Improved U-Net++ with Patch Split for Micro-Defect Inspection in Silk Screen Printing

Round 1

Reviewer 1 Report

Comments

1 – Line 95 to line 99

This paragraph (lines 95 to 99) stating the organization of the remainder of the paper should be the last paragraph of section 1. Please exchange the order of this paragraph with the one that is spanning on lines 100 to 107.

 

In section 1, first summarize the key contributions of the paper. Then, state the organization of the remainder of the paper.

 

2 - On the caption of Figure 1 (line 129), please delete “(* Source: U-Net: Convolutional Networks for Biomedical Image Segmentation)” and instead provide the correct numeric reference.

 

3 – On the caption of Figure 2 (line 140), please delete “(* Source: U-Net++: A Nested U-Net Architecture for Medical Image Segmentation)” and instead provide the correct numeric reference.

 

4 – A suggestion: on the caption of Figure 4, line 170, please provide more details on the types of defects therein depicted.

 

5 - A suggestion: on the caption of Figure 5, line 183, please provide more details on the meaning of the yellow, red, and green squares over the input images.

 

6 – On lines 203 and 204, we have “When training is performed by resizing the original 2448x2048 pixels image to 256x256 pixels”….However, in Figure 7, the depicted input image has a 1024x1024 resolution. This is confusing. Please check.

 

7 – On line 208, we have “randomly cropped into patch units 256x256 pixels that can”….However, in Figure 7, the depicted image patched have a 254x254 resolution. This is also confusing. Please check.

 

8 – On line 214, we have “crop, clahe, and resize“. I think that with should be something like “crop, contrast enhancement with the Contrast Limited Adaptive Histogram Equalization (CLAHE) technique [26], and resize”. Moreover, since there are different contrast improvement techniques, the authors can explain on why they choose this particular technique.

 

9 – Line 218. Please correct “image can be seen in Figure 0”. Is it Figure 8?

 

10 – On the Caption of Figure 8, the authors may add the literature reference number [26] to the (CLAHE) technique.

 

11 – On lines 227 to 229, we have “The purpose is to increase the defect area compared to the entire input data area to improve the model's inference performance.” If possible, please provide more details on this. A suggestion: over Figure 9, please provide an example of on how this works.

 

12 – On lines 254 and 255, we have “The method and reason for the patch split will be omitted as it has been described in this paper.”. I think that it is better to state that “The method of patch split was described in section ## and it will not be further discussed.”.

 

13 – Line 262. Please, provide appropriate literature references to the HoughLinesP and the Canny Edge Detection techniques.

 

14 – Figure 10. On the flowchart, there is no clear begin state/action and end state/action. Please, clarify on this beginning and ending states (like a state machine).

 

15 - Figure 10. On the approach depicted in Figure 10, for how many different products the proposed solution is intended to work.

 

16 – Line 280. Please, provide a definition for the Dice score on its first occurrence in the text.

 

17 – Line 281. Please, define the FPN acronym on its first occurrence. It is defined as Feature Pyramid Network later in line 400.

 

18 – Lines 292 and 293. We have “because it was judged to be suitable for the length problem.”

Please, clarify what is the length problem?

 

19 - On the caption of Figure 11 (line 304), please delete “(* Source: Focal Loss for Dense Object Detection)” and instead provide the correct numeric reference.

 

20 – Please treat equations as elements of text. Equation (1) should end with a final dot.

 

21 – Please, explain the meaning of all the symbols in each equation. For instance, in equation (1), what is the meaning of \alpha, \beta, and all the other symbols? In equation (2) what is the meaning of y and \tilde{y}?

 

22 – Figure 12 depicts the global architecture of the proposed solution, and it is placed on Section 4 that addresses the experimental evaluation of the method. I suggest placing this Figure in Section 3, in which the authors describe the proposed solution.

 

23 – Section 4.1 and elsewhere on the paper. The authors use the terms “data set” and “dataset”. Please, use the same consistent designation/notation across the paper.

 

24 – Lines 365 and 366. Please, define the TP, FP, FN, and TN acronyms.

 

25 – On line 367 we have “Zijdenbos et al. were among the first”. Please, add the proper numeric literature reference.

 

26 – Figure 16. On the experimental results on this figure, it would be nice to know the minimum number of epochs that reaches the maximum Dice score value.

 

27 – Lines 422 and 423. We have “Looking at Figure 19, our proposed method only finds defects and the segmentation area is also correct.” My question here is that the proposed method can find *all* the defects, with no exception?

 

 

Some comments on writing

 

Line 62

Please, define the GPU acronym

 

Line 63

Please, define the CNN acronym

 

Line 73

Patch Split Method is a method to solve a problem

->

Patch Split Method solves a problem

 

Line 77

First, We will explain

->

First, we will explain

 

 

Line 117

Convolutional Auto Encoder [19, 20]. CAE is used in various

->

Convolutional Auto Encoder (CAE) [19, 20]. CAE is used in various

 

Lines 123 and 124

to the original image, A segmentation map

->

to the original image, a segmentation map

 

Line 198

CNNs

->

CNN

 

Lines 201 and 202

The sentence “Introduces the methodology of training and inference of artificial intelligence models that can detect micro-defects using patch splits.” seems to be misplaced. It looks like a summary/description on the contents of the section. Please, check on this.

