Unsupervised Anomaly Detection in Printed Circuit Boards through Student–Teacher Feature Pyramid Matching

Round 1

Reviewer 1 Report

1. Authors have presented An Unsupervised Anomaly Detection in Printed Circuit Boards through Student-Teacher Feature Pyramid Matching.

2. More Literature review should be added with of Latest Papers.

3. Reference Number [38] was cited in the running text but not in the Reference List. So, I request the Authors to update the same.

4. Increase the size of the figures in the results section and should be elaborated in detail about the analysis part.

    

Author Response

Reviewer 1

An Unsupervised Anomaly Detection in Printed Circuit Boards through Student-Teacher Feature Pyramid Matching 

Venkat Anil Adibhatla ^{1, 3}, Yu-Chieh Huang ², Ming-Chung Chang ², Hsu-Chi Kuo², Abhijeet Utekar ², Huan-Chuang Chih ³, Maysam F. Abbod ^4,* and Jiann-Shing Shieh ^1,*

More Literature review should be added with of Latest Papers.

 Answer:   Extra literature has been added according to your suggestion from lines 77-85 at page 2 and lines 94-100 at pages 2-3.

Reference Number [38] was cited in the running text but not in the Reference List. So, I request the Authors to update the same.

Answer: We have already updated and corrected the mistake. It has been edited in line 390 at page 14 because the reference for PaDIM and PatchCore was same.

Increase the size of the figures in the results section and should be elaborated in detail about the analysis part.

Answer: As per your suggestion, we have increased the size of figures in result which is displayed at page numbers 9,10,11,12 and have elaborated the explanation at lines 288-297 at page 8, and lines365-373 at page 13.  

Reviewer 2 Report

General Comments:

The paper deals with the detection of defects on printed circuit boards (PCBs) and presents an unsupervised learning approach to handle the problem. The approach is based on using pre-trained image classification model that learns the distribution of faultless images. The topic is important for industrial electronics and the proposed approach is more efficient than supervised methods. However, clarity of presentation is the main issue.

Specific Comments:

1. Sub-Section 2.2.1: This sub-section is written in a very brief manner. There should be proper reasoning and explanation for the proposed framework and its elements. For example, why do we select low-level information for encoding? Before that what is the definition of low-level information?

2. Sub-Section 2.2.2:

a) It is unclear how the feature map is generated.

b) It is unclear what the feature vectors are; apart from “the most important feature vectors”.

c) It is unclear how the distance measure in Equation (1) is considered as cosine similarity. In addition, why such a measure is selected and used?

d) The coefficients {\alpha_l} in Equation (3) as well as their impact are unclear.

e) Please check whether there is a mix-up between bold-type and normal symbols.

3. Sub-Section 2.2.3:

a) The upsampling process in Equation (4) is in need for explanation.

b) The definition of the anomaly score should be justified.

c) I think there is a mix-up between bold-type and normal symbols in this sub-section.

Language Usage:

Moderate language revision is necessary. Just as examples:

1. Line 43, “different electronic device” should be “different electronic devices”.
2. Line 44, “any electronics product” should be “any electronic product”.
3. Line 49, “have been proved that” should be “proved that”.
4. Line 50, “have been surpassed” should be “surpass” or “outperform”.
5. Lines 67-68, “there are many drawbacks for the above methods it needs a lot of data” should be “there are many drawbacks in the above methods, for example, they need a lot of data”.
6. Line 134, “anomaly free images” should be “anomaly-free images”.
7. Line 143, I think that “(I,j)” should be “(i,j)”.
8. Line 147, “It is excellence mentioning that” should be “It is worth mentioning that”.

Author Response

Reviewer 2

An Unsupervised Anomaly Detection in Printed Circuit Boards through Student-Teacher Feature Pyramid Matching 

Venkat Anil Adibhatla ^{1, 3}, Yu-Chieh Huang ², Ming-Chung Chang ², Hsu-Chi Kuo², Abhijeet Utekar ², Huan-Chuang Chih ³, Maysam F. Abbod ^4,* and Jiann-Shing Shieh ^1,*

1. Sub-Section 2.2.1: This sub-section is written in a very brief manner. There should be proper reasoning and explanation for the proposed framework and its elements. For example, why do we select low-level information for encoding? Before that what is the definition of low-level information?

Answer:

The explanation of the above comments regarding sub-section 2.2.1 has been explained in lines 140-159 at pages 3-4.

2. Sub-Section 2.2.2:

1. a) It is unclear how the feature map is generated.

1. b) It is unclear what the feature vectors are; apart from “the most important feature vectors”.

Answer:

The answer for a) and b) are explained in lines 201-205 at page 5.

1. c) It is unclear how the distance measure in Equation (1) is considered as cosine similarity. In addition, why such a measure is selected and used?

Answer: The explanation is added to lines 184-190 at page 5.

1. d) The coefficients {\alpha_l} in Equation (3) as well as their impact are unclear.

Answer: The impact is explained in lines 196-198 at page 5.

1. e) Please check whether there is a mix-up between bold-type and normal symbols.

Answer: In line 180 at page 4, there was a mistake previously it was (I, j), but now we have changed to (i, j).

3. Sub-Section 2.2.3:
4. a) The upsampling process in Equation (4) is in need for explanation.
5. b) The definition of the anomaly score should be justified.

Answer: As you have asked for explanation for equation 4 and to elaborate the definition of anomaly score. We have added the explanation to lines 218-225 at pages 5-6.

1. c) I think there is a mix-up between bold-type and normal symbols in this sub-section.

Answer: As you have suggested me to check the mix up between bold type and bold type, I have checked it and made changes in line 208 at page 5  where it was “J” and we have changed it to “ J “ and in line  226 at  page 6 it was “J” and we have changed it to “ J “.

Moderate language revision is necessary. Just as examples:

Line 43, “different electronic device” should be “different electronic devices”.

Answer: In line 43 at page 1, “different electronic device” has changed to “different electronic devices”.

Line 44, “any electronics product” should be “any electronic product”.

Answer: In line 44 at page 1, “any electronics product” has changed to “any electronic product”.

Line 49, “have been proved that” should be “proved that”.

Answer: In line 49 at page 2, “have been proved that” has changed to “proved that”.

Line 50, “have been surpassed” should be “surpass” or “outperform”.

Answer: In line 50 at page 2, “have been surpassed” has changed to “outperform”.

Lines 67-68, “there are many drawbacks for the above methods it needs a lot of data” should be “there are many drawbacks in the above methods, for example, they need a lot of data”.

Answer: In lines 66-67 at page 2, “there are many drawbacks for the above methods it needs a lot of data” has changed to “there are many drawbacks in the above methods, for example, they need a lot of data”.

Line 134, “anomaly free images” should be “anomaly-free images”.

Answer:  In line 170 at page 4 “anomaly free images” should be “anomaly-free images”.

Line 143, I think that “(I,j)” should be “(i,j)”.

Answer: In line 179 at page 4 “(I,j)” has changed to “(i,j)”.

Line 147, “It is excellence mentioning that” should be “It is worth mentioning that”.

Answer: In line 183 at page 5 “It is excellence mentioning that” has been changed to “It is worth mentioning that”.

Round 2

Reviewer 2 Report

The Authors have carefully addressed all of the Reviewer’s comments and modified the manuscript successfully. I find the current version useful and suitable for publication in MDPI Electronics.