Strip Steel Surface Defects Classification Based on Generative Adversarial Network and Attention Mechanism

Round 1

Reviewer 1 Report

This paper presents a solid study and a credible proposal for classification of defects in metals processing by enriching data form the use of GAM. The paper as commented is complex, but even, if it fits well in the area of data-science, it is not so evident for specialists in material science and manufacturing processes.

 

One of the main contributions of the authors is replacing the Jansen-Shanon distance by the one of Wassertein, however, such a choice needs a more illustrative presentation and motivation.

 

The rest of the proposals quite technical, sometimes the proposals are poorly introduced, justified and discussed, for instance section 2.2. The same in sections 2.4. The authors should try to describe before interacting all the technical aspects, why the different bricks, how to interconnect them.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Some parts are clearly written, in other parts, English could be improved to ensure a good understanding.

In the introduction and main text, the authors refer to "defects". It would be helpful for readers to know if what kind of defects, if the results depend on the kind of defect, their size, etc.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This manuscript shows that the surface defects on strip steel can be detected with high efficiency using the multi-SE-ResNet34 model. In addition, data augmentation through WGAN was used to increase the accuracy of classification, and it was shown that the classification criteria of the Multi-SE-ResNet34 model can be indirectly checked using the class activation map (CAM) function.

 

The composition and content of the article are overall excellent, and the results also adequately explained. In addition, it was well demonstrated that the developed model was effective through comparative analysis with several existing models.

Therefore, it is judged that the paper is sufficient to be published in the Metals journal.

 

Some of the things that need improvement are as follows.

 

1. (lines 87, 88): Regardless of abstract, when the abbreviations ‘GAN’ and ‘WGAN’ first appear in the main text, the full name must be written together.

 

2. (line 132): In the “1 * 1 * C/r”, is ‘r’ the reduction ratio in the formula? It should be written along with the formula.

 

3. (line 140): Because the abbreviation CAM first appeared, the full name must be written together.

 

4. In the “2.3. Feature visualization” section, a more detailed explanation of Grad-CAM (along with formulas) would be helpful to readers.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf