Identification and Classification of Aluminum Scrap Grades Based on the Resnet18 Model

Round 1

Reviewer 1 Report

In this paper, the authors propose a method based on deep learning to achieve classification of different grades of non-ferrous aluminum according to surface properties. Using a deep learning algorithm, the authors compare the surface properties of RGB, HSV and LBP images of three aluminum blocks with grades 1060, 5052 and 6061. The topic is clearly interesting and current, but I have some reservations about the work.

-Figures 9. 10. 12. : There is a histogram in the pictures. A histogram represents data that breaks down a range of classes into columns along the horizontal x-axis, so I recommend drawing the graphs discretely rather than continuously.

-row 254 : In the mentioned line, you claim that the data distribution ratio is 8:2. Is that a error and did you mean 80:20? Please described the division of the data into sets. Why are you only using the training and validation set? On what data is the testing taking place?

-Figure 13. :The description for the picture does not include what is in the picture under point b)

-Figure 14. : The image is unclear and unreadable. For better readability, I recommend redesigning.

-Figure 15. : The graph shows that you achieved 100% on RGB images during the training process. Is it possible that you overtrained the network? My assumption is that when a network achieves such high results, it is overtrained on the given data. Please explain to me why you think the network is not overtrained and prove that the results are reliable.

-Figure 17. 18. : I understand that in the graphs you compare the results with different learning rates and different batch sizes, but it is not clear in the text or from the graph on which data this test was performed on RGB, HSV or LBP. Given the above-average accuracy, I'm assuming it's RGB data, but it's not clear in the text.

-At work, you only evaluate the results of the training process. I can't see the testing results anywhere. It is clearly necessary to supplement the results after testing the network and to compare them with other approaches of other authors.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

The authors have been working on Identification and classification of aluminum scrap grades based on the Resnet18 model. The research is interesting however the manuscript should have some modifications to be published.

 To make vision identification by camera, the samples should have any preparation?

 The authors should specify clearly what units are using in axes X and Y of the figures. brightness -Intensity a. u.?

 

On figure 9 the authors mentioned that “ grayscale images cannot achieve the classification of the three grades”. I think that in grayscale image is possible get the classification of three aluminum types. Due to in fig 9c appear another well defined peak at 190 brightnes

The authors should explain clearly how converted the RGB images to HSV images

 I recommend to change the title because the author are talking about Identification and classification of aluminum (they are implaying grouping or separating). In the paper all procedure is about identification the do not nothing about classification.

 to classify is to group objects or people with respect to a characteristic that they have in common

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors incorporated most of the comments, but I still have major reservations about the processing of the achieved results.

-Figure 15: From your correction, I rather perceive that you rewrote some results from the training process to the testing process. It does not make sense to have in one image in a) accuracy results from the testing process and in b) to have loss results from the training process. These results need to be clarified and divided. It is necessary to clearly divide the results into results that were acquired within the training process and results that were acquired within the testing process. With such a layout of the article, where you mix training and testing results, the article is very unclear.

-In their response, the authors offered me a graph where I can see the training and testing results, but this graph is nowhere to be found in this article. The article should clearly contain such a graph.

-I also find a discrepancy between the results, where the authors sent me a graph where they compare the results of the training process with the results of the testing process, but everywhere in the article I find these results as training results.

-For better clarity of the results, I recommend adding a confusion matrix to the work, which expresses the results of the testing.

Author Response

Thank you for your decision and constructive comments on my manuscript. We have carefully considered the suggestion of Reviewer and make some changes, We have tried our best to improve and made some changes in the manuscript.

Please see the attachment

Author Response File: Author Response.docx

Round 3

Reviewer 1 Report

The authors incorporated all my comments and I agree with the publication of the article, but I recommend an even smaller modification:

Figure.17.,18. : in both figures I would give separately the loss and accuracy results during the training process and the loss and accuracy functions separately during the testing process.

Author Response

Please see the attachment

Author Response File: Author Response.docx