Reinforced Residual Encoder–Decoder Network for Image Denoising via Deeper Encoding and Balanced Skip Connections
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe material is clearly presented, presenting an iterative improvement on denoising techniques. It would be good to have more detail on the limitations of the system, as well as a more guided approach to the image figures presented, in which it was not immediately clear how the representative images were different from each other. Since the system was compared with other denoising approaches in terms of PSNR and SSIM, it would have been good to include images processed with these approaches to provide a more direct presentation of the improvements made.
---
The paper presents a modification of the REDNET noise reduction algorithm in which additional convolution levels have been added, and additive skip layers have been replaced with averaged concatenated skip layers.
This is presented as an iterative improvement on this architecture, with metrics (PSNR, SSIM) comparing it with the base algorithm demonstrating improvements demonstrating improvement relative to the baseline REDNET.
However, there is no statistical analysis of these tests, means (but no standard deviation or error) are presented from 100 images from one dataset and only 10 from another, and there is no discussion of direct relevance of this level of improvement for actual de-noising tasks. Thus it is not clear as presented what the significance of this improvement for the field is.
Additionally, Figures 4 and 5 are intended to show the difference between ground truth and noisy images (Fig 4), and ground truth, noise, and cleaned images (Fig 5), however it is difficult to tell by eye what the differences in the images are, at least in the pdf for review. It would be beneficial if there are significant differences between the images that are not so readily apparent, that the authors draw attention to them in either the text or in the presentation of the images.
For example, a box indicating an area of the image to focus on to observe the impact of the algorithm. Additionally, it would be helpful for visual clarity to have a border between the images in Figures 3 and 4, though probably a smaller border than that used in Figure 5. It was also unclear to me which datasets or subsets were used for training - training is listed in Figure 2, but the training dataset was not clearly identified in the text.
Overall, the paper represents a modification of an existing approach that yields improved results based on the two metrics used, but a small dataset was compared, there was no statistical analysis, nor was the improvement shown to be significant in relation to some specific de-noising task. It was also not clear which data the system was trained on.
Some editing issues I noticed as well:
Fig 2 – comparison is mis-spelled in the Figure text. Section 4.2 has an unnecessary comma after the word "network": "The architecture of the proposed network, is shown in Figure 2.". Section 4.2.1 is missing the letter “I” in “in”
Author Response
Please check response in the attached PDF
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe article is clearly written and its contents are well organized. The introduction makes a good job of introducing the problem of image denoising and succeeds at underlying the significant step further to the field brought by machine learning based approaches, specially the ones based in the use of neural networks.
The section describing related work presents several key advances in the field and is supported by relevant citations. I feel this section could be somewhat more organized, using subsections for related approaches and connective text to avoid the feeling that we are being presented with a list of unconnected advancements. While I enjoyed the background section, and believe it to be useful to readers not familiar with the field, I think it could be better developed and probably presented before the related work section, as some of introduced concepts could be useful to better follow this latter section.
The article proceeds to introduce the proposed architecture, the R-REDNet model, going into some detail over the suggested improvements. The section is written with enough detail to allow the reproduction of the described work. The experimental setup is well described and I consider it to be adequate to evaluate the usefulness of the contribution to the field. The use of PSNR and SSIM metrics for evaluation is adequate and the use of two different datasets is sufficient to identify potential improvements from results obtained by previous approaches.
Results seem to validate the new model, specially in terms of PSNR performance. The discussion of the effects of the different additions to the REDNet model are interesting and help to access the contribution of each modification to the overall results. I have some doubts about the usefulness of the images presented to help illustrate the results, but they don’t negatively affect the article in any way.
Overall I think this is an interesting approach, with experimental results that strongly support the improvements made to REDNet, and its many strong points clearly outweigh some minor issues.
Author Response
Please check response in the attached PDF
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsIn this reviewed manuscript, the authors present an enhanced denoising residual network (REDNet), termed R-REDNet by incorporating deeper convolutional layers in the encoder and replacing additive skip connections with averaging operations to improve feature extraction and noise suppression. The manuscript is well written and of interest to the scientific community in areas of imaging and characterisation. the authors consider the following comments:
- Lines 1-5 of the abstract should be removed and taken to introduction section. The abstract should start from, "This paper presents ..."
- Sections 1-3 should be merged into a single section under introduction and streamlined into much more condensed and fewer pages.
- Section 5.3 should include results of comparison with existing methods such as REDNET, RESNET etc.
- Tables of nomenclature and acronyms should be included
Author Response
Please check response in the attached PDF
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for Authors- Why does replacing additive skip connections with averaging operations enhance denoising performance? Does it reduce noise amplification? How does it impact feature fusion?
- How does increasing the depth of the encoder improve feature extraction?
- What is the computational complexity of R-REDNet compared to standard CNN-based denoising models?
- How does increasing encoder depth affect model size and training time?
- What are the major limitations of R-REDNet?
- How can R-REDNet be extended to video denoising?
- Hybrid approaches combining Transformers or self-attention mechanisms with R-REDNet could be explored in future work.
Author Response
Please check response in the attached PDF
Author Response File: Author Response.pdf
Round 2
Reviewer 4 Report
Comments and Suggestions for Authors1. The title needs to be revised the match with the proposed model.
Abstract
2. Do not use so many abbreviations in the title, abstract, and elsewhere. An abbreviation is used only if the term appears at least five times in the main text. (The Abstract, conclusion, figures, and tables don't count.) If the term or phrase is used only two, three, or four times it should not be abbreviated (The Chicago Manual of Style). Thus, should the abstract should be revised as per the given instructions.
Introduction
3. The introduction should present the proposed method and why it is needed, rather than only pieces of information which are already available in the general media.
4. Before going to present related works, you should mention your target approaches, and lastly, your concluding remarks with mention the research gaps, to which you are going to contribute.
Related work
5.The provided literature review seems to be a list and not an in-depth analysis focused on the paper's objectives. Please reorganize the already reported literature review. In addition, there is a lot of literature being done on the different outbreak forecasting issues. In case the authors should expand the related work section as follows:
-A new COVID-19 classification approach based on Bayesian optimization SVM kernel using chest X-ray datasets.
- A hybridized LSTM-ANN-RSA based deep learning models for prediction of COVID-19 cases in Eastern European countries
Author Response
Dear Reviewer, please find all our responses in the attached PDF
Author Response File: Author Response.pdf
