Underwater Image Enhancement Algorithm Based on Adversarial Training
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors1. I suggest you write the content in Key words as “Underwater image enhancement”, instead of “Underwater” and “Image enhancement”.
2. The sections of “Introduction” and ”Related work”should be more enriched, and the authors need to provide an overview of the latest developments and research situation. Currently, there is still a lack of some latest research work.
3. In your references ([6] and [7]), they should be “In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition”instead of “In Proceedings of the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition”. The authors must double check this issue.
4. The tables and Figures displayed in this paper, the authors named “This paper” for their results. I do suggest the authors changing “This paper” into “The proposed method”(e.g. Table 1 and Figure 9 etc.).
5. The compared algorithms include two deep learning algorithms, CycleGAN [7] and FUGAN [9], as well as three non-deep learning algorithms, BLOT [23], CB [24], and RGHS [25]. These methods are proposed from 2017 to 2020, which can not be regarded as the state-of-the-art methods. These comparison methods are not novel enough and lack sufficient persuasiveness. Two methods (at least) published in recent years should be used to compare.
6. Besides, experiments should be added to prove the effectiveness of preprocessing method. They can be the preprocessing method, the whole framework without the preprocessing method, and the whole framework.
Comments on the Quality of English LanguageAs shown in the comments and suggestions.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper proposes an adversarial learning-based approach for Underwater Image Enhancement (UIE). Additionally, a high-dimensional semantic cycle loss is utilized to effectively deal with key issues such as blurriness, color distortion and brightness artifacts. A helpful preprocessing is also applied to address the significant color casting in diverse datasets, thus enhancing the feature learning ability for subsequent style transfer. Various experiments in UIE datasets validate the generality of the proposed method. I would like to recommend accepting this work if the author can address my following concerns.
1. Punctuation is required between formulas and the text for better presentation and variables should be clarified more clearly. For example, are the sigma and sigma. in Eq. (8) the same?
2. Image caption descriptions, such as for Fig. 5-7, are simply presented in this manuscript and drawings of all images are also relatively coarse. Uniform punctuation is required for all captions in this article.
3. The evaluation metrics of this paper is somewhat old, e.g. UIQM. Can your method work well in more advanced metrics such as UIF?
4. The references are somewhat old. It is suggested to refer more recent UIE works such as LCNet, CURE-Net, UIE-Convformer, etc.
5. The compared algorithms seem not to be the latest methods. Can the authors provide a stronger comparison with state-of-the-art methods?
6. The character size and font of the images in this article are obviously inconsistent with the body text.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper approaches the enhancement of underwater images using an adversarial learning-based approach.
The paper is technically sound, the results are correct and are compared with state-of-the-art. The paper is well organized and generally written in a clear style. The references are well chosen and relevant for the subject.
Revise the following issues:
1) Please explain for the reader the terms "color casting" and "style transfer".
2) Verify equation (3) and explain it in more detail; the expression N(I − G * I) is not clear; what represents the symbol * ? (product or convolution?) What represents the letter G?
3) Verify equation (4), it is rather unclear. Explain it clearly in text, it is really difficult to understand. As they are written, the terms − G1{I(x)} and G1{I(x)}, − G2{I(x)} and G2{I(x)} seem to cancel out (?); in the final sum, it should be Ll{I(x)} (the index "l" of the sum), not L1{I(x)} etc.
Please explain for the reader the relation between the Laplacian and Gaussian pyramids and why exactly are they used here.
After equation (4): "The weight WL is a combination of Li" (?) ; this weight WL does not appear anywhere in equation (4); please revise and correct if necessary.
4) Regarding equation (5): what exactly is meant by "pixel block (x,y)" and "pixel vectors p and q"?
After equation (5): "downsampling the image using a 10 × 10 kernel" ; please explain why the image is filtered and downsampled by a 10x10 kernel, and what kind of kernel is it, what does it look like? This is too vague and difficult to understand for the reader.
5) Verify equation (8); what represents "K.σ" (product between k and σ ? ) ; the symbol * represents a product? If yes, it should be replaced by point or x (* is commonly used for convolution).
After equation (8): "where .σ is" (?) - the point before where σ is a typo or what does it mean?
6) After equation (9): "features are decomposed into a Laplacian pyramid, while the corresponding weights are decomposed into a Gaussian pyramid" - please explain why, and what is the difference between the two decompositions.
7) Explain more clearly equation (11); what represents Lcon(G) and the notations Ex∼Pdata(x), Ey∼ Pdata (y)? Explain the difference between Φ(F(G(x))) and Φ(G(F(y))).
