Detecting Underwater Concrete Cracks with Machine Learning: A Clear Vision of a Murky Problem

Round 1

Reviewer 1 Report

The authors present the development of an underwater crack detection system, and machine learning based approaches are used. The existing networks performance are improved, and the effectiveness and accuracy of the proposed approach in detecting cracks in underwater concrete structures is demonstrated. The structure is well-designed and the writing is good.

Author Response

Reviewer 1 -

 

 The authors present the development of an underwater crack detection system, and machine learning based approaches are used. The existing networks performance are improved, and the effectiveness and accuracy of the proposed approach in detecting cracks in underwater concrete structures is demonstrated. The structure is well-designed and the writing is good.

 

Thanks for the reviewer comments.

Reviewer 2 Report

This paper presents the development of an underwater crack detection system for structural integrity assessment of submerged structures. But only visible cracks are suitable for this method.

Comments:

1. For Fig.7, the overall accuracy is wrong. And the correct average should be calculated.

2. In Fig.8, 959 of 6000 negative examples are predicted as positive. The reason should be explained in detailed. 

Minor editing of English language required.

Author Response

Reviewer 2:

 

 This paper presents the development of an underwater crack detection system for structural integrity assessment of submerged structures. But only visible cracks are suitable for this method.

 

 Comments:

 

1. For Fig.7, the overall accuracy is wrong. And the correct average should be calculated.

 

2. In Fig.8, 959 of 6000 negative examples are predicted as positive. The reason should be explained in detailed.

 

Thanks for the reviewer comments. The figures and the text has been revised as suggested.

Author Response File: Author Response.docx

Reviewer 3 Report

A detailed literature review of the state of the current methods for underwater surface crack detection is presented. An overview of the image augmentation approach for the creation of underwater optical effects is also presented. Experimental results using a standard network based on a machine learning approach, used for surface crack detection in an onshore environment, are presented. The research process is completed, but not innovative enough. The authors need to add the following content.

1)       What are the innovative points of this paper? What does it do well compared to existing studies?

2)       Is it possible to combine Section 3 and 4 in the writing process?

3)       Is there a lot of knowledge about optics in Section 5 that is relevant to image enhancement?

4)       What is in Section 5 that the authors propose? Regarding the methods of image enhancement, the authors just introduced and used methods proposed by others.

5)       Detailed experimental results should be given in Section 6.

6)       How the model in Section 6 was determined, multiple sets of comparison tests with other models should be added.

7)       How is the system in the title represented in the text?

 

8)       The format of the references should be rechecked and standardized.

There are still problems with improper language presentation in the text.

Author Response

 Reviewer 3:

 

 A detailed literature review of the state of the current methods for underwater surface crack detection is presented. An overview of the image augmentation approach for the creation of underwater optical effects is also presented. Experimental results using a standard network based on a machine learning approach, used for surface crack detection in an onshore environment, are presented. The research process is completed, but not innovative enough. The authors need to add the following content.

 

Thanks for the reviewers comment please see our response below:

 

• What are the innovative points of this paper? What does it do well compared to existing studies?
  1. Innovativeness is emphasized by the augmentation of the dataset of existing images - shallow water optical effects. Simulation of shallow water optical effects and integration with existing database when the network is able to deal with the images - this is innovative.

 

• Is it possible to combine Section 3 and 4 in the writing process?
  1. Thank you for noticing. Sections 3 and 4 are short, it's really worth combining them into one.

 

• Is there a lot of knowledge about optics in Section 5 that is relevant to image enhancement?
  1. Yes, of course. Without this knowledge, we would not be able to perform a realistic simulation of underwater optical effects. All illustrations, all formulas are particularly important because we rely on the existing state of the art in the field of underwater optical effects simulation. It is important for us to value the wave spectrum, wave height, and wind speed, these are very important parameters for modeling realistic underwater optical effects. Without proper realistic modeling this image enhancement would be worthless.

 

• What is in Section 5 that the authors propose? Regarding the methods of image enhancement, the authors just introduced and used methods proposed by others.
  1. We would like to emphasize that we use validated, underwater optical effects simulation results. In this place, we clarify the necessary knowledge, and clearly demonstrate how that knowledge is used for modeling effects. The novelty and our contribution is that with the help of these optical effects we significantly expand the training dataset, which allows the Mendeley database to expand from 40000 images to 40000 more images with optical illumination effects. Such a realistic expansion of the database has lasting value and is also one of the important results of the paper.

 

• Detailed experimental results should be given in Section 6.
  1. Additional plots have been added to provide details of the analysis.

 

• How the model in Section 6 was determined, multiple sets of comparison tests with other models should be added.
  1. An explanation on how the model for deep learning has been chosen is added to the text.

 

• How is the system in the title represented in the text?
  1. The system or deep learning method proposed in the paper addresses the challenges related to concrete crack detection in underwater images. The text explains how it is achieved with help of an image augmented database and use of deep learning methods.
  2. This is a direct representation of the flickering optical effects in shallow water changing the quality of the trained CNN; we demonstrate how to bring that quality back to the highest standards. This is the main purpose of this article.

 

 

 

• The format of the references should be rechecked and standardized.
  1.  done

 

 

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

1)       It is necessary to describe how to select the two networks based on the results in Figure 2.

2)       Please explain Figure 7 and Figure 8.

 

3)       Why are 30% of the data set used for the training set, validation set, and test set? Were they selected randomly?

Check the language thoroughly.

Author Response

Dear Reviewer,

Thanks for your comments please see our response below:

It is necessary to describe how to select the two networks based on the results in Figure 2.

• An explanation on choice of network has been added in text. We select AlexNet and Squeezenet as two deep learning networks for prototyping. As these networks are quick to implement. We look at accuracy vs. simulation time that's' how we decided to go with Alexnet and Squeezenet.
• Also since the aim of this work is not network designing, which is a broad topic we have decided to use existing networks for this work. 

  Please explain Figure 7 and Figure 8.

• An explanation for the figure 7 and 8 has been now added. 

  Why are 30% of the data set used for the training set, validation set, and test set? Were they selected randomly?

• 30% of the data is used for training, then 30% is used for testing and 30% of data is used for validation. So total 90% of the data is used. And yes the data is selected randomly.