Missing Data Recovery via Deep Networks for Limited Ground Penetrating Radar Measurements
Round 1
Reviewer 1 Report
This paper presents two methods to extract the prior distribution information in GPR data based on learning and learning-free process respectively, afterwards successfully applied to the recovery of GPR images with missing data. In general, this paper is well organized and written, and the proposed method has better performance than the traditional recovery algorithms. However, the following points still need to be considered and explained:
- PSNR is used to evaluate the quality of image restoration which is based on comparing the entire images. However, in GPR images, the hyperbolic signature actually occupies only a small part. Obviously, a relatively high PSNR value will also be obtained when the background data recover well, with the hyperbolic signature has not been accurately recovered, therefore, it is necessary to carry out targeted evaluation on the hyperbolic signature.
- In the original paper of DIP method, it is clearly pointed out that adding skip connections to deep learning network will damage the effect of image data reconstruction. Whether the skip block in the SkipNet used in this paper also have an adverse impact on the GPR data recovery?
- In addition, the depth of the network also plays an important role in the quality of data recovery when using DIP method. How the author determines the network with four layers also needs to be explained.
- During GPR detection, due to noise, medium and other factors, the form of data fading in the measured GPR image is not only pixel-wise missing or column-wise missing in most cases. Have the authors tried to repair the data missing image caused by more common reasons such as noise and shielding effect of metal impurities?
- The hyper-parameters, optimizer and the level of weight parameters update in the network all have important impacts on the image recovery, but authors fail to show the PEN-Net training process and training curves in detail, at the same time, the particular process of parameter update is also not clearly displayed when using DIP algorithm.
- Why use Patch GAN as a discriminator? Obviously, it is unreasonable to use the Generative Adversarial Network as the discriminator in this data recovery algorithm.
Author Response
The comments are attached.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
1) Abstract – “This study proposes 2 methods based on deep networks for the missing data recovery” -> “This study proposes two methods based on deep networks for the missing data recovery”.
2) Abstract – “The first method uses pyramid-context encoder network (PEN-Net) architecture which consists of 3 parts” -> “The first method uses pyramid-context encoder network (PEN-Net) architecture which consists of three parts”.
3) The related work discussion should be complemented with a table categorizing the reviewed works along their main distinctive characteristics so as to better position the present study.
4) When discussing the missing data problem in GPR, I would like the authors also to better highlight its importance (and impact) on imaging algorithms:
"Time reversal imaging of obscured targets from multistatic data." IEEE Transactions on Antennas and Propagation 53.5 (2005): 1600-1610.
“Performance analysis of time-reversal MUSIC." IEEE Transactions on Signal Processing 63.10 (2015): 2650-2662.
5) Please add a notation paragraph at the end of Sec. I.
6) For the sake of improved readability, it would be very useful adding a system model figure which depicts the considered missing data problem in GPR applications at a bird’s eye view.
7) By looking at the formulation in Sec. 2, it is not completely clear to me whether there is a rank assumption (i.e. M is known to some extent) in the proposed methods.
8) At the end of Sec. 3, it would be very useful if the authors could add a summarizing qualitative comparison from the complexity viewpoint of the two proposed approaches against UNet.
9) Sec. 4.1 – Please collect all the parameters concurring to describe the considered simulation setup in a given table for readers’ convenience.
10) I would recommend the authors relating pixel-wise and column-wise missing data scenarios to practical setups.
11) It is not clear to me whether the experimental data used in the second part of the numerical assessment are (will be) made publicly available by the authors. In the latter case, this would greatly foster reproducibility and further advances on the topic.
12) At the end of conclusions section, please add a paragraph describing future research directions foreseen. One interesting extension may be the use of XAI techniques to interpret and improve the considered two methods, following:
"Peeking inside the black-box: a survey on explainable artificial intelligence (XAI)." IEEE access 6 (2018): 52138-52160.
"XAI meets mobile traffic classification: Understanding and improving multimodal deep learning architectures." IEEE Transactions on Network and Service Management (2021).
Author Response
The comments are attached.
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors proposed a method for interpolating the missing data and the obtained accuracies of the proposed method were high. In a general way, the manuscript is interesting, well-written and easy to read. Experiments are clear and report on good quality work that has been well executed.
I list here below some comments to improve the presentation, but in general I would say that the manuscript requires minor revisions only.
Figures
Pleased add the color ramp legend.
LL.210-212
Could you provide the reasons why you chose SkipNet architecture? What kind of features were effective?
4. Experimental Results
Could you also evaluate the qualities based on the Structural Similarity (SSIM)?
Were the experimental data 8-bit color graphics?
Please provide more details on the datasets.
Author Response
The comments are attached.
Author Response File: Author Response.pdf
Reviewer 4 Report
Missing Data Recovery via Deep Networks for Limited Ground Penetrating Radar Measurements
A brief summary
The paper proposes 2 methods based on deep networks for the missing data recovery. The first method uses pyramid-context encoder network (PEN-Net) architecture. The method needs training, and it requires considerably less data compared to the existing U-Net based method. The second method, named deep image prior (DIP), is a regularization based data recovery method which uses an untrained network as a prior.
Comments
Strengths of the paper:
- Professionally laid out paper – both academic and practical, comprehensive, clear, and solid in terms of descriptive content and proposed considerations/applications.
- Extensive review references (45) on the considered topic and a clear review of published research works presented.
- Well thought out the methodology by using different approaches and methods.
- The proposed methods seem to be innovative and contains well-known hints of originality.
- The results show that proposed methods have noticeably better performance than conventional methods for the pixel-wise and moderate level column-wise missing data case. Besides, they can also deal with high rate column-wise missing data cases where the conventional methods seem completely to fail.
Weakness of the paper:
- The authors should clarify if the proposed methods work well only with the commercial GPR antenna (SIR-3000 with 1.5-GHz antennas, GSSI) or can use also data measured with other devices.
- The paper should be revised paying attention to the position of the figures.
- The most significant numerical results should also be reported in the abstract and in the conclusions.
- It is preferable to use an impersonal style (i.e., avoiding I and we).
- Abbreviations must be defined when first used and the abstract should not contain any references.
- … no more weaknesses!
The overall merit of presented research works and findings is high and definitely worth publishing after incorporation the above minor suggestions.
Author Response
The comments are attached.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
All questions have been properly responded.
Author Response
We appreciate for the time and effort that you have put into reviewing the previous version of the manuscript.
Reviewer 2 Report
The authors have improve their manuscript. Still, there are some pending comments that have been left unaddressed:
1) When discussing the missing data problem in GPR, I would like the authors also to better highlight its importance (and impact) on imaging algorithms:
"Time reversal imaging of obscured targets from multistatic data." IEEE Transactions on Antennas and Propagation 53.5 (2005): 1600-1610.
“Performance analysis of time-reversal MUSIC." IEEE Transactions on Signal Processing 63.10 (2015): 2650-2662.
Note: The authors have claimed to having discussed these references but I could not find the related discussion in the revised manuscript.
2) At the end of Sec. 3, it would be very useful if the authors could add a summarizing qualitative comparison from the complexity viewpoint of the two proposed approaches against UNet. One simple option may be comparing their training runtimes.
Author Response
We appreciate the reviewers for the time and effort that they have put into reviewing the previous version of the manuscript.
Author Response File: Author Response.pdf
