PHCNet: Pyramid Hierarchical-Convolution-Based U-Net for Crack Detection with Mixed Global Attention Module and Edge Feature Extractor
Round 1
Reviewer 1 Report
In this paper, a crack detection model named Pyramid Hierarchical Convolution based U-Net (PHCNet) with MGAM, EFEM, and SAM is proposed. This topic research has the certain theory significance. The detailed comments are as follows. Please consider them for further revision.
1. Page 1, section 1, in the introduction section, more of the latest literature in the field of Crack detection needs to be cited.
2. Page 4, the clarity of Figure 2 needs to be improved.
3. Page 5, the parameter b of Equation 1 needs to be introduced.
4. Page 7, line 213, it is mentioned that "In ELEM, the input feature…", is this a misrepresentation? The module is referred to as "EFEM".
5. Page 6, check Figures 4 and 5 for correctness and make corrections to them.
6. Page 8, section 2.7, Incorporating references or experimental evidence to support "the Sigmoid activation function has some drawbacks. Firstly, it has the problem of gradient disappearance. Secondly, the computational complexity of the Sigmoid activation function is relatively high. "
7. Page 8, section 2.7, "Google proposed a new activation function named Swish" is missing the reference.
8. Page 10, Section 3.2, Lacking some important references, e.g., FCN, ConvNet, Split-Attention Network.
9. The authors should provide more ablation experiments to demonstrate the validity of each module mentioned in section 2, e.g., PHCM, EFEM.
10.There is only one evaluation indicator in this paper, and it is suggested to add the evaluation indicator in the segmentation field, e.g., MIOU.
In this paper, a crack detection model named Pyramid Hierarchical Convolution based U-Net (PHCNet) with MGAM, EFEM, and SAM is proposed. This topic research has the certain theory significance. The detailed comments are as follows. Please consider them for further revision.
1. Page 1, section 1, in the introduction section, more of the latest literature in the field of Crack detection needs to be cited.
2. Page 4, the clarity of Figure 2 needs to be improved.
3. Page 5, the parameter b of Equation 1 needs to be introduced.
4. Page 7, line 213, it is mentioned that "In ELEM, the input feature…", is this a misrepresentation? The module is referred to as "EFEM".
5. Page 6, check Figures 4 and 5 for correctness and make corrections to them.
6. Page 8, section 2.7, Incorporating references or experimental evidence to support "the Sigmoid activation function has some drawbacks. Firstly, it has the problem of gradient disappearance. Secondly, the computational complexity of the Sigmoid activation function is relatively high. "
7. Page 8, section 2.7, "Google proposed a new activation function named Swish" is missing the reference.
8. Page 10, Section 3.2, Lacking some important references, e.g., FCN, ConvNet, Split-Attention Network.
9. The authors should provide more ablation experiments to demonstrate the validity of each module mentioned in section 2, e.g., PHCM, EFEM.
10.There is only one evaluation indicator in this paper, and it is suggested to add the evaluation indicator in the segmentation field, e.g., MIOU.
Author Response
- Page 1, section 1, in the introduction section, more of the latest literature in the field of Crack detection needs to be cited.
Answer: I add the related works section and I cite many latest literature in the related works.
- Page 4, the clarity of Figure 2 needs to be improved.
Answer: I draw a new Figure and improve the clarity of the Figure.
- Page 5, the parameter b of Equation 1 needs to be introduced.
Answer: I have revised it in the paper.
- Page 7, line 213, it is mentioned that "In ELEM, the input feature…", is this a misrepresentation? The module is referred to as "EFEM".
Answer: I have revised it in the paper.
- Page 6, check Figures 4 and 5 for correctness and make corrections to them.
Answer: I have revised it in the paper.
- Page 8, section 2.7, Incorporating references or experimental evidence to support "the Sigmoid activation function has some drawbacks. Firstly, it has the problem of gradient disappearance. Secondly, the computational complexity of the Sigmoid activation function is relatively high. "
Answer: I have added a reference.
- Page 8, section 2.7, "Google proposed a new activation function named Swish" is missing the reference.
Answer: I have added a reference.
- Page 10, Section 3.2, Lacking some important references, e.g., FCN, ConvNet, Split-Attention Network.
Answer: I have added references.
- The authors should provide more ablation experiments to demonstrate the validity of each module mentioned in section 2, e.g., PHCM, EFEM.
Answer: I have revised it in the paper and I have added some experiments about the PHCM, and EFEM.
10.There is only one evaluation indicator in this paper, and it is suggested to add the evaluation indicator in the segmentation field, e.g., MIOU.
Answer: This is a very good suggestion. Sorry, due to the length limitation of the article, I am considering not conducting comparative experiments on the MIOU indicator for the time being. In the future, my research will use multiple indicators for comparison.
Reviewer 2 Report
The authors’ work seems very meaningful. However, the confusing structure makes it difficult to understand the author's work. Here below are some specific questions or suggestions:
1. Generally, the first section is an introduction. And at present, this section is too limited to show the development of this area. And the structure of the paper should be clear at the end of the introduction.
2. The logic of the structure should be: Introduction, Methods, Data Description, Results and Discussion (including figures and tables), Conclusions.
