A Novel Deep Learning Network with Deformable Convolution and Attention Mechanisms for Complex Scenes Ship Detection in SAR Images

Round 1

Reviewer 1 Report

The manuscript “A novel deep learning network with deformable convolution and attention mechanisms for complex scenes ship detection in SAR images” proposed a method for complex scenes ship detection in SAR images, and tested on the public SAR image ship dataset showed that the detection accuracy of the proposed model. The proposed method is innovative and effective and will save time for detecting target as samples in the deep learning process. However, there are problems in the manuscript that need to be addressed and clarified.

1. The content between the abstract and introduction sections is poorly connected and the logic is not very clear.

2.  From the introduction section Figure 1 and Figure 2 are samples of datasets that will move these two figures into a “dataset section” to represent the dataset containing that kind of ship such as complex scenes and high-density small ships.

3.  A scientific paper always has 5 sections such as, introduction, material and method, result, discussion, and then the conclusion, but this paper is completely missing the organization. So therefore the structure and organization of this article are poorly organized, and the expression is not clear. The proposed method suggests adding more details, such as a description of the backbone feature extraction network of the proposed method. They should describe in more detail their additional architecture.

4.  A more in-depth literature review is needed, especially on state-of-the-art methods. The literature should be reviewed in light of the difficulties and challenges addressed in this paper, rather than being simply listed, which leads to insufficient emphasis on the key points. 

5.  The experimental results are not convincing, given only adding an attention layer with high accuracy. Substantial work (experiments and data) is required to validate the model improvements.

6.  The motivation for using the GloU loss function is ambiguous. It is unclear what impact it has on the results and whether this can be explained. Additionally, in the experimental results, it is recommended to compare other loss functions such as CloU, EIOU, and normal loU loss.

7.  Model evaluation: is cross-validation performed during the training process only using training and develop set or using all dataset? Usually, the final model is evaluated using independent testing set, producing confusion matrix, then calculate accuracy measurement metrics, for example, overall accuracy, user’s and producer’s accuracy, Kappa. For deep learning approach, it might use different metrics like F1 score.

8.  Figure 6, 7 needs a more logical explanation to support the proposed model motivations.

9.  Hyperparameters have a significant impact on machine learning methods. The authors should add a new section in the paper to discuss how to choose hyperparameters for their method or other deep learning models, making the structure of the paper clearer, such as the input size of the images, the ratio of dataset division, training parameters, and specific training environment.

10.  With Yolov 7, and 8 released, it seems necessary to judge whether it is correct to say that currently, the newer target detection model line 274. It would be nice to highlight the advantages of v5 over the latest Yolo.

11.  It seems logical to compare with detection studies using CNNs.

12.  The font size in Table 4 is inconsistent.

13.  The detection results and discussion of the proposed method should be revised to demonstrate the derivation of the current research, and the description should be clearer and more comprehensible.

14.  If am not wrong, please clear see the line 309 and line 394 as the same section insert “3”. It is requesting to your author for clear check double time the framework of the paper and check against the template of the journal.

15.  Conclusion should be re-written to 1) explicitly describe the essential features/advantages of the review that other reviews do not have, and 2) describe the limitation(s) of the review.

The authors need to mind the English language the manuscript is very poorly written.

Author Response

We would like to thanks for your hard work and excellent comments and suggestions during the reviewing process. We have gone through your comments carefully and made the corresponding corrections to the manuscript. It has taken a considerable effort to obtain and process additional datasets in the literature and change the manuscript accordingly, which considerably improved the quality and overall presentation of our work. The main corrections to the manuscript and reviewers’ reply are discussed in attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors sought to combine deformable convolution with channel attention mechanism to achieve improved performance on small near-shore ship detection, which is complicated by near-shore complex backgrounds. The work is of interest to potential audiences, but I still have some concerns as listed below.

 

1. The review of related works is not thorough. There are recent works based on deep learning for ship detection that need to be discussed, which is listed below. Also, some of these works have open-source codes, please compare the performance of your model to these works.

[1] Rostami, M., Kolouri, S., Eaton, E. and Kim, K., 2019. Deep transfer learning for few-shot SAR image classification. Remote Sensing, 11(11), p.1374.  

[2] Zhang, Y., Guo, L., Wang, Z., Yu, Y., Liu, X. and Xu, F., 2020. Intelligent ship detection in remote sensing images based on multi-layer convolutional feature fusion. Remote Sensing, 12(20), p.3316.  

