20. Line 586: removed to -> reduced by
21. Line 587: “bu”
22. Line 593: network -> processing chain or “system”
23. Line 602-606: you can summarise the quantitative evaluation in the conclusions section, but this is not considered as “discussion”.
24. Line 608: lesser?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

I can see the manuscript has been improved - it looks much better than the previous submission - thanks for considering my suggestions and resubmit.

However, I still have two major concerns.

Firstly, where is the description of the network architecture? (what layers you are using? how many of them? number of feature maps? kernel size? how is your training set up? what are the loss functions? ) Please check some of the CVPR papers on how to describe a network architecture if you are using a CNN based method. However, if you are just reusing an existing network with a pre-trained model, the reference and github link need to be clearly cited.

Last but not least, you really need to remake Figure 4 - make the images bigger with full-page-width for the two columns or reproduce using other images - this is the first and a key figure to demonstrate the efficacy of your method, but it is not convincing at all at its current form, because I can hardly observe any differences or improvement between each of the subimages other than a minor tone difference. At the form of this figure, I am not convinced for why we need these processing steps if a few simple image editing operations or a few lines of opencv image enhancement code can produce similar output.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 3

Reviewer 2 Report

Thanks for addressing the two major issues - the method description and result presentation look much better now. I suggest the paper to be accepted in its present form, leaving the minor English and formatting issues with the post-production editors.

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

The paper proposes a method to address these issues by de-noising and increasing the resolution of the image using a model network trained on similar data. The network extracts frames from a video and filters them with a Trigonometric – Gaussian filter to eliminate the noise in the image. It then applies Contrast Limited Adaptive Histogram Equalization-CLAHE to improvise the image contrast and finally enhances the image resolution. Experimental results show that the proposed method could effectively produce enhanced images from degraded underwater images. 1.Figure one should be exported proper and not just screen grabbed from Microsoft word to avoid the red and blue error markings under "sun" and "Rayleigh" .

A few grammatical errors in the paper needs the authors attention

The research objectives and contributions should be clearly outlined in the introduction section.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper "Deep Learning based Filtering Algorithm for Noise Removal in Underwater Images" has an interesting topic, however, there are several flaws that need to be considered in order to fulfill a scientific piece of a research paper. Also, it was difficult to grasp the ideas and the integrity of the work due to the way the manuscript was written. Below are some essential points that seriously need to be considered for the publication. for example:

1- The introduction was weak. It didn't cover the topic, trends, focus of the study, approached used also there was no aim mentioned. It is necessary to address the aforementioned points to show the significance of the study and link it with the contribution of the study. 

2- Related work also was not comprehensive. It did not explain the problem clearly. What are the common approaches and main methods used for the Filtering Algorithm for Noise Removal in Underwater Images?  what are the techniques in each category? What is the gap? All these points need to be considered in this section, otherwise, the reader can not find a clue between the methods and the proposed method. 

3-  The metrics used should be mentioned in the last part of the method section. 

4- The FLowchart of the study is not clear and needs more improvements by adding more details so the reader can find the idea smoothly. 

5- What and where are the architecture details used for deep learning used?

6 - What data was used for the study? what is the source? how did the authors collect the data? All these items should be mentioned.

7- In the result, the authors need to address why, what and how they got such a result? What factors might affect the result? and more details..

8- The paper does not have a discussion section which is very important. Where this study does stand compared with previous works? What has been improved and what is the elimination of the proposed method?

9 -The contribution of the work and the gap is not clear. 

 In general, the paper needs extensive improvement before publication in "Water''.

 

 

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I think the paper in its current state is acceptable 

Reviewer 2 Report

Although the article has been improved compared to the previous version, It still has several flaws. I am mentioning some of those essential points;

The introduction has improved, but the motivation and significance of the study were not well presented as a standard scientific form. Also, the background and related studies still need extra work. As I mentioned in my earlier review, the gap of the study was not clearly presented. You do not need to put figures and tables in the introduction and background and related studies; Instead, you may put them in the methods section if necessary ( however, the presented figures are still generic and do not add extra value to the manuscript).

Only general deep learning information was mentioned in the manuscript. The specific configurations and hyperparameters used for this study must be reported. Also, the search space used for the best values and parameters needs to be presented to see the best values for the optimization. 

In the discussion, you should compare the finding of your studies with at least 3 recent related studies to show what improvement has been made. Also, what can be done to improve the method further. As I mentioned in the previous review, you need to show where your research stands compared to those studies? 

I appreciate the author's effort in improving the manuscript, but the current version may not be the best version of this manuscript. I hope the comments could be helpful for the authors.