Automated Crack Detection in Micro-CT Scanning for Fiber-Reinforced Concrete Using Super-Resolution and Deep Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper focuses on the micro-crack detection of fiber-reinforced concrete, utilizing Micro-CT scanning combined with super-resolution and deep learning technologies to achieve automated and high-precision crack detection. The study trains and evaluates the performance of the DETR model, and highlights the advantages of the proposed method by comparing it with existing studies.

1. The abstract mentions "committee-based post-processing" but fails to briefly explain its core logic (e.g., multi-scale verification rules), making it difficult for readers to quickly understand the value of this post-processing step.
2. In the literature review section, the description of limitations of studies like Li et al. [25] and Tian et al. [26] is too vague. It does not specifically analyze their shortcomings in "adaptability to fiber-reinforced concrete," "dataset openness," and "adaptability to detection tasks.
3. Missing details and logical supplements in the method section. Section 2.1.1 mentions that the dataset includes "XZ plane slices of 10 volumes" but does not specify the number of specimens corresponding to each volume, the proportion of fiber types.
4. The crack detection experiment in Section 3.2 does not explain the "data augmentation strategies" used during training, it is can be improved to determine whether the model's generalization ability is limited by insufficient data diversity.
5. The discussion section does not analyze the model's limitations. For example, the paper mentions that "the model has a high missed detection rate for small cracks" but does not quantify the specific scale range of "small cracks" or discuss optimization directions.
6. Some tables in the paper have non-standard formats: in Tables 5 and 6, "Yolov10" is inconsistent with the previous expression "Yolo (v8 and v11)," indicating a typo.
7. The following literatures related to this topic are recommended：

• Xiang Z, He X, Zou Y, et al. An active learning method for crack detection based on subset searching and weighted sampling[J]. Structural Health Monitoring, 2023: 14759217231183661.
• Xu Y, Wei S, Bao Y, et al. Automatic seismic damage identification of reinforced concrete columns from images by a region-based deep convolutional neural network[J]. Structural Control and Health Monitoring, 2019, 26(3): e2313.

Author Response

Dear reviewer,

We greatly appreciate your comments on our work and the experience you have shared, which have contributed to the enrichment of the paper.

Each comment will be discussed separately below.

Comments 1: The abstract mentions "committee-based post-processing" but fails to briefly explain its core logic (e.g., multi-scale verification rules), making it difficult for readers to quickly understand the value of this post-processing step.

Response 1: We agree that the Introduction is vague. We have added more explanatory details about the method's stages and provided a more detailed explanation of the post-processing stage, which involves voting by a committee of models trained at different resolutions. More details can be found in subsection 2.5.

Comments 2: In the literature review section, the description of limitations of studies like Li et al. [25] and Tian et al. [26] is too vague. It does not specifically analyze their shortcomings in "adaptability to fiber-reinforced concrete," "dataset openness," and "adaptability to detection tasks.

Response 2: We discussed the adaptability of the methods presented in the literature in the Introduction. We added more details about this in lines 106-133 (highlighted in red). We hope this clarifies the issue.

Comments 3: Missing details and logical supplements in the method section. Section 2.1.1 mentions that the dataset includes "XZ plane slices of 10 volumes" but does not specify the number of specimens corresponding to each volume, the proportion of fiber types.

Response 3: More details about the distribution of matrix types and fibers present in the base were added in subsection 2.1.1.

Comments 4: The crack detection experiment in Section 3.2 does not explain the "data augmentation strategies" used during training, it is can be improved to determine whether the model's generalization ability is limited by insufficient data diversity.

Response 4:  We performed an online data augmentation during training. The parameters were added in Table 5.

Comments 5: The discussion section does not analyze the model's limitations. For example, the paper mentions that "the model has a high missed detection rate for small cracks" but does not quantify the specific scale range of "small cracks" or discuss optimization directions.

Response 5: According to our expert, the minor cracks are about 5 µm (microns) in size. We have added more information to the discussion in lines 382-393.

Comments 6: Some tables in the paper have non-standard formats: in Tables 5 and 6, "Yolov10" is inconsistent with the previous expression "Yolo (v8 and v11)," indicating a typo.

Response 6: You are correct. It was indeed a typo. The version of YOLO used in our work is v10. We have corrected this and standardized all tables in terms of style and captions.