8) Equation (12) is not explained at all; what represent G, D, X, Y in Ladv(G,D,X,Y), log D(y), log(1 − D(G(x))) ? Equations (13) and (14) are not properly explained either.
In the linear combination in equation (17), specify the possible or optimal values for the coefficients c1, c2 and c3.
Comments on the Quality of English LanguageGrammar, language and style are correct. Make just the following corrections as suggested:
In Introduction:
Under such adverse underwater environments -> In such adverse underwater environments / Under such adverse underwater conditions
In section 2:
with the leap improvement -> with the rapid improvement
page 5:
and α in equations () and () -> and α in equations (1) and (2)
takes a value of 1 -> takes the value 1
scaling factor„ thus -> scaling factor, thus
After the sharpening step, the next step is fusion. -> After sharpening, the next step is fusion.
page 7: The generator and discriminator are composed of multilayer perceptions (?) -> multilayer perceptrons
page 11: trend, The process transitions -> trend. The process transitions ; UCIQE , combines -> UCIQE combines
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsAn actual research study - it is presented a new enhanced algorithm for underwater image improvement based on adversarial training.
Introduction is sufficient and presents main historical algorithms concerning the subject. Included references are relevant and sufficient but mainly authored by Asian researchers.
The used methods and models and algorithms are relevant as well as presented and described in a detailed style.
Results are presented in numerical and graphical way. There is analysis and description of the results.
Conclusions are relevant and supported by the results.
I would like to propose to the authors to avoid some repetitions of known methods either in introduction or in experimental section for better readiness. Some tables and figures are placed far from its text description (Table 5 and Figure 14 are in the Conclusion section) - this has to be corrected. A better and concise formulation of the original contribution of this study in the abstract and as the aim after introduction will improve the quality of the paper.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 5 Report
Comments and Suggestions for AuthorsThe paper deals with methods of improving the quality of photos taken in low light conditions underwater, affected by turbulence and showing blurring. The natural goal is to improve their brightness, contrast and sharpness. A number of techniques are described in the literature, which the authors describe in a clear way, together with their possible limitations, most often based on classical deep learning frameworks. The authors extend them with a color-enhancing preprocessing step, which then tunes the next stage of the computation by setting specific weight coefficients in the training and learning steps to maximize the quality of the photos in terms of brightness, contrast, sharpness and color fidelity while suppressing noise. Here they use a number of non-trivial methods, such as image fusion strategy, Laplacian filtering, entropy evaluation. Tests on a representative dataset in all cases show improvements over competing algorithms, which the authors document with tables and concise graphs, including detailed verbal evaluations.
The results evoke a number of questions:
· Can the proposed methods also be used to improve photographs that were not taken underwater, but under difficult lighting conditions (e.g., in fog, in the early evening, against sunlight)?
· Many of the operations to improve the quality of images (increasing brightness, contrast, sharpness, applying Gamma correction) are also offered by professional graphics programs such as Photoshop. Although their know-how is not available, would it be possible to indicate how the algorithm proposed by the authors would compare with the tools offered by post-professional graphics programs?
Some parts of the text could be completed or clarified:
Page 4, The meaning of the symbol $\alpha$ in Eq. (1) is not explained until Eq. (2), when it is stated that it takes the value 1.
p. 5: "$\alpha$ in equations () and () takes a value of 1" - references to the numbers of equations are missing
p. 5: weight $S(x,y)$ is not used in the following text
p. 12: "The UIQM is a linear combination of UICM, UISM, and UIConM". However, information about the determination of the coefficients $c_1$, $c_2$ and $c_3$ is missing in Eq. (17).
Format:
page 6, line 200: “dimensions..” – "dimensions." (one dot)
p. 7, l. 240: "At this point,the" – "At this point, the" (inserted space)
p. 11, l. 340: "similar trend, The process" – "similar trend, the process" (lowercase)
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsNo comments
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors' responses were limited to simply repeating the content that was already present in the original manuscript, without providing additional analysis or valuable experiments to demonstrate the reliability and innovativeness of this paper. In academic publishing, comprehensive and insightful responses to reviewers' questions are necessary to demonstrate the contribution to related field. This article failed to sufficiently demonstrate the credibility and innovativeness of the authors' views.
Comments on the Quality of English LanguageNo comments.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 2 Report
Comments and Suggestions for AuthorsNo comments.
Comments on the Quality of English LanguageNo comments.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