 

Line 225

Defects are found using a U-net++ model trained

->

Defects are found using a U-Net++ model trained

 

Line 252

Clahe pre-processing is

->

CLAHE pre-processing is

 

Lines 285 and 286

The architecture of the backbone network used EfficientNet-b0.

->

The architecture of the backbone network used is EfficientNet-b0.

Line 300

as shown in Fig. 11.

->

as shown in Figure 11.

 

Line 301

bonding box

->

bounding box

 

Line 309

f2 score

->

F2 score

 

dice score, so it is suitable for optimizing dice score

->

Dice score, so it is suitable for optimizing Dice score

 

Line 350

Each image consists of 2448 (w) x 2048 (v)

->

Each image consists of 2448 (w) x 2048 (h)

 

Line 400

We have “shown in Table 0,”. Please correct this.

 

Lines 410 to 412, we have these two sentences:

 

“The Loss Function we propose is a Combine Loss that combines Focal Loss and Tversky Loss. Finally, we use a loss function that combines focal loss and Tversky loss.”

 

The second sentence is repeating the first one. Please revise.

 

Line 415

the Tversky loss function, so I used to combine loss.

->

the Tversky loss function, so we used it to combine loss.

 

Line 427

an improved U-Net++ applying

->

an improved U-Net++ model applying

 

Line 430

We have “that the Ice Score was”. I think it should be “Dice score”. Please check.~

 

Reference list

Please check your references.

Ref. 16. u-net -> U-Net

Ref. 17 ieee -> IEEE

Ref. 20 bayes -> Bayes

Ref. 24 and 34 have repeated year

Ref. 29 Iou loss for 2d/3d -> IoU loss for 2D/3D

Ref. 35 seems to be incomplete

Author Response

Dear Reviewer 1,

Special Issue " Applications of Deep Learning and Artificial Intelligence Methods" of Multidisciplinary Digital Publishing Institute (MDPI) Applied Sciences

Thank you very much for your letter of April 21^st, 2022 regarding my paper ID applsci-1666655 entitled “Improved U-Net++ with Patch Split for Micro-defect Inspection in Silk Screen Printing”.

 

 

Based on the reviewer’s comments, I have incorporated them in the revised version. The comments regarding our manuscript were extremely helpful to us in preparing a clearer version. We have rewritten many paragraphs according to the recommendations of the referees. Thank you very much for your advice. Attached is a copy of the revised version of the manuscript and a list of the revisions. Your acknowledgment will be greatly appreciated. Thanks once again for your significant help with this paper.

Thanks once again for your significant help with this paper.

Sincerely yours,

Byungguan Yoon

Author Response File: Author Response.pdf

Reviewer 2 Report

The work presented in this manuscript is about investigation and examination of an improved U-Net++ utilizing Patch Split Method applied on micro-defect inspection in Korean silk screen printing.  The work is well presented with the following merits:

(1) Clearly identification of the engineering needs with the proposed method supported by sufficient literature survey.

(2) The work has potential engineering implications in improving economic values for Korean silk screen printing.  Very interesting applied research work indeed.

(3) The approaches used in investigate the proposed method are fine and reasonable and the results support the claim stated in the conclusion session.

The only two things are needed to be improved for possible publication in this journal are:

(1) Mostly likely no English errors are found in this manuscript but enhancing writing style to be more academic is recommended.

(2) The quality of the figure 16,  and figure 19 should be improved.

Author Response

Dear Reviewer 2,

Special Issue " Applications of Deep Learning and Artificial Intelligence Methods" of Multidisciplinary Digital Publishing Institute (MDPI) Applied Sciences

Thank you very much for your letter of April 21^st, 2022 regarding my paper ID applsci-1666655 entitled “Improved U-Net++ with Patch Split for Micro-defect Inspection in Silk Screen Printing”.

 

 

Based on the reviewer’s comments, I have incorporated them in the revised version. The comments regarding our manuscript were extremely helpful to us in preparing a clearer version. We have rewritten many paragraphs according to the recommendations of the referees. Thank you very much for your advice. Attached is a copy of the revised version of the manuscript and a list of the revisions. Your acknowledgment will be greatly appreciated. Thanks once again for your significant help with this paper.

Thanks once again for your significant help with this paper.

Sincerely yours,

Byungguan Yoon

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors propose an automated quality inspection system based on an improved U-Net++. Instead of using the original image level,  Patch level has been used which illustrated good accuracy with Dice score of 0.728. It demonstrates that the Patch Split Method-based AI is useful for relatively small dataset. The current study is interesting. In general, the main conclusions presented in the paper are supported by the figures and supporting text. However, to meet the journal quality standards, the following comments need to be addressed.

