A Texture Feature Removal Network for Sonar Image Classification and Detection

Round 1

Reviewer 1 Report

In this paper, the authors presented an automatic system based on transfer learning using feature separation function that can help identify and classify images collected using side-scan and acoustic sonar systems. The methods, results and discussion are for the most part sound. The authors used novel machine learning platforms to help clarify images and the end-result is the ability to determine more accurate descriptions of underwater structures. The justifications of using transfer learning combined with the other described platforms seem sound and reasonable. The idea presented is very interesting and can be useful in image recognition applications. Overall, the paper is well written and clearly presented but some areas of improvement are needed. See attached pdf for specific comments and suggestions for improvement.

Comments for author File: Comments.pdf

Author Response

Dear reviewer:

Thank you very much for your thorough and detailed review of our paper，we have revised the paper in response to your comments and suggestion, including adding a brief description about the proposed method, change the usage style of references，fix the caption of the figures, corrected some misused words and modified the format of the reference in accordance with the requirements of the journal.

We have moved the t-SNE visualization part to result section according to your opinion.

And for your last comment, “Are there any plans to market this technology?”, we are going to make the code public so that it can be used as a reference for other researchers.

The revised paper will be uploaded to the system.

Thanks again.

Best regard.

Reviewer 2 Report

In this paper, a method is proposed to process remote sensing images and sonar images, retain contour features, remove color and texture features, by doing this, the domain gap could be narrowed between the two types of images, so as to achieve the purpose of knowledge transfer, which is innovative and progressiveness.

 

Aiming at the problem that directly using whitening transform could not effectively remove domain specific features (color and texture feature). You have proposed two improvements. the first one is to use white gaussian noise to pollute the deep features, and it is useful. But there is not enough theory to explain why this improvement is effective. Please supplement the relevant theoretical content of this improvement measure.

 

Also, I noticed that some images are blurry, such as Figure 17, the bounding box of the detection result cannot be seen clearly. Please adjust them.

Author Response

Dear reviewer：

Thank you very much for your thorough and detailed review of our paper，we have revised the paper in response to your comments and suggestion.

For the detail of the first improvement, we decompose the whitened deep features by wavelet transform, and found that after filtering out high-frequency components, the texture and color feature are enhanced, which is contrary to our needs, so we propose an improvement to add noise pollution to the depth features.

Relevant analysis procedures and results have been added to the paper.

We also processed the unclear pictures according to your opinion.

The revised paper will be uploaded to the system.

Best regard.

Reviewer 3 Report

Minor revision

In this manuscript the authors propose a sonar image target detection method based on texture feature removal.

It is significant context for scientists.

The manuscript is well organized. Mathematically the paper is well supported.

Although, the authors have to do the following:

1)      A brief of the presentation of the paper in the last paragraph in introduction.

2)      It is better to put more information in the caption of the figures.

3)      They should not start with the number of the References, i.e in line 50, in line 51, in line 53 etc. They should write i.e. in [25].

4)      In lines 487 and 515, it is better to use Section 4.1.3 and Section 3.1 instead Chapter 4.1.3 and Chapter 3.1, respectively.

5)      In line 604, they should remove the Section 6. Patents

6)      The citation of the figures inside the text must have the same appearance i.e. lines 570, 575.

7)      The text needs grammatical corrections. i.e. line 423 starts with 9. Line 552, should write we instead of We.   

8)      The references should be in the Journal’s format.

Author Response

Dear reviewer：

Thank you very much for your thorough and detailed review of our paper，we have revised the paper in response to your comments and suggestions，including adding a brief description about the proposed method, change the usage style of references，fix the caption of the figures, corrected some misused words and modified the format of the reference in accordance with the requirements of the journal.

The revised paper will be uploaded to the system.

Best regard.