A Survey of Vision-Based Methods for Surface Defects’ Detection and Classification in Steel Products

Round 1

Reviewer 1 Report

In this paper, the detection and classification methods of steel surface defects are reviewed. However, there are still the following problems：

1. Table 1: It seems that the literature [245] should be classified in the category of machine learning or deep learning rather than statistical models. The authors are advised to double-check.

2. Line 534-535: The forward and back propagation represent the different flow directions of the data, and the neural network architecture seems to be the same for both, suggesting the authors consider the rigor of this description again.

3. Line 660-664: The method proposed in the literature [46] seems to use only unsupervised methods for pre-training to obtain feature extraction ability when using the autoencoder, and the final classification function is obtained by training a softmax with the supervised learning method. The following is a description of DAN in literature [46] "DAN with two auto-encoders (hidden layers) was constructed to classify NEU surface defects of the hot-rolled steel strip images. First, the two hidden layers were trained individually in an unsupervised method using auto-encoders. Then a final softmax layer was trained in a supervised method, and the layers were joined together to form the proposed DAN as in Figure 6.". The author discussed the unsupervised defect classification method in this part, so it is not very reasonable to refer to the literature [46] here, and further consideration is suggested.

4.  Line 664-668: The method proposed in the literature  [188] and [189] seems to be only used for defect detection rather than defect classification, so the author is advised to verify it again.

5.   Line 710-712: It seems appropriate to refer to Figure 8 instead of Figure 3 here.

6.   It is suggested to add a summary of steel surface defect detection methods based on self-supervised methods in this paper.

Comments for author File: Comments.pdf

Authors are advised to check the entire paper for minor problems in the English language

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

This article reviews vision-based steel surface defect detection and classification methods. The authors surveyed more than 200 relevant studies and evaluated state-of-the-art algorithms for the detection and classification of steel surface defects. In the process of comparison, their advantages and disadvantages are emphasized, as well as the evaluation results. This article is a comprehensive review article, but there are still some problems that need to be improved, as follows:

1. In Figure 5, "Methods based Learning" should be written as " Machine Learning based Methods" or just "Machine Learning". Otherwise, ambiguity will result, and it is recommended to make corrections to ensure the accuracy of the chart content.

2. The authors divide the categories of classification methods into: supervised methods，semi- supervised methods,  and unsupervised methods. Please consider carefully whether this classification is appropriate as it also applies to detection methods.

3. The picture in the article is not clear enough, better to provide a clear version. Also, carefully review image citation for compliance with relevant norms and compliance.

4. The text in most of the tables in this article is not neat enough and it is recommended to re-simplify the formatting.

5. "Future work" is written simply and generally. Authors should combine the existing article framework to refine clearer research directions for methodological issues.

6. Figure 8 and Figure 9 , it is recommended to use charts with a unified format, such as line charts or bar charts. It is important to keep the display format consistent.

7. In Defects Detection Methods Categories, there is a statement called "Model based methods". It may be misleading to the reader because there are also a large number of models in " machine learning based methods ". I don’t think it’s appropriate to express it this way.

The quality of English language is good.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

In this paper, The authors conducted an in-depth investigation into existing methods for detecting and classifying surface defects in steel, discussed the latest developments in this field, investigated the progress made in automated steel surface detection, and expressed their views on subsequent research in this field. The following comments are listed below for reference:

1. "Fig.4" in line 74 is not highlighted.

2. Suggest authors to beautify the lines in Figures 3 and 5.

3. There are symbol error in formula 3,  please check carefully

4. Suggest the authors mentioning “Conclusions and future directions” in the abstract

5. The review does not give an in-depth survey of unsupervised defect detection methods based on deep learning

no

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 4 Report

This article is a comprehensive review of the visual basic methods for the detection and classification of surface defects in steel products. The following are several problems outlined:

1. The sequence of the literature is unreasonable and somewhat chaotic.

2. Figure 2 is cited in the introduction section, but the figure is placed in the second section? The same applies to Figure 4, please check all figures and tables in the text.

3. On line 196, the annotation of Formula 5 should be written flush left, please check the annotations of all formulas in the text.

On line 227, 'Choi et al. citeb57 proposed', it seems that there is an error in the citation of the literature.

The resolution of Figures 6 and 7 is low, please replace them with high-resolution images.

The model accuracy in the table should maintain the same precision, for example, all retain two decimal places.

4. I noticed that the article mentioned the abbreviations of professional methods, which should be spelled out when they are mentioned for the first time.

5. On line 711, 'As shown in Figure 3', should this refer to Figure 8?

6. Conclusion: At present, when there is sufficient data, after data enhancement, model fine-tuning and other means, the model accuracy is usually not very low. How to improve the model recognition accuracy of small batch data sets is still a problem.

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The article has been revised as suggested, and authors are advised to double check for minor grammatical errors before publication.