Adaptive-Attention Completing Network for Remote Sensing Image
Round 1
Reviewer 1 Report
In the abstract, to emphasize the results obtained, include the databases that are analyzed, and the metrics shown in the article.
In the introduction, it is not usual for explanations to be given with figures (with a very extensive caption, the description can be done when citing the reference), although I understand the need to explain your own method and its advantages over another, it seems appropriate to do so in the part where he exposes his methodology.
Also in this section, citations are repeated in the same paragraph, they are not necessary. the contributions of the article correspond to the conclusions.
Regarding the methodology, which is an innovation to the attention model, I should introduce it a little more: suddenly the query, key value appear without any other presentation.
In the paragraph of line 197 it does not describe H, W and C, it does so advanced in the text.
The formulation seems well founded. Tables could be adjusted.
The results are well exposed, however the analysis of the authors in the conclusions should be more conclusive, I find it brief for the extensive work they have done.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
1. Comparison of the proposed technique should be made with the existing techniques in the same area i.e. deep learning based techniques
2. Proposed technique merely seem just an implementation of existing deep learning model RS images . It lacks in novelty.
3. Literature review of the work can be improved.
4. Robustness of the proposed technique is neither discussed nor compared with existing techniques. The complexity and cost of the supposed method must be given.
5. Pseudocodes will be useful for future readers to reproduce your method.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
In this paper authors propose Adaptive Attention (Ada-attention), based on self-attention and based on Ada-attention, customize a u-shaped Adaptive-attention Completing Network (AACNet) to reconstruct missing regions for remote sensing images. Experiments on several digital remote sensing and natural image datasets demonstrate that the proposed AACNet outperforms the state-of-the-art methods in quantitative and qualitative measures.
Remarks for the paper:
1. Several other objective measures might be added to compare the proposed method with existing methods, as well as the discussion about their results:
-full reference: MS-SSIM [1], FSIM [2], PSNR-HVS-M [3], LPIPS [4]
2. Please can you explain about the image resolution and does it affect the proposed method? I.e. will it be equally effective for both smaller and higher resolution images? Also, can it be used for the multispectral images?
3. Please can you also give inference time for the proposed method?
4. Can you share the proposed implementation code, to be publicly available?
[1] Z. Wang, E.P. Simoncelli, A.C. Bovik, "Multiscale structural similarity for image quality assessment", 37th Proc. IEEE Asilomar Conf. on Signals, Systems and Computers, Vol. 2, pp. 1398-1402. (2003)
[2] Lin Zhang, Lei Zhang, X. Mou and D. Zhang, "FSIM: A Feature Similarity Index for Image Quality Assessment", IEEE Trans. Image Processing, vol. 20, no. 8, pp. 2378-2386, 2011.
[3] N. Ponomarenko, F. Silvestri, K.Egiazarian, M. Carli, V. Lukin, "On Between-Coefficient Contrast Masking of DCT Basis Functions", CD-ROM proceedings of Third International Workshop on Video Processing and Quality Metrics for Consumer Electronics VPQM-07, January, 2007
[4] R. Zhang, P. Isola, A. A. Efros, E. Shechtman and O. Wang, "The Unreasonable Effectiveness of Deep Features as a Perceptual Metric," 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 2018, pp. 586-595, doi: 10.1109/CVPR.2018.00068.
Summarizing, the paper proposes a new method to reconstruct missing regions (i.e. due to the clouds) on the remote sensing images. In comparison with other methods, the proposed method has better PSNR, SSIM and MAE measures, compared with uncorrupted images.
In the experimental section more objective measures might be added to compare the newly proposed method with the existing methods. Proposed references can be also added then. Statistical significance of the mean values can be also added. In the ablation study, different resolution images might be tested to check the robustness of the proposed method.Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Paper review
The authors introduced an Adaptive-attention Completing Network (AACNet) with an Ada-attention module to address the inpainting of remote sensing images, where Ada-attention adaptively selects more relevant features and models more informative dependencies for restoring missing data. The Ada-attention uses a data-dependent offset position subnet to adapt to inputs and calculate the attention map as a softmax of the dot product of the query and key. Experiments show that the AACNet outperforms other inpainting methods and effectively removes clouds in remote sensing images.
The paper is technically sound and provides a new architecture, as claimed. It is well-motivated and addresses a significant problem in remote sensing. The authors describe their architecture in detail and provide several experiments to assess its performance.
However, several issues need to be addressed.
Regarding the network architecture (Figure 2), the authors proposed using the attention module between the end of the downsampling layer and before the input of the upsampling layers.
The authors should have discussed whether adding the attention modules between intermediate layers would be beneficial.
The name assigned to the proposed approach as adaptive attention needs to be clarified. Why is this attention module tagged as adaptive? This could be clarified in different places in the paper to unveil the logic to the reader.
In the experiments, there is a need to add a table presenting all datasets and their characteristics.
Also, the GitHub repo of the code must be referenced to be able to reproduce results.
The authors compared it with several other approaches, which is good. However, there are several results and experiments tables, and there is a need to aggregate all results to have an overall view of the average performance among all results and get the general lessons learned. A good approach is to use a Sunburst to aggregate and present all results into an illustrative graph.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Authors have answered all the questions.
Author Response
We would like to show our sincere appreciation to you for valuable feedback and contributions to our manuscript.
Reviewer 4 Report
The authors have addressed most of my comments, but there is still an issue with the aggregated results in Figure 4. The reviewer finds the sunburst in Figure 4 to be unreadable and suggests that the size of the sunburst should be increased to improve readability. Additionally, the colors used in the sunburst are not mapped to the values, and the reviewer recommends using a unified gradient color that varies in intensity based on the value. Once these issues are fixed, the paper should be ready for publication.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
