Defining Structural Cracks in Exterior Walls of Concrete Buildings Using an Unmanned Aerial Vehicle

Round 1

Reviewer 1 Report

In this paper, the crack detection method combining UAV photography and deep learning is used to classify the detection results. The technical route  is clear, and each step solves the corresponding problem.

The followings are revise suggestions:

1.     The structural cracks and non-structural cracks introduced in this paper only have text descriptions, the corresponding picture examples should be given.

2. The experimental results of 98.5% in this paper are only compared with the actual values, and are not compared with other similar methods. Comparison with other methods, especially the Maguire method cited in this article, should be made.

3.     It is recommended to explain in detail the causes of misclassification and missing of crack detection results.

 

 

 

Author Response

Reviewer #1

→ We thank the editor and reviewers for their detailed comments and suggestions. We have thoroughly revised and improved our manuscript. Our responses to the reviewer comments are appended below.

 

1. The structural cracks and non-structural cracks introduced in this paper only have text descriptions, the corresponding picture examples should be given.

→ Structural cracks and non-structural cracks have different forms depending on the structure of the building, so it was difficult to present a comprehensive picture example, so only have text descriptions. Thanks for the good comments.

 

2. The experimental results of 98.5% in this paper are only compared with the actual values, and are not compared with other similar methods. Comparison with other methods, especially the Maguire method cited in this article, should be made.

→ Maguire's research cited in this paper is a study that provides SDNET2018, a deep learning training dataset. Also, this study used a drone and deep learning to define structural cracks in the exterior walls of a concrete building. Since there is no study of a method similar to the method proposed in this study, comparison was not made. Thanks for the good comments.

 

3. It is recommended to explain in detail the causes of misclassification and missing of crack detection results.

→ In Chapter 3, We have additionally written about the contents. (In line 312-316). Thanks for the good comments.

Author Response File: Author Response.docx

Reviewer 2 Report

This study proposes a method for defining structural cracks in the exterior walls of concrete buildings. In summary, the research is exciting and provides valuable results, but the current document has several weaknesses that must be strengthened to obtain a documentary result that is equal to the value of the publication.

 

Chapter 1: Introduction

(1) What are your main contributions? What is your scientific novelty? Please clarify.

(2) The research gap is not well presented, and the total number of references is insufficient.

(3) The first paragraph may present a much broad view of the problems with citations to computer vision application references.

 

Chapter  2:Data Acquisition

(4) In lines 140-141, you have mentioned taking at a distance of 2m. Please add why you chose this distance to take photos. A hypothesis needs to be presented, When the weather is terrible,  e.g. in windy-to-rainy weather, how far away to ensure safety?

(5) In line 148, the reference is set to a few groups of image overlap data to construct spatial information. These data need to be briefly explained. And why 65% was chosen as the critical value?

(6) Why choose YOLO to detect cracks? In addition to introducing the principle of the improved algorithm, the paper should also fully explain the reasons for choosing the algorithm. For example, the advantages of this algorithm, its differences from traditional algorithms, and its disadvantages. 

(7) For crack detection, please refer to integrated generative adversarial networks and improved VGG model keyword-related articles, and mark-free vision method related ones.

Chapter 3: Experiment

(8) Overall, this section does not comprehensively describe the Deep Learning-based training results. Only figure 4 illustrates the results, which does not seem very convincing; please add data to illustrate.

Chapter 5: Conclusions

(9) The novelty of the study is not apparent enough. There is no experimental comparison of the algorithm with previously known work, so it is impossible to judge whether the algorithm is an improvement on previous work. 

Author Response

Reviewer#2

→ We thank the editor and reviewers for their detailed comments and suggestions. We have thoroughly revised and improved our manuscript. Our responses to the reviewer comments are appended below.

 

Chapter 1: Introduction

1. What are your main contributions? What is your scientific novelty? Please clarify.

→ Korea is incurring substantial social costs in maintaining deteriorating facilities. Thus, developing technologies for an economical and efficient building safety inspection is necessary. Therefore, this study has significance as a basic research on how to prioritize the risk of building damage for efficient maintenance and repair and reinforcement. In addition, studies related to crack detection in existing buildings focus on detection using UAV and deep learning. However, this study not only detects cracks in buildings, but also detects and classifies cracks in structural and non-structural parts of buildings using orthoimages and design drawings. Thanks for the good comments.