 

1. Abstract: Should be improved and extended. The authors talk lot about the problem formulation, but novelty of the proposed model is missing. Also provided the general applicability of their model. Please be specific what are the main quantitative results to attract general audiences.
2. The introduction can be improved. The authors should focus on extending the novelty of the current study. Emphasize should be given in improvement of the model (in quantitative  sense)  compared to   existing  state-of-the art models.
3. The authors should cover some  state-of-the-art CNN   detection algorithms used in various scenarios  that can be extended to the present study (see : Expert Sys App (2021) 172, 114602 https://doi.org/10.1016/j.eswa.2021.114602; Comp  Elect in Agri (2022), 193 106694 https://doi.org/10.1016/j.compag.2022.106694). So that the general readers aware of it. Hence they should be refereed.
4. Section 2.1: Not very clear what modification have been done on original UNet ++ network. So this section can be followed by section 3. Also subsection 2.2 can be in a separate section.
5. More details about network architecture and complexity of the model should be provided.
6. Table 4: what about comparison of the result with current state-of-the art models? Did authors perform ablation study to compare with different models?
7. Section 4.1. Methodology: Did the authors employ any data enhancement methods (augumentation) before training? If so, it should be mentioned. Also, all hyperparameters (learning rate, mini-batch size, number of epochs, optimizer) and model complexity should be detailed.

 

8. Conclusion parts needs to be strengthened.
9. Please provide a fair weakness and limitation of the model, and how it can be improved.
10. Typographical errors: There are several minor grammatical errors and incorrect sentence structures. Please run this through a spell checker.

Author Response

Dear Reviewer 3,

Special Issue " Applications of Deep Learning and Artificial Intelligence Methods" of Multidisciplinary Digital Publishing Institute (MDPI) Applied Sciences

Thank you very much for your letter of April 21^st, 2022 regarding my paper ID applsci-1666655 entitled “Improved U-Net++ with Patch Split for Micro-defect Inspection in Silk Screen Printing”.

Based on the reviewer’s comments, I have incorporated them in the revised version. The comments regarding our manuscript were extremely helpful to us in preparing a clearer version. We have rewritten many paragraphs according to the recommendations of the referees. Thank you very much for your advice. Attached is a copy of the revised version of the manuscript and a list of the revisions. Your acknowledgment will be greatly appreciated. Thanks once again for your significant help with this paper.

Thanks once again for your significant help with this paper.

Sincerely yours,

Byungguan Yoon

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Overview

This version of the paper is clearly improved, as compared to the first version. The authors have improved the paper in different aspects, following the reviewers comments. We now have more figures, more references, and a better explained approach.

The authors have also provided a very satisfactory response letter to my comments.

 

I have only a few comments on this version of the paper.

1 – Regarding the rebuttal report, I agree with all the responses except Response 16. Notice that the Dice Score definition is still in section 4.2 on equation (2), line 423. Thus, the Dice Score is still not defined on its first occurrence (Line 316).

 

2 – Line 238

of the local region after CLAHE

->

of the local region after Contrast Limited Adaptive Histogram Equalization (CLAHE)

 

3 – Line 327

Learning rate is 0.0001 in float form, The optimizer adopts AdamW

->

Learning rate is 0.0001 in float form. The optimizer adopts AdamW

 

4 – Line 330

used for various tasks, Partial restarts are also

->

used for various tasks. Partial restarts are also

 

5 – Line 332

with ill-conditioned functions Therefore, as a scheduler

->

with ill-conditioned functions. Therefore, as a scheduler

 

6 – Line 421

between two segmentations ? and ?̃ including.

->

between two segmentations named ? and ?̃.

 

7 – On the definition of the Dice score, please show the range of the values of the score and their meaning.

Author Response

Thank you very much for your letter of April 21st, 2022 regarding my paper ID applsci- 1666655 entitled “Improved U-Net++ with Patch Split for Micro-defect Inspection in Silk Screen Printing”.

Based on the reviewer’s comments, I have incorporated them in the revised version. The comments regarding our manuscript were extremely helpful to us in preparing a clearer version. We have rewritten many paragraphs according to the recommendations of the referees. Thank you very much for your advice. Attached is a copy of the revised version of the manuscript and a list of the revisions. Your acknowledgment will be greatly appreciated. Thanks once again for your significant help with this paper.

Thanks once again for your significant help with this paper. Sincerely yours,

Byungguan Yoon, Graduate Student.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors incorporated reviewers comments satisfactorily. The manuscript can be accepted in its current form. 

Author Response

Dear Reviewer 3,
Special Issue " Applications of Deep Learning and Artificial Intelligence Methods" of Multidisciplinary Digital Publishing Institute (MDPI) Applied Sciences

Thank you very much for your letter of April 21st, 2022 regarding my paper ID applsci- 1666655 entitled “Improved U-Net++ with Patch Split for Micro-defect Inspection in Silk Screen Printing”.

Thank you very much for your advice. Your acknowledgment will be greatly appreciated. Thanks once again for your significant help with this paper.

Sincerely yours,

Byungguan Yoon, Graduate Student.

Author Response File: Author Response.pdf