3. List of acronyms must be provided before introductions.
4. Please clearly highlight how your work advances the field from the present state of knowledge and you should provide a clear justification for your work. The impact or advancement of the work can also appear in the conclusion.
Author Response
- Generally, the first section is an introduction. And at present, this section is too limited to show the development of this area. And the structure of the paper should be clear at the end of the introduction.
Answer: I add the related works section and I cite many latest literature in the related works.
- The logic of the structure should be: Introduction, Methods, Data Description, Results and Discussion (including figures and tables), Conclusions.
Answer: I have revised it in the paper.
- List of acronyms must be provided before introductions.
Answer: I have revised it in the paper.
- Methods
To illustrate our proposed methods clearly, the list of abbreviations is shown as follows:
PCHNet: Pyramid Hierarchical Convolution based U-Net model with Mix Global Attention Module (MGAM), Edge Feature Extractor Module (EFEM) and Supplementary Attention Module (SAM)
PHCM: Pyramid Hierarchical Convolution Module
MGAM: Mix Global Attention Module
SAM: Supplementary Attention Module
EFEM: Edge Feature Extractor Module
- Please clearly highlight how your work advances the field from the present state of knowledge and you should provide a clear justification for your work. The impact or advancement of the work can also appear in the conclusion.
Answer: I have revised it in the conclusion of the paper.
This article improves the architecture of existing deep learning based crack detection models, solving the problems of multi-scale crack detection, insufficient refinement of crack edge feature extraction, and the influence of interference features in crack samples.
Reviewer 3 Report
The manuscript entitled “PHCNet: Pyramid Hierarchical Convolution based U-Net for Crack Detection with Mix Global Attention Module and Edge Feature Extractor” proposes Pyramid Hierarchical Convolution Module (PHCM) to extract features of crack with different size. Indeed, three modules Mix Global Attention Module (MGAM), Edge Feature Extractor Module (EFEM) and Supplementary Attention Module (SAM) are used to improve the precision of crack detection. The role of MGAM, EFEM and SAM are fusing global feature information, learning edge features of cracks and solving the interference of stains in crack images, respectively. The contribution of this work is almost enough and the organization of the manuscript is well so it is suitable to publish in the Applied Science Journal. Strong points of the manuscript can be concluded as:
1. The manuscript is well organized and well written.
2. Contribution of the manuscript is enough to publish in Applied Science journal.
3. The references are appropriate and in field of the paper subject.
4. Experimental results are reported in an appropriate manner.
5. The conclusion is consistent with the evidence and results reported in the manuscript.
6. Problem definition, solution, and methodology are presented very well.
However, some of the ideas those need to be thoroughly clarified before publication:
1. Lines 92-93, it is better that the sentence “Section 2 introduces the datesets we used.” is replaced with the “Section 2 introduce the materials and methods”.
2. Fig. 2 is vague. The color of layers is not in an appropriate manner.
3. It is strongly recommended to add a new section “Related works” after Introduction and discuss related papers in this section. More new papers should be discussed such as:
a. U-Net-Based CNN Architecture for Road Crack Segmentation
b. Two-stage convolutional neural network for road crack detection and segmentation
c. ECSNet: An Accelerated Real-Time Image Segmentation CNN Architecture for Pavement Crack Detection
4. Compare your results with state-of-the-art works such as “U-Net-Based CNN Architecture for Road Crack Segmentation”.
5. Some typos:
a. Lines 92, in the sentence “Section 2 introduces the datesets we used.”, “datesets” should be “datasets”.
b. Line 120, insert a space between “Fig.” and “1” in the sentence “Some sample images of the OAD_CRACK dataset are shown in Fig.1”.
Comments for author File: 
Comments.pdf
Author Response
- Lines 92-93, it is better that the sentence “Section 2 introduces the datesets we used.” is replaced with the “Section 2 introduce the materials and methods”.
Answer: I have revised it in the paper.
- 2 is vague. The color of layers is not in an appropriate manner.
Answer: I draw a new Figure and improve the clarity of the Figure.
- It is strongly recommended to add a new section “Related works” after Introduction and discuss related papers in this section. More new papers should be discussed such as:
- U-Net-Based CNN Architecture for Road Crack Segmentation
- Two-stage convolutional neural network for road crack detection and segmentation
- ECSNet: An Accelerated Real-Time Image Segmentation CNN Architecture for Pavement Crack Detection
Answer: I have revised it in the paper.
- Compare your results with state-of-the-art works such as “U-Net-Based CNN Architecture for Road Crack Segmentation”.
Answer: I have revised it in the paper in Table 1. Comparison with the state-of-the-art methods.
- Some typos:
- Lines 92, in the sentence “Section 2 introduces the datesets we used.”, “datesets” should be “datasets”.
- Line 120, insert a space between “Fig.” and “1” in the sentence “Some sample images of the OAD_CRACK dataset are shown in Fig.1”.
Answer: I have revised it in the paper.
Round 2
Reviewer 1 Report
In the new version of the paper, the author addressed all of the issues raised in my 1st review. So I advise to accept this paper.
In the new version of the paper, the author addressed all of the issues raised in my 1st review. So I advise to accept this paper.
Reviewer 3 Report
All comments have been addressed. ACCEPT
