A Large-Crack Image-Stitching Method with Cracks as the Regions of Interest
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper proposed a method for stitching crack images of concrete surfaces through multiple steps, including training a segmentation model, feature extraction and keypoint matching. Addressing the following comments can improve the overall quality of the paper:
1. Introduction:
- In the first paragraph, it had better to introduce the concept of stitching images scientifically and its importance, rather than providing a numerical example.
- You may refer to some key review papers on the data-driven SHM and concrete damage as references in this section.
- The quantitative results provided in the last paragraph of the section seem unnecessary.
2. Materials and Methods
- The research process in Fig. 1 needs to be redrawn attractively and clearly.
- It is suggested to include a pseudo-code to help in the replicability of the method.
- How does the method address the problem of the distortions and changing the angle of the camera during imaging?
3. Results
- In Table 10, the values of the Ground Truth column need to be clarified.
4. Conclusions
- The disadvantages and drawbacks of the method need to be further clarified. Which images cause potential problems for the method?
Comments on the Quality of English Language
The paper needs comprehensive proofreading to handle the grammar, typo, citation and coherency issues. Here are some examples:
- The title of the paper has a typo (Cracks s!).
- The type and size of the font should be unified. For example, see the affiliation of the fourth author and the caption of Figure 8.
- In the abstract, which metric has been saved 98.6% and 58.7%?
- There should be a space between the last word and the citation brackets.
- In Figure 9's caption, there are double dots (matcher ..!).
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe research aims to perform an efficient image stitching by employing Mask R-CNN model that creates the ROIs with the purpose to reduce the feature points. After reading this manuscript, the reviewer comes up with two critical comments. One is that the presentation of methods is not clear. For example, in Tables 4-10, what are the traditional methods exactly? What are the differences of proposed method compared to “Matching using only ROI images” (Table 4)? Another is the CNN model used before image stitching. There are a lot of state-of-the-art image semantic segmentation models like DeepLabv3+, FCN. Why do you select Mask-RCNN? Other comments are also listed below, which suggests a modification of the present manuscript:
Please correct all the typos throughout the paper, such as “Cracks s Regions”; “the BF matcher.. (a)”.
Abstract can be polished by avoiding the repeated words like “However”; In addition, describe the abbreviations “SIFT keypoints”; “BF matcher”, “FLANN matcher”, “SSIM values”. It is difficult to understand the meaning of “saves 98.6% and 58.7%” for its first time to read. The content of the research output is not easy to be followed.
Avoid repeated keywords “image stitching”.
As to “This implies that for cracks exceeding 1 meter in length, dense imaging is required for detailed documentation.”, why not adjust the shooting distance to increase the scale of image?
As to “However, these approaches involve computation on entire images, …, requires significant matching computation time to find all matching point pairs.”, please mention the typical computation time required for the task of image stitching.
As to “detection results of CNN models can serve as Regions of Interest (ROI) for crack image stitching.”, is this a good logic for application. The CNN models may generate some errors in case of false positive results. This may cause seriously false information for image stitching.
In Figure 8(e) and (f), what is the difference of matching performance?
It is better to explain the underlying mechanism of “However, the performance of using only ROI image matching method is only better with the FLANN matcher, indicating its insufficient robustness.”.
It would be better to integrate all Figs. 11-14 to get an overview and whole comparison of stitching results.
Comments on the Quality of English LanguageExtensive editing of English language required.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper can be considered for publication.
Author Response
Thank you for your suggestion.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe revised paper shows some improvements and the comments have been addressed to some exent. But a few issues still remain: 1. The performance of image segmentation models of Mask-CNN, DeepLabv3+ and others are not compared to demosntrate that Mask-CNN is more powerful for crack segmentation. 2. Adjusting the shooting distance to increase the scale of the image and the resolurion of image depends upon the quality of camera. 3. Computation time required for the task of image stitching could be affected by the choice of devices (CPU or GPU). 4. False positive identification results can be even generated without stains and the produced masks usually have irregular patterns. 5. As to "while the matching result in Figure 8 (f) appears to be successful, in comparison with Figures 8 (c), 8 (d), and 8 (f), the first image still shows slight differences in the upper right corner of the image.", this is not clearly presented. Where is "Figures 8 (e)" mentioned in the sentence?
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have answered the questions and made a bit revisions of the manuscript. The revision can be accepted for publication in the journal.
Comments on the Quality of English LanguageMinor editing of English language required.
