Ship Recognition for SAR Scene Images under Imbalance Data

Round 1

Reviewer 1 Report

This paper manuscript introduces a ship recognition method on the basis of deep network for SAR scene images under imbalance data. The squeeze-and-excitation (SE) module is introduced for amplifying the difference features as well as reducing the similarity features among various SAR ship classes. A loss function Central Focal Loss (CEFL) based on depth feature aggregation is constructed to reduce the differences within classes.

From the detection results, it seems that the effect is well. I think it is helpful to improve the accuracy of ship detection. However, some major modifications are still necessary to further improve the quality of the manuscript. I hope that my review helps the authors to improve the manuscript. I recommend to addressing the following specific concerns and recommendations:

Abstract

Page: 1

“SAR” appear for the first time, please provide the full name.

Abstract

Page: 1

It is suggested to explain how much precision the method in this paper can improve compared with the traditional method at the end of the abstract.

1. Introduction

Page: 1

“ATR” appear for the first time, please provide the full name.

2. Method

Page: 4

Please introduce the specific methods of image preprocessing in 2.1.1

3. Experiments

Page: 8-9

Many factors, such as incidence angle, ship size, wind speed, and metocean parameters influence the ship detectability. Typically, it is easy to detect ships in high incidence angle, low wind speed, or low sea state. Figures 7 and 8 results show that the detection effect seems to be well, but they both seems to be in the low sea state. Has the author tested the effectiveness of your method under complex sea conditions? In addition, in nearshore complex background images, such as the port area containing strong reflective targets, is your method in this paper effective?

4. Discussion

Page: 12

Please explain the limitations of your method. Especially the detection effect of complex sea conditions and nearshore areas with strong reflective targets mentioned in the previous opinion. Whether there are missing ships and false alarms?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript proposes a ship recognition method for SAR scene images under unbalanced conditions, which can detect and classify the ships in SAR scene images. In the proposed method, target detection and classification are implemented in an integrated framework of deep neural network, and the loss function and network structure are improved. The effectiveness is validated by OpenSARShip and Sentinel-1 SAR datasets. The paper is well-written, but before it could be accepted for publication, I think the following problems should be considered:

1. Whether the performance of the model has been tested on other datasets?

2. The manuscript aims at ship targets detection and classification in SAR scenes under end-to-end model, the idea is novel, but section 2.3 points out that there is a class imbalance problem in recognizing ships in SAR scene images. Is there a causal relationship between the class imbalance problem and recognizing ships in SAR scene images? The authors should explain it more clearly to facilitate understanding.

3. In the manuscript, for the class imbalance problem in the SAR scene image recognition, a new loss function named Central Focal Loss is proposed based on the original classification loss function called Focal Loss, which achieves good results. However, it is seen from Equation 6 (p. 7) that the first half of the proposed new loss function still uses the optimal values of the parameters of the original Focal Loss, the feasibility of such a selection should be explained.

4. I suggest that add the results of Focal Loss to Table 1 (p11), so that the difference between Central Focal Loss and Focal Loss can be more clearly manifested.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The paper carries out some novelties and contributions. It seems to be eligible for publications after doing the minor modifications as follows

The abstract must be revised because is too ambiguous and it must refer to the problem properly.

Although, the paper is written concisely, the writing aspect still conveys some problems related to technical language of the field and the expected knowledge of the journal's readership.

The paper doesn’t cite appropriate and up to date literature sources. Also, the presented papers are not enough for an enriched literature. Although the findings and results are clearly stated, I believe that the authors can extend this section more than before

At first, the above major modifications should be done, accurately. Then, the paper must be reviewed, again. The raised comments in “Summary, Conclusion and Future Research” should be considered, perfectly. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx