GPR B-Scan Image Denoising via Multi-Scale Convolutional Autoencoder with Data Augmentation

Round 1

Reviewer 1 Report

In this paper, the Authors propose a Multi-scale Convolutional AutoEncoder (MCAE) algorithm, based on deep learning techniques, to reduce the noise in Ground Penetrating Radar (GPR) images. A specific data augmentation strategy was designed to solve the problem of insufficient training dataset. The proposed method is described in detail and then compared with some of the other commonly employed image denoising methods, highlighting that the proposed method is generally more robust to various noise intensities than the other methods, and even more efficient in image processing.

This paper is very well written, the methods are properly presented and the comparison with the previous methods is fair. In my opinion, this paper is ready for the publication on Electronics, once the Authors have addressed the following minor issues.

 

1. Line 35: to help non specialist audience, please explain B-scan.
2. L 62: please explain acronyms the first time they are mentioned in the main text (BM3D).
3. L 115: please add a space between “…the network” and “[30,31]”.
4. L 118-119: in this sentence there are two verbs and a single subject, please check and rephrase.
5. L 135: outputs in place of output.
6. L 230: to help non specialist audience, please add a brief description of the ReLu activation.

Author Response

Dear reviewer,

Thanks for your review. Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear Authors

I revised with pleasure your manuscript “GPR B-scan image denoising via Multi-scale Convolutional Autoencoder with data augmentation” and I suggest the manuscript will be accepted after minor revisions.

I think of the manuscript is organized well, but the text can be difficult to follow for a geophysicist interested to GPR.

You propose a MCAE method that seems to work well with strong noise random data (see tab. 1 for PSNR and tab.2 for SSIM). I think it is necessary to comment because not in a low noise environment (BM3D is better for example). Is it a problem of algorithm? (increse discussion)

I hope your work can also be applied on GPR complex targets (for example archeological buried ruins). Is it possible to insert in your paper a GPR noisy image more complex than hyperboles?  A complex target would greatly enrich the paper. 

 In the Introduction, line 29 you have [1-4] references about GPR. Please insert an archaeological reference [5], GPR is successfully applied.  

 Line 69 Check reference 

Good work

Author Response

Dear reviewer,

Thanks for your review. Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript shows that GAN-based data augmentation can improve the denoising performance of multi-scale convolutional autoencoder (MCAE) in context of 
Ground penetrating radar (GPR) B-scan image denoising. The manuscript is very weak in following aspects:

1) The novelty of the proposed method is not sufficient. By now extensive works on GAN have shown that GAN-based augmentation can improve performance of almost anything.
There is no further novelty in the proposed method or how the GAN is used.

2) The experimental validation is insufficient. The manuscript did not even demonstrate whether GAN-based augmentation is superior to other types of augmentation. Moreover,
several types of GAN exist and they could be compared.

3) Manuscript claims in abstract that existing methods suffer when "image is severaly contaminated." However, there is no proof provided that the proposed method is a better
alternative to deal with "severely contaminated" images.

4) Figure 5, "with data augmentation" is always pretty much below "without data augmentation", even at the starting epochs. This does not seem quite right.

5) Insufficient references to existing works on denoising and GAN-based augmentation.

Author Response

Dear reviewer,

Thanks for your review. Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The manuscript is improved from previous version, however it still needs to improve/clarify some points:

1) Experimental validation: more methods should be added.There were two points in the previous comment:
Point 1: The manuscript did not even demonstrate
whether GAN-based augmentation is superior to other types of augmentation.
Point 2: Moreover, several types of GAN exist and they could be compared.
I did not see authors have acted upon point 1.
For points 2 also, at least 1 more type of GAN must be compared.

2) Authors say "Due to the limited space of the manuscript, we did not add this part to the manuscript, thank you again." - what is the reason of limited space in
a journal? My apologies if I am misunderstanding something!

3) GAN-based augmentations, referneces are still inadequate. Moreover those included are not quite related to the broad subject of this manuscript. Here
are some references that can be included:
i) Unsupervised multiple-change detection in VHR multisensor images via deep-learning based adaptation  (CycleGAN-based augmentation in Earth observation)
ii) Semantic-Fusion Gans for Semi-Supervised Satellite Image Classification  (ACGAN-based augmentation in Earth observation)
iii) Attention GANs: Unsupervised deep feature learning for aerial scene classification  (GAN in Earth observation)
iv) Generative modeling of spatio-temporal weather patterns with extreme event conditioning  (GAN in climate science)

4) In Algorithm 1, t (non-italic) and t (italic) are two different parameters? One is used as an iterator, other as summing t*... + (1-t)*...
Please make them visibly different.

5)Please include a brief description of "mean preprocessing method"

6) "For better display, the loss value is recorded starting from epoch of 10" - I do not see the point why loss value is really recorded starting from epoch 10. At most 1 epoch could have been excluded if authors are really concerned about undesired jumps at the beginning.

Author Response

Dear reviewer,

  We have responded to your comments. Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 3 Report

I am satisfied with the revised manuscript.