A Deep Convolutional Generative Adversarial Networks-Based Method for Defect Detection in Small Sample Industrial Parts Images

Round 1

Reviewer 1 Report

The work, entitled "A DCGAN and Lightweight CNN Based Method for Defect Detection in Small Sample Industrial Parts Images" needs a major revision prior to being considered for publication in Applied Sciences.

Comments:

• Recently, it is recommended not to use abbreviations in the titles of manuscripts submitted to journals. The submitted manuscript contains two abbreviations in the title. Can you opt out of them?
• The title states that the manuscript is about "small samples". Unfortunately, in the manuscript, I have nowhere found any information about the dimensions of the gear in question. It would also be good to add how these samples were made and what material they were made of.
• Does the method of identification texture defects presented in the paper allow for the identification of products with wrong dimensions? If not, and this assessment is manual (thus allowing for the removal of defective products for texture defects), why add redundant action to the quality control procedure?
• Abstract, line 20: What does F1macro mean?
• Introduction, line 40: What does MODWOT mean?
• In the description, under equation 1, there is divergence, unfortunately it does not appear in the equation. Does this mean there is an error in the equation?
• The discussion related to Table 3 mentions "256 channels". Why does channel 384 appear in the table?
• Suggests putting pictures 7a and 7b one above the other. Since in the case of Figure 7a, the manuscript refers to percentages - scale to 1, why is the y-axis scale to 2 in Figure 7b?
• Please look at the first row in table 9.
• Please correct the captions under the figures. Their parts are located between the drawings that make up a figure with a specific number.
• Please check references. Errors have crept into their notation (e.g. position 25).

The main recommendation is careful reading of text by authors in order eliminates mistakes which occur in manuscript.

Author Response

请参阅附件

Author Response File: Author Response.docx

Reviewer 2 Report

The paper „A DCGAN and Lightweight CNN Based Method for Defect Detection in Small Sample Industrial Parts Images” by Gao et al. highlights important issues, but it is incorrectly written in terms of language and editing, as well as substantively. As the manuscript does not meet the standards of scientific work, I cannot recommend it for publication in its current form. In order to reconsider this article for publication, the following remarks should be considered at the outset.

- The presented method of defect detection is a technological solution, the study does not have the features of a scientific publication, only the features of the proposed method are presented here, which is interesting, original and effective, but this article lacks a broader scientific analysis.

- According to generally accepted rules, the presented method for defect detection as well as the research methodology should be precisely described in a scientific publication, so that the reader could potentially repeat the experiment. Meanwhile, the method and the methodology of experimental research are insufficiently described, it is not fully clear how the authors conducted the research.

- Conclusions are too general. It is necessary to compare results obtained with results from references.

- Conclusions should contain plans for further research.

- Old references should be replaced by more recent journal paper references.

- The formatting does not fully comply with the guidelines imposed by the journal.

- Generally, the language needs revision.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

The paper proposes to tackle the problem of a limited dataset for training machine learning tools by applying DCGAN and Lightweight CNN. The paper generally addresses well the method, and results and presents good readiness. I have just some comments for the authors.

1) I advise authors to include the following references and discuss them a bit in order to highlight the advances in light of those papers.

https://doi.org/10.1016/j.measurement.2020.108234

https://link.springer.com/article/10.1007/s10845-020-01579-w

https://link.springer.com/article/10.1007/s10489-021-02475-3

2) The authors could add the time and training time for the methods presented in the paper (comparison).

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 4 Report

The paper presents an approach to solving the defect classification problem when having a small sample size. Data regarding gear face defects were recorded using two cameras in controlled conditions. The method based on convolutional generative adversarial network and CNN was proposed. The proposed method was compared to VGG11 and VGG16 architecture and an approach based on traditional augmentation methods and a modified VGG11 network of input data. Results obtained on standard metrics indicate that the proposed method achieves better results than existing methods.

After reading the paper, I am left with a few concerns regarding the used methods.

In section 2.1.2. when training DCGAN, it is stated that random oversampling was used to increase data set size. Firstly, random oversampling is a technique that is commonly used to solve to problem of an unbalanced dataset but the specific dataset is balanced (by design), and secondly, it is not clear how will duplication of existing images help GAN to learn the diverse distribution of images. Could traditional data augmentation be used on input data for training GAN? 

In section 2.2., an explanation of how hyperparameters of proposed CNN such as a number of neurons in dense layers (and channels in Conv. layers) were selected, is missing? Which selection method was used (e.g. grid search, random search, Bayesian optimization, etc.)? Also, the introduction of dropout layers is motivated by the goal of reducing network overfitting, but what was the motivation behind the introduction of leaky ReLU instead of ReLU activation? These explanations should be included in the paper. 

In Figure 7, it would be helpful to plot training accuracy and training loss together with validation accuracy and validation loss to show if the proposed model (and also vgg11 and vgg16) is not overfitting because of the small dataset size. 

In section 3.4 it was mentioned that VGG11 and VGG16 were trained on data generated by DCGAN and augmented with geometric transformations. Were weights of these two models randomly initialized, or were these models initialized with pre-trained weights obtained on some larger dataset such as ImageNet, and then only the final layers of models were fine-tuned with the created gear defect dataset? These explanations should be added to the paper.

 

 

 

 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Thank you for your responses to comments and corrections to the manuscript. I recommend the manuscript for publication in the Applied Sciences journal.

Author Response

请参阅附件。

Author Response File: Author Response.docx

Reviewer 4 Report

 

Line (350 - 351) Page (12 - 13).

"Using the cross-entropy loss function and the weights of all network models are updated by random initialization ..."

I think that weights were randomly initialized, and not randomly updated. It is necessary to correct this in the final version of the paper.

Other modifications to the paper are adequate. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx