Insulated Gate Bipolar Transistor Solder Layer Defect Detection Research Based on Improved YOLOv5

Round 1

Reviewer 1 Report

In this manuscript, the authors present the defect detection model of IGBT based on the deep neural network. I have a couple of comments regarding this manuscript.

1) Overall English expressions should be revised. There are typos and unnatural expressions.

2) It is believed that how the authors obtained the raw images for the dataset of IGBT solder layer defects should be explained more thoroughly.

3) It would be better to explain the network structure of YOLOv5 with more details including how this network can detect defects with which output node.

4) I think how YOLOv5 has been imporved can be highlighted in Figure 3.

5) The raw image and processed images can be compared for a better understanding.

Author Response

Dear Editors and reviewers:

Re: Manuscript ID: applsci-1977334 and Title: Insulated Gate Bipolar Transistor solder layer defect detection

Thank you for your precious comments and advice. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our research. We have studied the comments carefully and have made a correction which we hope meets with approval. Revised portions are marked in blue on the paper. The main corrections in the paper and the responses to the reviewer’s comments are as flowing:

 

1. Overall English expressions should be revised. There are typos and unnatural expressions.

Response: Thank you for your valuable and thoughtful comments. We apologize for the poor language of our manuscript. Previously we worked on the manuscript for a long time and repeatedly revised sentences resulting in poor readability of the article. We have carefully checked and improved the English writing in the revised manuscript. We hope that the flow and language level have been substantially improved.

2. It is believed that how the authors obtained the raw images for the dataset of IGBT solder layer defects should be explained more thoroughly.

Response: Thank you for your valuable and thoughtful comments. To be more clear and follow the reviewer's concerns, We have provided a thorough explanation of the network structure of YOLOv5 in section 2.3. See lines 170-217 for details of the changes.

3. It would be better to explain the network structure of YOLOv5 with more details including how this network can detect defects with which output node.

Response: Thank you for your valuable and thoughtful comments. We appreciate it very much for this good suggestion, and we have done it according to your ideas. We have explained the YOLOv5 network structure in more detail in section 2.3 in response to your comments, see lines 170-217 for details of the changes.

4. I think how YOLOv5 has been imporved can be highlighted in Figure 3.

Response: Thank you for your valuable and thoughtful comments. According to the reviewer's suggestion, We have added the structure diagram of the C3STR module to make the improved details in Figure 3 more prominent by using different colors to distinguish the added detection layers.

5. The raw image and processed images can be compared for a better understanding.

Response: Thank you for your valuable and thoughtful comments. According to the reviewer's suggestion, we have discussed that due to the characteristics of the bubble defect, comparing the raw image with the processed image will make the results understandable. Therefore we have put the raw image in Figure 5, please see Figure 5 for details of the modifications.

 

Thank you for your careful review. We appreciate your efforts in reviewing our manuscript during this unprecedented and challenging time. We wish good health to you, your family, and your community. Your careful review has helped to make our study clearer and more comprehensive.

Reviewer 2 Report

In this manuscript, an algorithm based on an improved model of YOLOv5 is proposed to detect bubble defects. The paper focuses on the methodology and the result. I have the following comments.

1. The authors are encouraged to clearly mention the novelty of this work compared to previous works, such as adding a detection layer in the model or optimizing the parameters.

2. Texts in Figures 7 and 9 should be enlarged.

3. The resolution of Figures 1, 3, and 5 should be improved.

4. Why does the addition of the small target detection layer improve mAP and accuracy? Is this just about the depth?

5. Can the proposed model be applied to the detection of other defects instead of bubble defects? Please mention the expandability of the model.

Author Response

Dear Editors and reviewers:

Re: Manuscript ID: applsci-1977334 and Title: Insulated Gate Bipolar Transistor solder layer defect detection

Thank you for your precious comments and advice. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our research. We have studied the comments carefully and have made a correction which we hope meets with approval. Revised portions are marked in blue on the paper. The main corrections in the paper and the responses to the reviewer’s comments are as flowing:

 

1. The authors are encouraged to clearly mention the novelty of this work compared to previous works, such as adding a detection layer in the model or optimizing the parameters.

Response: Thank you for your valuable and thoughtful comments. We appreciate it very much for this good suggestion, and we have done it according to your ideas. According to the reviewer's suggestion, We have elaborate on the background of the study and the reasons for the improvement of our method. In lines 67-134 of the introduction, we present the differences between this paper and previous methods as well as the advantages, and in lines 136-148 of the related work, we summarize the studies of other scholars and present the improvements made in this study based on them.

2. Texts in Figures 7 and 9 should be enlarged.

Response: Thank you for your valuable and thoughtful comments. We appreciate it very much for this good suggestion, and we have done it according to your ideas. Redraw and reposition Figures 7 and 9, see pages 12 and 17 for detailed modifications.

3. The resolution of Figures 1, 3, and 5 should be improved.

Response: Thank you for your valuable and thoughtful comments. We appreciate it very much for this good suggestion, and we have done it according to your ideas. We took the original images of Figures 1, 3, and 5 and used the higher-resolution export settings to get a higher-resolution image. For detailed modifications, please see Figures 1, 2, and 5.

4. Why does the addition of the small target detection layer improve mAP and accuracy? Is this just about the depth?

Response: Thank you for your valuable and thoughtful comments. In addition to increasing the network depth, there are many reasons for adding a small target detection layer to make the mAP and accuracy increase. Before adding a small target detection layer, a larger feature map for detecting small targets needs to be obtained. Although enlarging the feature map can increase the deep semantic information, it may lose the shallow location information at the same time. Therefore， we increase the model's attention to small targets by merging small feature maps and enriching the feature map information to improve the detection accuracy of the model. In this part, we make corresponding changes in the paper, see lines 253-262 in the paper.

5. Can the proposed model be applied to the detection of other defects instead of bubble defects? Please mention the expandability of the model.

Response: Thank you for your valuable and thoughtful comments. We appreciate it very much for this good suggestion, and we have done it according to your ideas. We use the NEU-DET open-source dataset to validate the scalability of our proposed model, see lines 471-479 of the article for details of the validation results.

 

Thank you for your careful review. We appreciate your efforts in reviewing our manuscript during this unprecedented and challenging time. We wish good health to you, your family, and your community. Your careful review has helped to make our study clearer and more comprehensive.

 

Round 2

Reviewer 1 Report

The authors well addressed the comments raised by the reviewers. But, I wonder if a complicated neural network can lead to enormous energy consumption, so it might be conducted by hardware-based neural network, so-called neuromorphic system. It is also believed that this whole process can be efficiently done in one-chip if IGBT and it can be integrated together. Can the authors introduce the possibility and feasibility of hardware-based neural network for this purpose with additional ciations?

- Menelaos Skontranis, et al., "Time-Multiplexed Spiking Convolutional Neural Network Based on VCSELs for Unsupervised Image Classification," Applied Science, vol. 11, p. 1383, 2021.

- T.-H. Kim, et al., "Effect of Program Error in Memristive Neural Network with Weight Quantization," IEEE Transactions on Electron Devices, vol. 69, no. 6, pp. 3151-3157, 2022.

- I. Chakraborty, et al., "Technology Aware Training in Memristive Neuromorphic Systems for Nonideal Synaptic Crossbars," IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 2, no. 5, pp. 335-344, 2018. 

Author Response

Dear Editors and reviewers:

Re: Manuscript ID: applsci-1977334 and Title: Insulated Gate Bipolar Transistor solder layer defect detection

Thank you for your precious comments and advice. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our research. We have studied the comments carefully and have made a correction which we hope meets with approval. Revised portions are marked in red on the paper.

We appreciate it very much for this good suggestion, and we have done it according to your ideas.

See lines 99 - 160 for details of the changes. We also did a more detailed check on the spelling of words.

Thank you for your careful review. We appreciate your efforts in reviewing our manuscript during this unprecedented and challenging time. We wish good health to you, your family, and your community. Your careful review has helped to make our study clearer and more comprehensive.