A Semi-Supervised Inspection Approach of Textured Surface Defects under Limited Labeled Samples

Round 1

Reviewer 1 Report

1.- Please improve the quality of some figures (Figure 1, 2, and 5) too blurry, low resolution.

2.- Line 151 Where...  is "something is missing". Please correct.

3.- Line  162 (Be..) in capital letter.

4.- Please explain why use ReELU function in the Generator of GAN architecture and LeakyReLU in the discriminator.

5.- Line 175 "...distribution of the variance is 0". Please rewrite.

6.- "...included into labeled ones as training data." and unlabeled data (U) right?. Please, complete.

7.- The description of the dataset begins in line 288 where some references to them appears before in the text as shown in Figure 3. Please reorder these references.

8.-Line 232: How do you define the threshold for estimating if there are enough samples for avoid over-fitting?

9.- Line 265: Symbol representing the threshold is missing.

10.- Line 280: Errata, please remove "is".

11: Line 282: Even it is well know, I suggest to a  add a ref. to Adam optimization algorithm.

12: Line 316: Term fuzzy refers to "blurry"?. Please clarify.

13: Line 328: replace "Classification accuracy has been increase" by "classification accuracy increases".

14.: Line 354-355. Please explain better the sentence. It's not clear.

15.- Line 376: "Training configuration is described in section 2". Should be section 3 right?

16.- Table 3. The description of the number of parameters is wrong. Please correct.

17.- Line 380: "Consumption time"?. Please correct.

18: Line 384: "We observe that each additional block in the ALLnet brings approximately 1% increase in the overall accuracy, and this trend will continue until there are five blocks in the model". Not true. Starting from 4B the accuracy converges, as shown in Table 3.

19: Line 387: In the same paragraph "computation time", "computation speed" and "consumption time" are used. Please use the same term for referring the same concept.

20: Line 395: "We selected the supervised methods are" should be "The selected supervised methods are..."

Author Response

Thanks a lot for your kind suggestions. We have fully checked and modified the manuscript. More details can be found in the revised manuscript. The point-by-point responses are as following:

Q1: Please improve the quality of some figures (Figure 1, 2, and 5) too blurry, low resolution

A1: We have replaced the clearer figures for the Figure 1, 2, and 5.

Q2: Line 151 Where...  is "something is missing". Please correct.

A2: We have added the symbol E in line 151.

Q3: Line 162 (Be..) in capital letter.

A3: We have corrected this error in line 162.

Q4: Please explain why use ReELU function in the Generator of GAN architecture and LeakyReLU in the discriminator.

A4：The settings of output activation functions are followed by the outstanding conclusions in [27]. We add this reason in line 170.

Q5: Line 175 "...distribution of the variance is 0". Please rewrite.

A5: We have corrected the wrong way of writing in line 175.

Q6: "...included into labeled ones as training data." and unlabeled data (U) right?. Please, complete.

A6: The unlabeled data will be assigned to the pseudo labels according to the class scores. Therefore, these samples are unlabeled before training.

Q7: The description of the dataset begins in line 288 where some references to them appears before in the text as shown in Figure 3. Please reorder these references.

A7: We have c reorder these references according to their numbers in line 291.

Q8: Line 232: How do you define the threshold for estimating if there are enough samples for avoid over-fitting?

A8: Because there are kinds of defect datasets, the thresholds are different. We cannot define a widely applicable threshold. The current research is also difficult to achieve it. Therefore, judge the training quality should refer to the specific training results, and ensure as many samples as possible.

Q9: Line 265: Symbol representing the threshold is missing.

A9: We have added the symbol τ in line 266.

Q10: Line 280: Errata, please remove "is"

A10: We have removed the “is” in line 281.

Q11: Line 282: Even it is well know, I suggest to a add a ref. to Adam optimization algorithm

A11: We have added this ref. in [32].

Q12: Line 316: Term fuzzy refers to "blurry"?. Please clarify.

A12: We have rewritten the controversial sentences in line 316

Q13: Line 328: replace "Classification accuracy has been increase" by "classification accuracy increases".

A13: We have corrected this error in line 329.