 

2. The research gap is not well presented, and the total number of references is insufficient.

→ Korea is incurring substantial social costs in maintaining deteriorating facilities. Thus, developing technologies for an economical and efficient building safety inspection is necessary. Therefore, this study has significance as a basic research on how to prioritize the risk of building damage for efficient maintenance and repair and reinforcement. In addition, studies related to crack detection in existing buildings focus on detection using UAV and deep learning. However, this study not only detects cracks in buildings, but also detects and classifies cracks in structural and non-structural parts of buildings using orthoimages and design drawings. Thanks for the good comments.

 

3. The first paragraph may present a much broad view of the problems with citations to computer vision application references.

→ The purpose of this study is not to develop deep learning algorithms. This is an applied study related to building damage and maintenance using previously developed algorithms. Therefore, in this study, it was judged that the content related to computer vision application reference was unnecessary in the introduction. Thanks for the good comments.

 

Chapter 2: Data Acquisition

4. In lines 140-141, you have mentioned taking at a distance of 2m. Please add why you chose this distance to take photos. A hypothesis needs to be presented, When the weather is terrible, e.g. in windy-to-rainy weather, how far away to ensure safety?

→ We reviewed previous studies for crack detection that used UAVs to photograph the outer walls of buildings, and analyzed the shooting distance of UAVs during aerial photography. In addition, in order to acquire the size of 0.3~0.4mm, which is the defect standard for cracks under the current Korean law, the shooting distance was set in consideration of the specifications and resolution of commercial drone cameras that are commonly used (1” CMOS, 20MP). Thanks for the good comments.

 

5. In line 148, the reference is set to a few groups of image overlap data to construct spatial information. These data need to be briefly explained. And why 65% was chosen as the critical value?

→ “a few groups of image overlap data to construct spatial information.” is judged to be a misunderstanding. In this study, spatial information was constructed using the point cloud technique. When constructing such spatial information, each image was set as one group, and overlap between images was set to 65%. This is the result of considering previous studies. Thanks for the good comments.

 

6. Why choose YOLO to detect cracks? In addition to introducing the principle of the improved algorithm, the paper should also fully explain the reasons for choosing the algorithm. For example, the advantages of this algorithm, its differences from traditional algorithms, and its disadvantages.

→ This study is not a study to evaluate the accuracy of deep learning algorithms. The purpose of this study is to define the location of cracks in structural and non-structural parts of buildings by applying deep learning. In the future, we are preparing a study to evaluate accuracy using various deep learning algorithms. Thanks for the good comments.

 

7. For crack detection, please refer to integrated generative adversarial networks and improved VGG model keyword-related articles, and mark-free vision method related ones.

→ In Chapter 1, we added references about to integrated generative adversarial networks and improved VGG model keyword-related articles, and mark-free vision method related ones. Also, this will be referred to in future research. Thanks for the good questions and references from the reviewers.

 

Chapter 3: Experiment

8. Overall, this section does not comprehensively describe the Deep Learning-based training results. Only figure 4 illustrates the results, which does not seem very convincing; please add data to illustrate.

→ In Figure 4, we revised figure and manuscript. Thanks for the good comments.

 

Chapter 5: Conclusions

9. The novelty of the study is not apparent enough. There is no experimental comparison of the algorithm with previously known work, so it is impossible to judge whether the algorithm is an improvement on previous work.

→ This is an applied study related to building damage and maintenance using previously developed algorithms. Also, the purpose of this study is not to improve the algorithm. UAV and deep learning algorithms are research tools for detecting and classifying structural and non-structural cracks in buildings. As mentioned in Chapter 4, this study differs from previous research in that it provides a method for judging the quality and risk level of cracks. Thanks for the good comments.

 

Author Response File: Author Response.docx

Reviewer 3 Report

I. My first issue is related to the construction of the dataset. Why the authors have selected only 12000 images from the SDNET2018 dataset, that contains initially over 56000 images?

II. The learning curves, that are presented in figure 4 have several issues :

II.1. First of all, the curves do not reach a stability performance plateau enabling to judge whether the used number of training epochs was a good choice or no.

II.2. Second, I can't tell from the curves about the underfitting/overfitting behavior of the model. I suggest adding the learning curves also for the validation dataset along the training one to enable a comparison between the two.

II.3. Third, the curves have to be resized and presented more clearly.

II.4. Fourth, the authors do not explain what is the box_loss , the obj_loss as well as the expression of the mAP. 

II.5. Why there are two mAP learning curves? and what is the difference between the two?

II.6. I cannot understand how the authors state a "rate" of 98.5%, while the precision and recall curves are at around 41%, a first mAP is around 35% and another one is around 10%.  

Author Response

Reviewer #3

→ We thank the editor and reviewers for their detailed comments and suggestions. We have thoroughly revised and improved our manuscript. Our responses to the reviewer comments are appended below.

