An Optimized Intelligent Segmentation Algorithm for Concrete Cracks Based on Transformer

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors,

Thank you for the manuscript which is very interesting.

This work describes novel improvements to the automated crack detection technology in concrete utilizing image processing, to face many challenges encountered in   practical applications. A new algorithm is presented for which an enhanced crack recognition is claimed under a variety of environments with many advantages in precision, recall, and robustness over existing algorithms.

In general, this manuscript is of very good quality. It is easy to read and well structured.  There are some remarks concerning minor mistakes, and proposed impacts on future developments.

These remarks are:

1.- Line 81. It’s saying “parallelly process” should be better “parallel processing”?

2.- Lines 296-297. Where were these datasets taken from?  A pipeline concrete?

3.- Fig. 1. Mispealled word “Segmatataion”
4.- Reference 1 was not found in Google search.
5.- Can you briefly described the blocks (RULE, BN, etc) for convolution modules of Fig. 2?
6.- There is no mention of preprocessing images for quality and consistency. Is this feature embeded in the flowwork of Fig. 1?
7.- Is there a possibility to estimate the crack depth using ResNet101?. Otherwise, can the manually measured crack depth be used to further categorize cracks?
8.- Could it be possible to implement in  Resnet the capability of detecting the probabilistic localization of hidden cracks in concrete structures?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. Title: The proposed method in the manuscript seems to mainly use CNN as a framework and only a small part of the Transformer module. Is the title appropriate?

 

2. Abstract: The abstract section should contain some specific findings, such as algorithmic accuracy.

 

3. Related work: Section 2.1 is titled Image Segmentation, but the authors state a great deal about object detection models. It is recommended that the title of Section 2.1 be changed.

 

4. Methodology: In the manuscript, the author's description of the method in Image 3 is very poor, as is the definition of RA and PDD.

 

5. Experiments and Results: The model presented in the manuscript seems to be similar in name to transUNet, what is their difference, and is there any citation in the relevant literature for transUNet?

 

6. Experiments and Results: Is there a rational explanation for the fact that FCN, a pioneering method for semantic segmentation, has achieved a performance that is more than second only to transFissNet?

 

7. Experiments and Results: The dataset used for comparative model performance tests in the manuscript is poorly described.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper presents transFissNet, a deep learning model for concrete crack segmentation, demonstrating clear innovation and practical relevance for structural health monitoring. The methodology and experiments are well-structured. One of my concerns is that only 437 raw images (augmented to 2,185) may insufficiently capture real-world diversity.

1. Abstract: The abstract lacks quantifiable results. It would be better to add metrics to highlight contributions.
2. Related work: Comparisons between cited methods are qualitative. It would be better to add a table summarizing strengths/weaknesses.
3. Line 230: ResNet101 is chosen as the backbone, it would be better to detail the reason.
4. Section 3.2: The Transformer module’s architecture is unspecified. It would be better to provide more details.
5. Section 3.2: The "intelligent repair" component (mentioned in contributions) is absent from the methodology. It would be better to describe how segmentation outputs drive repair decisions.
6. Line 296: The concrete crack dataset (437 images) seems small.
7. Line 317: For Crack500, splitting images into 256×256 patches may lose context. It would be better to explain how patch size affects performance.
8. Line 349: Authors mentioned the input image resolution size is 204x204, however in line 317, the segmented images is 256x256. It would be better to justify the inconsistency.
9. Line 350: It would be better to give more details about the loss function.
10. It would be better to double check “Ai, BI, SR, SO”.
11. Section 4.6: The crack identification system’s hardware is described, but its integration with transFissNet is unclear. It would be better to provide a workflow diagram. 
12. Section 4.6: The crack identification system is interesting. I hope authors can provide more details, such as the decision-making frameworks.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Okay