Q14: Line 354-355. Please explain better the sentence. It's not clear

A14: We have rewritten the controversial sentences in line 354-355.

Q15: Line 376: "Training configuration is described in section 2". Should be section 3 right?

A15: We have corrected this error in line 378.

Q16: Table 3. The description of the number of parameters is wrong. Please correct.

A16: We have corrected these wrong contents in Table 3.

Q17: Line 380: "Consumption time"? Please correct.

A17: We have replaced with “computation time” in line 382.

Q18: Line 384: "We observe that each additional block in the ALLnet brings approximately 1% increase in the overall accuracy, and this trend will continue until there are five blocks in the model". Not true. Starting from 4B the accuracy converges, as shown in Table 3.

A18: We have corrected this error in line 388.

Q19: Line 387: In the same paragraph "computation time", "computation speed" and "consumption time" are used. Please use the same term for referring the same concept.

A19. have corrected this error. We only use the “computation time” to avoid readers confusion.

Q20: Line 395: "We selected the supervised methods are" should be "The selected supervised methods are..."

A20: We have rewritten the controversial sentences in line 398-399.

Reviewer 2 Report

Researchers proposes a semi-supervised framework, based on a generative adversarial network (GAN) and a convolutional neural network (CNN), to classify defects of textured surfaces. The proposed approach is promising and the corresponding accuracy rates are higher when compared to other methods. I just want to advice a further discussion of the literature section. Following papers can be considered for putting a full picture defect detection.

1-Y Kahraman, A Durmuşoğlu, Deep learning-based fabric defect detection: A review, Textile Research Journal, 00405175221130773

2-Y Kahraman, A Durmuşoğlu, Classification of Defective Fabrics Using Capsule Networks, Applied Sciences 12 (10), 5285

Author Response

Thank you very much for your valuable comments. We think these papers are very suitable for this manuscript. They have been added in the Reference.

Reviewer 3 Report

In the introduction,  What major controversy, or paradox does this research address?

Further, in the introduction, what is the current knowledge gap of the main literature that the authors need to write this research? What we have known and what we have not known? What is missing from the reported works in literature?  Why does it need to be addressed?

Have other deep learning techniques tested  on same dataset ?

Why GAN is selected only.

  Compare results in latest state of art

Author Response

Thank you very much for your kind suggestions. We have fully checked and modified the manuscript. More details can be found in the revised manuscript. The point-by-point responses are as following:

For the first question, this paper mainly deals with the problem of limited labeled samples in industrial defect inspection tasks. Therefore, we propose a semi-supervised defect classification method which uses a GAN to generate new samples, a novel label assignment scheme to assign pseudo labels, and a CNN to perform defect classification task.

For the second question, we revised the introduction part to make the topic clearer. Specifically, we first analyze the shortcomings of the existing methods and introduce the necessity of the SSL methods for defect inspection in introduction part. To establish the SSL framework, we think there are two issues that need to be addressed——the acquisition of unlabeled samples（line 94）and how to include them into training (line104). In these corresponding paragraphs, we discuss the existing methods in the literature and point out their shortcomings. To overcoming their shortcomings, we write this research. The things what we have known that have been written in the introduction. Moreover, for the ones what we have known, we think it is whether the proposed SSL method can overcome the above-mentioned shortcomings as we discussed in the introduction part. Finally, we added the reasons why the problems should be addressed in line 40-43.

we mainly use the generated defect samples and the more accurate label assignment algorithm for defect inspection tasks. Our semi-supervised learning method can be used on kinds of different industrial production surfaces. The experiments can provide it. As mentioned in the introduction, the deep-learning-based methods for industrial defect classification are based on a single surface which lacks of generalization.

For the third question, the selected comparison methods in the experiment are all based on deep learning technique. They are all tested on the four industrial defect datasets in experiment part.

For the last question, we select GAN to generate defect samples because texture defects are Irregular. Many researches also prove that a GAN is more suitable for this situation than other models, e.g., auto-encoders. This manuscript focuses on the industrial inspection, so Calculation speed and accuracy are equally important. However, the state of art methods always use very large models but only a few precision improvements can be obtained, which can hardly meet the actual needs of industrial production.