Comparison of Residual Network and Other Classical Models for Classification of Interlayer Distresses in Pavement
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsIt's an important application of image mining. A workflow diagram of the whole experiment may help the reader. They may take imbalance into account from the very beginning, not in just in the discussion. The conclusions may be improved with the discussion issues and considering literature review. The authors may revise the numbers' format (lines 188-189 or others) for homogeneity.
Comments on the Quality of English LanguageJust minor revision is required.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsIn this work, the authors deal with the residual network evaluation of roadway structure properties using GPR images. However, the most important goal of the paper was not achieved, since the goal was a "fine-grained" classification, which was only touched upon in the paper, and at the end of the paper, in line 528, it was stated that "need further verification". I think that this state of the paper is not worthy of publication, let us wait for the development of a fine grained system.
However, the essay is logically disjointed, we do not see a coherent picture that is easy for the reader to follow.
An explanatory figure in the introduction would be necessary for the presentation of CNN.
In the case of background noise, we do not understand why there is nevertheless sometimes better prediction certainty than without it. Which part of the residual network provides this?
The system was recorded at 80 km/h, why is this important, is it not related to sampling?
The most important part of the first figure is the segmentation, it is not clear how it is realised. Here the propagation of uncertainties is also in question.
The second figure shows the various roadway phenomena without any explanation attached. Why do each recordings belong to a group in engineering terms?
On page 5, it is irrelevant exactly which of the thousands of labels is incorrect, if we do not use this information numerically any further.
The application of ResNet would have to be proven by comparing it to another network, it is not sufficient to test ResNet4-50 to the same network.
The low point of the paper is Figure 5, where nothing is readable, a curve of knots is vomited there without any deeper explanation.
In Table 5, the metrics are essentially the same. It's surprising that I can't distinguish between the data so well. The reviewer has been in the neural network business for over 30 years, but has never seen results like this.
Figure 7 is also illegible and unintelligible.
In summary, my suggestion is to revisit the results by more intensive processing, comparing with other nets and most importantly by uploading the code to github.
NA
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe article proposes and evaluates the performance of several models for the detection of pavement distress. The topic is important for field practitioners and authorities. The following issues need to be resolved before its acceptance.
The full forms of ResNET and Grad-CAM should be mentioned in the abstract and introduction. At least one statement should be added about these items in the abstract and some more details should be mentioned in the introduction.
Line 84: Can this observation be corroborated with some data or previous publications?
Line 147-162: References should be cited.
Figures 3, 5 and 8, Table 5: Titles are too long. The explanation can be incorporated in the text or as a footnote.
More explanation should be added about Figure 4.
Line 178-185: How was the labeling done here? How was the correctness of labeling determined?
Section 3.2, line 372 onwards: The explanation could be summarized and the accuracies of the models could be compared to other models. These models can be prepared in this study or found in the literature.
Conclusion: Future directions of research should be recommended at the end.
More references should be added from the last 5 years (2019 onwards).
Comments on the Quality of English LanguageThe English is generally good and understandable. The abstract could be improved in this regard.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors actually improved the paper.
Author Response
Thank you for accepting our work.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have obviously put a lot of work into improving their work and making it marketable, but the reviewer does not consider the results and their presentation to be publishable, as the questions have not been answered. The figures have been enhanced, but cannot be made more widely available, partly because they are illegible and partly because they are not interpretable.
I suggest starting another article on a smaller topic.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors have satisfied all my comments.
Author Response
Thank you for accepting our work.
Round 3
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have obviously put a lot of work into improving their Version 3 work and making it marketable, but the reviewer does not consider the results and their presentation to be publishable. The figures have been enhanced, but cannot be made more widely available, partly because they are illegible and partly because they are not interpretable.
After this third attempt, the reviewer has a growing feeling that the authors of the study should open a new, smaller project and start a new article.
Comments on the Quality of English LanguageNA
Author Response
Please see the attachment
Author Response File: Author Response.pdf