CPAM: Cross Patch Attention Module for Complex Texture Tile Block Defect Detection
, Quan Wang 1, Lufeng Luo 1, Yunzhi Zhang 1, Qinghua Lu 1, Wei-Chang Yeh 2 and Jiancheng Liang 1
Round 1
Reviewer 1 Report
The authors propose a new attention mechanism called the Cross Patch Attention Module to tile block defect detection.
Questions:
1. Regarding the method proposed by the authors and described in Figure 1, what are the innovations/contributions in relation to existing methods? Because the use of functions "pool", "split" and "conv" already existing in the literature.
2. In Section 3.4.1, the authors conclude that the YOLOv7 model is the best to apply the proposed method. Thus, it would be interesting to describe this model pointing out the differences in relation to the analyzed models. Will we always be able to apply the YOLOv7 model to any of the problems? Or are there requirements to use the YOLOv7 model?
3. The authors do not present the values of the parameters of the proposed method. It is important to present. In all case studies, were the parameters always the same? In addition, a sensitivity analysis of the parameters of the proposed method would be of interest to assess how its performance changes.
Author Response
Response to Reviewer 1 Comments
Dear Editor and Reviewer:
Thank you for your decision and constructive comments on my manuscript. We have carefully considered the suggestion of Reviewer and make some changes. We have tried our best to improve and made some changes in the manuscript. The following are our detailed responses to the issues raised by the reviewer.
Point 1: Regarding the method proposed by the authors and described in Figure 1, what are the innovations/contributions in relation to existing methods? Because the use of functions “pool”, “split” and “conv” already existing in the literature.
Response 1: Compared with the existing methods, we do not use new functions in the proposed method, but propose an attention mechanism for extracting patch information in view of the shortcomings of the existing methods in extracting feature information. The biggest difference between the proposed method and the existing method is the structure and the processing method of the extracted feature information.
Point 2: In Section 3.4.1, the authors conclude that the YOLOv7 model is the best to apply the proposed method. Thus, it would be interesting to describe this model pointing out the differences in relation to the analyzed models. Will we always be able to apply the YOLOv7 model to any of the problems? Or are there requirements to use the YOLOv7 model?
Response 2: Thank you for this very valuable question. In Section 3.4.1, we added some analysis of how YOLOv7 differs from other models. In addition, the YOLOv7 model can not always be applied to any problem. In this paper, YOLOv7 has obvious advantages for the detection of partial tile block defects.
Point 3: The authors do not present the values of the parameters of the proposed method. It is important to present. In all case studies, were the parameters always the same? In addition, a sensitivity analysis of the parameters of the proposed method would be of interest to assess how its performance changes.
Response 3: Thank you very much for your advice. In Section 3.2, we have provided our experimental environment and specific parameter values, and all experiments are conducted in the same environment. In addition, we analyzed the impact of different parameters on performance in Section 3.4.2.
Reviewer 2 Report
The submitted manuscript on “CPAM: Cross patch attention module for complex texture tile block defect detection” has presented a good work. However, there are few shortcomings as listed below that need to be revised by the author(s) before any further recommendation –
1. What is the main question addressed by the research? This needs to be included in the Introduction section. There are so many objectives claimed; however, same should also be specifically drawn herein.
2. The literature presented in introduction section need to be presented in the Tabular form as it is difficult to follow. Particularly for other defects are required to be outlined clearly.
3. Quality of the Heat maps needs to be improved.
4. The discussion section 4 is too small. Literature support to this section needs to be inculcated with an overall improvement as well as division into 2-3 subsections.
5. Add future recommendations for further investigation as well.
*Author(s) should highlight all the modifications carried out in the paper.
Author Response
Response to Reviewer 2 Comments
Dear Editor and Reviewer:
Thank you for your decision and constructive comments on my manuscript. We have carefully considered the suggestion of Reviewer and make some changes. We have tried our best to improve and made some changes in the manuscript. The following are our detailed responses to the issues raised by the reviewer.
Point 1: What is the main question addressed by the research? This needs to be included in the Introduction section. There are so many objectives claimed however same should also be specifically drawn herein.
Response 1: Thank you for this very valuable question. In this study, the main problem we solve is that the traditional convolutional neural network architecture is not easy to take into account the connection between regional features. In the introduction, we sort out the methods and problems in the tile block defect detection and propose the corresponding solution to the problem. Our method is to connect tile block defect features with regional features, and thus propose a new attention mechanism -- CPAM, so as to effectively establish the relationship between different tile block defect regions and extract tile block defect features.
Point 2: The literature presented in introduction section need to be presented in the Tabular form as it is difficult to follow. Particularly for other defects are required to be outlined clearly.
Response 2: Thank you for this very valuable question. In the introduction section, we have compiled information about the methods and problems in tile block defect detection and present these using Tabular.
Point 3: Quality of the Heat maps needs to be improved.
Response 3: Thank you for your advice. We have restocked Heat maps to improve the quality.
Point 4: The discussion section 4 is too small. Literature support to this section needs to be inculcated with an overall improvement as well as division into 2-3 subsections.
Response 4: Thank you for your suggestion. In the discussion section, we have reorganized the content about complex texture tile block defect detection. The content is further expanded into three sections from the application object to the enhancement of the method, which makes the discussion section more clear and full.
Point 5: Add future recommendations for further investigation as well.
Response 5: Thank you for your suggestion. In the conclusion section, we have expanded on future work in response to the limitations that currently exist.
Reviewer 3 Report
This paper proposes to use the CNN architecture with an attention mechanism to detect block defection, generally, it is well written and has good contribution compared to related works, however, it would be better if the authors adds some information about the environment used to implement the proposal as well as the error vs epoch.
Author Response
Response to Reviewer 3 Comments
Dear Editor and Reviewer:
Thank you for your decision and constructive comments on my manuscript. We have carefully considered the suggestion of Reviewer and make some changes. We have tried our best to improve and made some changes in the manuscript. In Section 3.2, we have provided our experimental environment and specific parameter values, and all experiments are conducted in the same environment. In addition, we analyzed the impact of different parameters on performance in Section 3.4.2. Thanks again.
Author Response File: 
Author Response.docx
Round 2
Reviewer 1 Report
The authors propose a new attention mechanism called the Cross Patch Attention Module to tile block defect detection.
The article has been improved, the contribution is good and all questions have been effectively answered.
Reviewer 2 Report
Author(s) have done well. In my opinion, the paper is now ready for publication.
