Review Reports

Round 1

Reviewer 1 Report

The authors contribution is technically sound. The paper structure is well articulated.

The papers contains mild innovative components. 

The merit of this contribution contribution can be further improved  by introducing physics (physical phenomena and underlying mechanisms description) to authors' modelling. As result, there room for further tunning of the kernel (transfer learning-based recognition method) used to train the backbone network.

As said earlier, this is a valuable study.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper can effectively improve the accuracy of sonar target recognition, and it has certain innovation.

This paper does not describe in detail the data association processing rules of optics and sonar and the target recognition of correlated strong noise states. It is suggested to add this part.

The database mentioned in the paper（https://github.com/HHUCzCz/-SeabedObjects-KLSG--II） was not found.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this paper, the authors mainly discuss multi-modal multi-stage underwater side-scan sonar target recognition based on synthetic images. To some degree, the work is interesting. However, the methods in this paper are just verified by well reconstructed targets. The conclusions based on current experiment are not convincing. Actually, many sonar images are seriously contaminated by noise. The authors should test their method based on these images.

1. Line 31 on page 1, the authors said that ‘sonar images are affected by various types of noise, such as Gaussian, impulse, and speckle noise’. Here, some citations about underwater impulsive noise like alpha, Middleton Class A and Class B should be carried out.

2. Target recognition based on sonar image is a hot topic. However, the methods about the target recognition are not comprehensively described in section 1 and section 2. The authors should comprehensively conduct the literature review in detail. More importantly, the advantages and disadvantages should be concentrated.

3. As the authors said in the line 31 on page 1, sonar images are affected by various types of noise, such as Gaussian, impulse, and speckle noise. And the whitening transformation process is shown in Eq.(2). The reviewer wanders to know whether this transformation is suitable for all various types of noise such as Gaussian, impulse, and speckle noise. If not, the reviewer wanders to know how to preprocess impulse and speckle noise individually.

4. In Eq. (7), the style control factor α is introduced. However, the authors did not clarify how to determine this parameter in detail. That is to say, the readers cannot know how to choose this parameter. At this point, the authors should clearly clarify the determination of this parameter.

5. On page 5, the VGG19 network is used by the authors suddenly. The authors should firstly clarify the reason why to choose VGG19 network.

6. In section 4.3, the Fig. 9 shows that the accuracy of authors’ method is always 1, which is not convincing based on a single experiment with well reconstructed targets. In practice, many reconstructed targets are seriously contaminated by noise, and the target recognition is a hard job. At this point, more complicated image should be used to validate the authors’ method.

7. The references 15 and 42 are the same.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The authors well address the reviewer's comments, and the reviewer suggests that this paper is accepted.