 

1. I) My first issue is related to the construction of the dataset. Why the authors have selected only 12000 images from the SDNET2018 dataset, that contains initially over 56000 images?

→ Thanks for the good comments. Among the SDNET2018 dataset, 6,000 high-quality crack images and 6,000 high-quality images without cracks were selected to build learning data for crack detection on the outer wall of a concrete building.

 

1. ll) The learning curves, that are presented in figure 4 have several issues:

II.1. First of all, the curves do not reach a stability performance plateau enabling to judge whether the used number of training epochs was a good choice or no.

→ This study is not a study to evaluate the accuracy of deep learning algorithms. The purpose of this study is to define the location of cracks in structural and non-structural parts of buildings by applying deep learning. Therefore, we set up 150 epochs to train to a level where crack detection is possible. Thanks for the good comments.

 

II.2. Second, I can't tell from the curves about the underfitting/overfitting behavior of the model. I suggest adding the learning curves also for the validation dataset along the training one to enable a comparison between the two.

→ Thanks for the good comments. We confirmed the performance of the model through the results of mAP, precision, recall, etc. that appeared through training of the deep learning dataset. Also, this will be referred to in future research.

 

II.3. Third, the curves have to be resized and presented more clearly.

→ We revised the size and image quality of Fiugre 4. Thanks for the good comments.

 

II.4. Fourth, the authors do not explain what is the box_loss , the obj_loss as well as the expression of the mAP.

→ We revised the manuscript about mAP, box_loss, and obj_loss In Section 3.2. (In line 312-316). Thanks for the good comments.

 

II.5. Why there are two mAP learning curves? and what is the difference between the two?

→ mAP 0.5 is obtained by setting the average of mAP to the IOU threshold of 0.5, and mAP 0.5:0.95 is the average of mAP measured by changing the value between the IOU threshold of 0.5 and 0.95 by 0.5. Thanks for the good comments.

 

II.6. I cannot understand how the authors state a "rate" of 98.5%, while the precision and recall curves are at around 41%, a first mAP is around 35% and another one is around 10%.

→ “98.5%” is the crack identification rate. It is the comparative analysis results based on cracks detected by deep learning and measured location information. Thanks for the good comments.

 

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Overall, the provided comments are imprecise and vague. 

The dataset has been reduced without a clear justification.

The validation curves have to be provided along the training curves to diagnose the model behavior.

The presented numeric values in the curves are inconsistent with the claimed results.

 

Author Response

Reviewer#3

Overall, the provided comments are imprecise and vague. 

→ We thank the reviewers for their detailed comments and suggestions. We have thoroughly revised and improved our manuscript. Our responses to the reviewer comments are appended below.

 

The dataset has been reduced without a clear justification.

→ The selection of only 12,000 images from the SDNET2018 dataset was made for practical reasons such as reducing processing time or limited computational resources. As a result, a deep learning model was built using a training data set of 12,500 images and was used to complete the crack detection. - We revised the manuscript about this context In Section 3.2.

The purpose of this study is to define structural and non-structural cracks by applying a layer-merging technique using various layers (orthoimage layer, structural design drawing layer, deep learning layer) to the existing crack detection technology using drones and deep learning.

This study is not a paper on the development of deep learning for crack detection. In other words, the accuracy of deep learning and the accuracy of the methodology proposed in this study are unrelated. Therefore, considering the resource limitations of the computer and efficient processing time, we constructed the training data set with 12,500 images.

 

The validation curves have to be provided along the training curves to diagnose the model behavior.

→ We understand the importance of including validation curves along with the training curves to diagnose the behavior of the model. We revised the size and image of Figure 4. In addition, we added the validation curves. - We revised the manuscript about this context In Figure 4. We are also considering deleting Figure 4, which may cause confusion to reviewers and future readers, since this paper is not aimed at deep learning development.

 

The presented numeric values in the curves are inconsistent with the claimed results.

→ As we revised the paper once again, we understood the reviewer's misunderstanding. This paper is not about a deep learning development study. The purpose of this study is to provide a methodology for defining structural and non-structural cracks by utilizing technologies such as location information of UAV images, point cloud technology, structural analysis of design drawings, and merging layers (design drawings, orthoimages, and deep learning-based crack detection images). The cracks defined by the methodology proposed in this study were compared with the on-site measured cracks, and it was confirmed that they all matched. We have revised the title, abstract, introduction, section 3.2, section 3.3, discussion, and conclusions because the contents of the paper may cause misunderstanding.

Author Response File: Author Response.docx