Aggregate Channel Features and Fast Regions CNN Approach for Classification of Ship and Iceberg

Round 1

Reviewer 1 Report

Aggregate Channel Features and Fast Regions CNN Approach 2 for classification of Ship and Iceberg

1) Since none of the authors appears to be native English speaker, I suggest a total revision of the English language and style for the entire article using Grammarly or another specialized tool. The grammar of the paper should be checked.

2) In Section 1, PCA and K Methods are mentioned as supervised learning methods, but these methods are unsupervised.

3) Some citations are too old in the related works section. I hope the author can modify this section to cover recent works.

4) Dataset descriptions should be given in a Table.

5) Although the titles of both sections 3 and 4 include material and dataset expressions, there is no explanation of SAR image data. The name of Chapter 3 should be changed to "Material and Method" and some information about the dataset should be given, adding related citations about the dataset in subsection 3.1.

6) Why was the confidence score of ROP selected as >50? It should be explained using references from the literature.

7) The steps of the proposed algorithm and the mathematical equations should be shown more technically.

8) The proposed method should also be compared with Faster RCNN, which, unlike Fast RCNN, uses RPN for region proposals.

English language should be improved.

Author Response

Dear Reviewers,

First of all, we would like to take this opportunity to express our sincere thanks to the editor and reviewers for their valuable time to make suggestions to improve the quality of this paper. We have revised the manuscript in accordance to the reviewers’ comments, and the changes in the revised manuscript are highlighted by yellow. In the following, we give a point-by-point reply to the reviewers’ comments:

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 2 Report

This paper proposed an object detection method based on channel feature aggregation and FRCNN for detecting icebergs and ships from SAR images. Interesting topic, indeed. However, it is not well-written. Many contents are not easy to follow, and the novelty is limited. My detailed comments are as follows:

 (1) The manuscript is not well-formatted, especially for the equations, algorithm, and references. The figures are low-resolution and not readable.

(2) The sample is of size 75 x 75 pixels in the experimental dataset. I doubt whether deep neural networks are applicable to such small-size samples.

(3) The motivation for developing an ACF detector is unclear. Although it can eliminate some unwanted regions, the contribution of FRCNN is also weakened. Besides, the novelty is limited because FRCNN is a common choice for object detection.

 

(4) The experimental results are not convincing enough. More datasets are needed, and other object detection methods such as SSD, CenterNet, Yolo, and DETR should be compared.

Not easy to follow. Writing flow and sentence expression should be optimized.

Author Response

Dear Reviewers,

First of all, we would like to take this opportunity to express our sincere thanks to the editor and reviewers for their valuable time to make suggestions to improve the quality of this paper. We have revised the manuscript in accordance to the reviewers’ comments, and the changes in the revised manuscript are highlighted by Green. In the following, we give a point-by-point reply to the reviewers’ comments:

Please See the Attachment.

Author Response File: Author Response.doc

Reviewer 3 Report

This paper has proposed the aims to locate and classify both ship and iceberg in a given SAR image with the help of Deep Learning (DL) and non-DL methods.  Authors used ACF detector to extract the candidate areas and used fast regions CNN to classify the objects. I have the following issues to improve this manuscript as: 

- Authors mentioned "Fast Regions CNN (FRCNN)". As I know, Fast R-CNN exists already with region proposal function. What is the difference from it.

-   In section 2, authors described some related works for small object detection. But the mentioned works were very basic and old. More literature survey is needed as:

. FE-YOLOv5: Feature enhancement network based on YOLOv5 for small object detection,  Journal of Visual Communication and Image Representation, Volume 90, February 2023, 103752

.  "FPN-GAN: Multi-class Small Object Detection in Remote Sensing Images," 2021 IEEE 6th International Conference on Cloud Computing and Big Data Analytics (ICCCBDA), Chengdu, China, 2021, pp. 478-482, doi: 10.1109/ICCCBDA51879.2021.9442506.

 . ssFPN: Scale Sequence (S2) Feature Based-Feature Pyramid Network for Object Detection, Sensors (MDPI), Sensors 2023, 23(9), 4432; April 30, 2023

- In the proposed scheme, what is the role of GAN? I didnot find the reason of it.

- Figure 2 shows the proposed ACF detector. but how to set the RoIs? Also, Figure 3 described the FRCNN. But the existing Fast RCNN can operates same and well. Why you need to design a new? What advantages of the proposed method?

- Most figures are very low quality. Enhance the resolutioin and quality.

- In experiments, authors used teh existing methods to compare the performance without citation of them. 

- In Fig. 12, there was no label in X and Y-axis. 

- Also, authors should analyze the effect of some variables and the performance in detail. Then we can see what is merit or drawback.

- More recent methods should be added to compare the performance evaluation.

- There are many typos. All should be polished. For example, "~ of layers that is Ƥ={ Ƥ1, 245 Ƥ2, Ƥ3, Ƥ4,…… Ƥn,}.The proposed system used 3 LUV colour ~"

English writing should be re-checked very carefully.

Author Response

Dear Reviewers,

First of all, we would like to take this opportunity to express our sincere thanks to the editor and reviewers for their valuable time to make suggestions to improve the quality of this paper. We have revised the manuscript in accordance to the reviewers’ comments, and the changes in the revised manuscript are highlighted by Blue. In the following, we give a point-by-point reply to the reviewers’ comments:

Please see the Attachment.

 

Author Response File: Author Response.doc

Round 2

Reviewer 1 Report

In this version of the study, the authors covered most of the issues. 

Thus, the study should be accepted.

Author Response

Dear Editor and Reviewers,

First of all, we would like to take this opportunity to express our sincere thanks to the editor and reviewers for their valuable time to make suggestions to improve the quality of this paper.

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 2 Report

The authors have made revisions to the descriptions of the methodology and experiments analysis. However, the novelty is limited and the presentation is not convincing enough.

1) Almost all equations are not well formatted. The experimental results are not well illustrated and the figures (e.g., Fig. 6, 8-10) have very low resolutions.

2) The test images are of size 75x75 pixels. I do not think clipping images into a small size is a good choice for deep learning. Today, the development of deep learning methods focuses on large amounts of images with large sizes.

3) FRCNN is a classical object detection method, and many studies have demonstrated its lower performance compared with more advanced approaches such as YOLOv7 and Transformer-based object detectors. Although the authors improve the accuracy of regional proposals of FRCNN, only one dataset with small sizes of images is not convincing enough for a scientific research paper. 

Overall, the proposed method has a very limited application scope, and I recommend the author redesign the research.

Not easy to follow. 

Author Response

Dear Reviewer,

First of all, we would like to take this opportunity to express our sincere thanks to the editor and reviewers for their valuable time to make suggestions to improve the quality of this paper. We have revised the manuscript in accordance with the reviewers’ comments, and the changes in the revised manuscript are highlighted by yellow.

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 3 Report

As my check, all issues have been well addressed. 

it needs minor correction.

Author Response

Dear Reviewer,

First of all, we would like to take this opportunity to express our sincere thanks to the editor and reviewers for their valuable time to make suggestions to improve the quality of this paper. We have revised the manuscript in accordance to the reviewers’ comments, and the changes in the revised manuscript are highlighted by Green.

Please see the attachment

Author Response File: Author Response.doc