Comments 7: The following literatures related to this topic are recommended：

Response 7: We appreciate the references to other studies. Because they use macro-scale images, they do not bring much to the discussion of the studies in the Introduction. However, we have added the references and provided more details on the types of techniques used in the studies, in lines 106-112.

We hope that these changes address the comments made, and thank you for your contributions.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors,

Paper is well presented but it has some limitations, which should be addressed. I am listing them below. 

1.) The introductory part is rather slim. A lot of work has happened in area of crack detection using AI methods in past few years. Particularly in area of concrete cracks. It should be addressed in the introduction part. Additional referenced should be added. 

2.) Assumptions are not clearly stated. 

3.) Deep learning network used in the paper is only briefly mentioned, where as it plays key role. No discussion about network training parameters and no metrics is shown. 

4.) Figures are poor quality see for example figure  6 to 8, some of the image aspects are not very clear. Please use image enhancement to understand it better. 

Thanks 

 

Author Response

Dear reviewer,

We greatly appreciate your comments on our work and the experience you have shared, which have contributed to the enrichment of the paper.

Each comment will be discussed separately below.

Comments 1: The introductory part is rather slim. A lot of work has happened in area of crack detection using AI methods in past few years. Particularly in area of concrete cracks. It should be addressed in the introduction part. Additional referenced should be added.

Response 1:  We agree that the Introduction is slim. We have added more details in the abstract. 

Comments 2:  Assumptions are not clearly stated. 

Response 2: We have revised the article and explained some points that needed further information. We hope this has clarified the issue.

Comments 3: Deep learning network used in the paper is only briefly mentioned, where as it plays key role. No discussion about network training parameters and no metrics is shown. 

Response 3: We detail the training parameters in subsections 3.1 (for classification) and 3.2 (for detection). The metrics are presented in Tables 2 and 4 (for classification) and 6-10 (for detection).

Comments 4: Figures are poor quality see for example figure  6 to 8, some of the image aspects are not very clear. Please use image enhancement to understand it better. 

Response 4: We agree that the image quality was not suitable. We generated the images again at a higher resolution and ensured that they all have a resolution of 300 dpi.

We hope that these changes address the comments made, and thank you for your contributions.

Reviewer 3 Report

Comments and Suggestions for Authors

Summary:
Fiber-reinforced concrete is a crucial material for civil construction, and monitoring its health is important for preserving structures and preventing accidents and financial losses. Among non-destructive monitoring methods, Micro-CT imaging stands out as an inexpensive method that is free from noise and external interference. However, manual inspection of these images is subjective and requires significant human effort. Therefore, this work proposes a framework for automatic crack detection that combines super-resolution-based preprocessing, Detection Transformers (DETR), and committee-based post-processing to reduce false positives. The model was trained on a new publicly available dataset, the FIRECON dataset, which consists of 4,064 images annotated by an expert, achieving metrics of 86.098% Intersection over Union, 89.37% Precision, 83.26% Recall, 84.99% F1-Score, and 44.69% Average Precision. The framework, therefore, significantly reduces analysis time and improves consistency compared to the manual methods used in previous studies. The results demonstrate the potential of Deep Learning to aid image analysis in damage assessments, providing valuable insights into the damage mechanisms of fiber-reinforced concrete and contributing to the development of durable, high-performance engineering materials.

Comments (minor):
1. Equations should be corrected (319,320 line).
2. Much clearer images are in need. (Figure 2)
3. Please explain what is the meaning of different color-bounding box in Figure 7. One is with confidence, oneis not.

Author Response

Dear reviewer,

We greatly appreciate your comments on our work and the experience you have shared, which have contributed to the enrichment of the paper.

Each comment will be discussed separately below.

 

Comments 1:  Equations should be corrected (319,320 line).

Response 1: We have reviewed all the equations. The comma character at the end of each line is confusing. We have removed it. Please let us know if anything is still wrong.

Comments 2: Much clearer images are in need. (Figure 2)

Response 2: We agree that the image quality was not suitable. We generated the images again at a higher resolution and ensured that they all have a resolution of 300 dpi.

Comments 3: Please explain what is the meaning of different color-bounding box in Figure 7. One is with confidence, oneis not.

Response 3: Figures 6 and 7 show the comparison between the model predictions (in red, with confidence) and the expert annotations (in green). Since the expert annotations represent the gold standard, there is no need for confidence. We explain this in the captions, but we also include it in the body of the text for added clarity.

 

We hope that these changes address the comments made, and thank you for your contributions.