SD-CapsNet: A Siamese Dense Capsule Network for SAR Image Registration with Complex Scenes
Round 1
Reviewer 1 Report
This manuscript proposes a Siamese Dense Capsule Network (SD-CapsNet) for SAR image registration. A texture constraint-based phase congruency is proposed to detect uniformly distributed keypoints of high repeatability. SD-CapsNet is used to implement feature descriptor extraction and matching which can get better semantic information. In this manuscript, the content is detailed and the experimental data are sufficient. I have some comments and suggestions.
(1) In the Page 8, Fig.7 shows the difference in back propagation between traditional connected way and densely connected way. Forward propagation of densely connected way leads to differences in back propagation. It would be better to add forward propagation arrows of densely connected way in Fig.7(b).
(2) In the lines 116-117 of Page 4, the keypoint is discarded when the RI-LBP value of the detected keypoint is higher than the average value of global RI-LBP. But In the Page 7, the Formula(5) shows that ${KP_{final}}$ value would be positive when the RI-LBP value is higher than the average value of global RI-LBP. Is there a conflict between these two representations?
(3) In the Page 9, the Formula(8) contains the variate $l$ in $\sum\limits^{32}\limits_{l=1}$. But the use of the variate $l$ could not be seen in Formula(8).
(4) In the lines 224-225 of Page 9, the feature descriptors are reconstructed as images. It is suggested to clarify how to reconstruct the images by using the feature descriptors.
(5) In the Page 5, the outputs of low-level capsules, medium-level capsules and high-level capsules are concat to be sent to primary capsules in the SD-CapsNet part of Fig.3. But in the Page 8, the input of primary capsules comes from three convolutional layers. So, how the SD-CapsNet densely connects? It is suggested to refiine the SD-CapsNet schematic diagram in Fig.3.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
In this paper, the authors proposed a SAR image registration method based on an improved capsule network, which consists of a texture constraint-based phase congruency keypoint detector and a siamese dense capsule network, to address the challenging task of SAR image registration in complex scenes. SAR image registration is quite essential to the SAR data interpretation. The topic is meaningful to the readers. From a general perspective, the manuscript is well written and organized, and it can be easily read and understood. I have some following suggestions as outlined below.
l It would be better to give some quantitative evaluation measures of the proposed method in the abstract.
l In the introduction, more newly published articles about the SAR image registration should be analyzed and discussed.
l In Figure 5, the authors give the results of keypoint detection after texture constraint, and it can be seen that RI-LBP works well. Therefore, it is suggested that the authors give the feature map of RI-LBP.
l Does the proposed method have any drawbacks compared to state-of-arts? Can the author talk about the future work to potentially improve the proposed method?
l Is it possible for the publication of the code for the re-implementation of the proposed method?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
See attachment
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
(1) The authors mentioned in the abstract: “our proposed method outperforms other state-of-the-art methods in terms of the number of correctly matched keypoint pairs (NCM), root mean square error (RMSE) and registration time.” It would be better if the quantitative comparisons can be given specifically.
(2) The authors mentioned in the abstract: “we define a novel distance metric for the feature descriptors in capsule form and feed it into the Hard L2 loss function for model training”, However, in the fourth section of the article, the authors only explained the effectiveness of the TCPC Keypoint Detector and the SD-CAPSNet Feature Descriptionor Extractor, but did not explain the effectiveness of this distance metric.
Author Response
Comments and Suggestions for Authors
- The authors mentioned in the abstract: “our proposed method outperforms other state-of-the-art methods in terms of the number of correctly matched keypoint pairs (NCM), root mean square error (RMSE) and registration time.” It would be better if the quantitative comparisons can be given specifically.
- Dear reviewer, many thanks for your comment. In the revised manuscript, we have given the quantitative results of the comparison method on NCM and RMSE.
- 2. The authors mentioned in the abstract: “we define a novel distance metric for the feature descriptors in capsule form and feed it into the Hard L2 loss function for model training”, However, in the fourth section of the article, the authors only explained the effectiveness of the TCPC Keypoint Detector and the SD-CAPSNet Feature Descriptionor Extractor, but did not explain the effectiveness of this distance metric.
- Many thanks for your comment. In fact, during our experiments, we tried to use Euclidean distance, L2 distance, and cosine distance as distance measures for the feature descriptors in capsule form, but none of them made the network converge. This is due to the special form of feature descriptors, which will lead to confusion of useful information of keypoints after being reshaped into one-dimensional vectors. Therefore, in this manuscript, we cannot provide comparative results for other distance measures.In the revised manuscript, we have explained this issue in detail.
Reviewer 3 Report
Dear authors,
Thank you for answering to almost all my points. Concerning the last point, I was asking you to generate a figure with a selected area (a crop), showing the reference image and the results of the different image registration. In this way the reader will be able to better appreciate the improvement by applying the proposed algorithm. There aren't many examples in the literature, but a visual result of a detail can strengthen your paper.
Author Response
Thank you very much for your comments concerning our manuscript entitled “SD-CapsNet: A Siamese Dense Capsule Network for SAR Image Registration with Complex Scenes”. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our research. We have studied the comments carefully and have made corrections that we hope to meet with approval. Corresponding changes have been made to the manuscript with red color. Responses to the reviewers’ comments are presented as follows.
Comments and Suggestions for Authors
- Thank you for answering to almost all my points. Concerning the last point, I was asking you to generate a figure with a selected area (a crop), showing the reference image and the results of the different image registration. In this way the reader will be able to better appreciate the improvement by applying the proposed algorithm. There aren't many examples in the literature, but a visual result of a detail can strengthen your paper.
- Dear reviewer, many thanks for your comment.We apologize that your comments were incorrectly understood in the previous response. In the revised manuscript, we have added a enlarged result figure of a selected area of the registration results to explain the link between the keypoint matching result and the registration result as a way to further indicate the superiority of the proposed method.