[3] Li, L., Zhou, Z., Wang, B., Miao, L., An, Z. and Xiao, X., 2021. Domain adaptive ship detection in optical remote sensing images. Remote Sensing, 13(16), p.3168.  

[4] Sun, Z., Meng, C., Cheng, J., Zhang, Z. and Chang, S., 2022. A Multi-Scale Feature Pyramid Network for Detection and Instance Segmentation of Marine Ships in SAR Images. Remote Sensing, 14(24), p.6312.

2. Could you also test your model on other more standard SAR datasets such as HRSID.

3. The conclusion and discussion part is not thorough.

4. English should be improved.

5. Commas (‘,’) or periods (‘.’) are missing at the end of all formulas.

6. Periods (‘.’) are missing at some figure legends, such as Figure 4 and 7.

7. Table 1-3 should be placed to the right.

English must be improved.

Author Response

We would like to thanks for your hard work and excellent comments and suggestions during the reviewing process. We have gone through your comments carefully and made the corresponding corrections to the manuscript. It has taken a considerable effort to obtain and process additional datasets in the literature and change the manuscript accordingly, which considerably improved the quality and overall presentation of our work. The main corrections to the manuscript and reviewers’ reply are discussed in attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The study sought to address the problem that nearshore ship targets in Synthetic Aperture Radar (SAR) images are affected by background clutter, low detection rate of ship targets in SAR images, and high false alarm and missed detection rates of small-scale ship targets. A novel deep learning network with deformable convolution and attention mechanisms was proposed for near-shore ship target detection in SAR images with complex backgrounds. Test experiments on the public SAR image ship dataset showed that the detection accuracy of the proposed model provided in this study was 87.95% in complex scenes, which is 8.46% higher than that of the original target detection model. The detection accuracy for small-scale ship targets was 95.14%, which was 12.28% higher than the original target detection model. However, the following problem need to be improved.

(1)   The English of paper is poor, please polish it again.

(2)   Each variable in the equation should be given the meaning.

(3)   Please highlight the novelty of the proposed method.

(4)   Why use Deformable CNNs?

(5)   Please review some recently method, such as

[*]FINet: A feature interaction network for SAR ship object-level and pixel-level detection [J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-15.

[**]Scale in Scale for SAR Ship Instance Segmentation[J]. Remote Sensing, 2023, 15.

[***] BANet: A balance attention network for anchor-free ship detection in SAR images [J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-12.

(6)   Please give all parameters of the proposed method.

(7)   More experiments should be given to show the advantage of the proposed method.

   The English of paper is poor, please polish it again.

Author Response

We would like to thanks for your hard work and excellent comments and suggestions during the reviewing process. We have gone through your comments carefully and made the corresponding corrections to the manuscript. It has taken a considerable effort to obtain and process additional datasets in the literature and change the manuscript accordingly, which considerably improved the quality and overall presentation of our work. The main corrections to the manuscript and reviewers’ reply are discussed in attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The paper proposed a modified FPN network, which with the addition of channel attention and deformable CNN modules seems to achieve improved small target detection, and the reduction of false positives caused from land features in near-shore scenarios.

I have primarily one specific comment to make, and it revolves around clarifying/discretising changes made to the FPN from the attention and convolutional CNN contributions.

In discussing the results of Tables 2 and 3, the authors should clarify what the results obtained for FPN only reflect; is this with the use of the updated FPN discussed in the beginning of Section 2 sans the attention/convolutional CNN mentioned later on? Could you elaborate more on the difference in effectiveness between the FPN architecture of other proposed methods and yours?

I feel the English language use is of satisfactory standard, there's a few small typos, a repeated sentence I saw somewhere etc but I trust these will be edited in the final proofread round.

Author Response

We would like to thanks for your hard work and excellent comments and suggestions during the reviewing process. We have gone through your comments carefully and made the corresponding corrections to the manuscript. It has taken a considerable effort to obtain and process additional datasets in the literature and change the manuscript accordingly, which considerably improved the quality and overall presentation of our work. The main corrections to the manuscript and reviewers’ reply are discussed in attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have not addressed my minors from the last round.

The authors have not addressed my minors about the language from the last round.

Reviewer 3 Report

The revision is good.