Surface Defects Recognition of Wheel Hub Based on Improved Faster R-CNN
Round 1
Reviewer 1 Report
Paper needs editorial proof reading for its language. Paper content should be enhanced for its scientific merit.
Author Response
Point 1: Paper needs editorial proof reading for its language.
Response 1: Thank you very much for your valuable advice, I have already invited a person whose native language is English for editing the paper. You can refer to the paper for detail.
Point 2: Paper content should be enhanced for its scientific merit.
Response 2:
First of all, thank you very much for your valuable advice. I have found a postdoctoral fellow whose native language is English to help us polish the language. For detailed modification, please refer to the part in red font in the article.
About enhancing its scientific merit. I think your suggestion is meaningful. After careful consideration, I think the scientific merit should be an exploration on the generalization of convolutional neural network for target recognition. Please refer to the details in line 55-59 in red.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper presents an improved of the Faster R-CNN network i n order to detect 10 method wheel hub defects (scratch, oil pollution, block, grinning). Results are compared with the ones obtained with the R-CNN and YOLOv3.
Please make the following updates and corrections:
-compare the obtained results with results from articles described in related work
-please explain abbreviation before use (eg. RPN)
-Fig. 5 is not referred in the text
-line 163 is not finalised
-correct end of line 203: t 。
-in Table 3 add unit measurement for all columns
-correct line 346: CNN、
Author Response
Response 1:
First of all, thank you very much for your advice. According to your suggestions, I made the following statistics for the references [13-25] as the following table.
| Reference | Objects | Methods | Results | |
| [13] | wheel hubs defects (air hole, shrinkage porosity, shrinkage cavity and crack) | improved peak location algorithm + BP | 96% | |
| [14] | PCB surface defect | similarity measure approach | 100% | |
| [15] | Steel surface | Gabor filter | 94% | |
| [16] | steel surface | Singular Value Decomposition | 89.3% | |
| [17] | Steel surface defects | CNN | 99.29% | |
| [18] | Ship detection | Rotated region based CNN (RR-CNN) | 75.7% | |
[19] | Solar cell surface defect | multispectral convolutional neural network | 88.41% | |
| [20] | Fasteners | DCNN | 92.78% | |
| [21] | Object detection (PASCALVOC2007) | semantic segmentation | 90.34% | |
| [22] | Object detection (PASCALVOC2007) | Relief R-CNN | 53.8% | |
| [23] | Object detection (PASCALVOC2007) | Fast R-CNN | 66.9% | |
| [24] | Object detection (PASCALVOC2007) | Faster R-CNN | 69.6% | |
| [25] | fabric defects | Faster R-CNN | 83.6% | |
| Ours | wheel hub defects | Faster R-CNN | 72.9% | |
Based on the information in the table, we find that the image datasets are different, and methods are different too, so the results are not comparable. But they follow a common thread of research into the recognition of multiple classes of objects. I hope you will be satisfied with my explanation. Analysis of experimental results has been done at 5.4, please refer to table 2, table 3, line 333-334.
Point 2: Please explain abbreviation before use (eg. RPN)
Response 2: First of all, thank you very much for your valuable advice. I have already checked the full text and made the corresponding explanation for the abbreviated part, as well as the reference notes. For the detailed information, please refer to line 105, 117, 179 ,427-430 in red font. The subsequent references have also been adjusted accordingly.
Point 3: Fig. 5 is not referred in the text.
Response 3: I have marked it in the previous paragraph of Figure 5. Please refer to the line 165 red font.
Point 4: line 163 is not finalised
Response 4: Thank you for reminding me. I'm very sorry, this is my mistake. I have already finished the modification. Please refer to the line 177-178 red font for details.
Point 5: correct end of line 203: t
Response 5: I have already corrected it. Please refer to the line 219 red font for details.
Point 6: in Table 3 add unit measurement for all columns.
Response 6: I have already corrected it. Please refer to the Table 3, line 334 red font for details.
Point 7: -correct line 346: CNN、
Response 7: I have already corrected it. Please refer to the line 345 red font for details.
Author Response File: Author Response.pdf
Reviewer 3 Report
This work presents an automated solution for recognizing wheel factory defects using machine learning and deep learning techniques.
The approach is interesting and the proposed method for correctly communicating different types of networks seems effective; results are interesting although improvements when compared with state-of-art detectors don't seem so exciting.
Minor issues:
Abstract: avoid writing personal considerations (we hope...)
English must be improved, especially the use of verbal forms, articles (the) and punctuation.
Examples (but not the only):
· line 53: This layout of this paper
· line 56: In Section 5, the training process was completed
· line 71: it was proposed in 2006,
· line 95: Time consuming of object proposal is the bottleneck
· line 286: a four-step training strategy were adopted
Figure 1 must enlarged and better described in the text.
R-CNN, RPN, ecc. are acronyms that must be explained.
line 41-42: what means “verification sets” ? From Table1 is not clear if the training and validation are confused or are two distinct phases.
line 155-157: rewrite this sentence
line 166: (reg-layer, cls-layer) are not pointed in Fig.5; it must be clarified whether these names are known in the literature, or they belong to the chosen implementation or are original adaptations proposed by the authors.
line 200 troth -> truth
line 198: from text it is not clear if the loss function has been proposed and edited by authors or comes from well-known literature.
line 215: better using “Stochastic” (SGD)
line 228-245: this paragraph is too obscure and must be clarified. Please make a sort of scheme to explain step by step this process.
line 247: what mean "resource" ?
Author Response
Point 1: Abstract: avoid writing personal considerations (we hope...)
Response 1: Thank you very much for your valuable comments, I have adjusted this sentence, please refer to line13-15 in red.
Point 2: line 53: This layout of this paper
Response 2: I have already change layout to structure, please refer to line 61 in red.
Point 3: line 56: In Section 5, the training process was completed.
Response 3: I have already changed, please refer to line 64-65 in red.
Point 4: line 71: it was proposed in 2006,
Response 4: I have already changed the sentence, please refer to line 85-86 in red.
Point 5: line 95: Time consuming of object proposal is the bottleneck
Response 5: I have already changed two prepositions, please refer to line 109 in red.
Point 6: line 286: a four-step training strategy were adopted
Response 6: I have already adjusted to was, please refer to line 286 in red.
Point 7: Figure 1 must enlarged and better described in the text.
Response 7: I have already enlarged the Figure.1. In addition, further explanation is be done for Figure 1(a) ,(b). please refer to line 27-28, 32 in red.
Point 8: R-CNN, RPN, ecc. are acronyms that must be explained.
Response 8: I have already done the explanation for R-CNN, RPN. please refer to line 105, 117 in red.
Point 9: line 41-42: what means “verification sets” ? From Table1 is not clear if the training and validation are confused or are two distinct phases.
Response 9: I have already changed verification sets to validation set. Please refer to line 159. And I'd like to explain Table1 as follows: Training set is used to set up the model, Validation set is used to tune the model, and Testing set is used to evaluate the performance of the model. They are different stages, first training, then verification, last testing. However, the role of the former two is to establish the optimal model, while the latter is to evaluate the model, so the dataset is used to put training and verification sets together. I have indicated the percentages of the three(training, validation, test) in the table1, Please refer to Table 1, line 161.
Point 10: -line 155-157: rewrite this sentence.
Response 10: After careful consideration, I think the statement in this part is a little verbose, so I deleted it, which will not affect the integrity of the whole article. Please refer to line 173.
Point 11: line 166: (reg-layer, cls-layer) are not pointed in Fig.5; it must be clarified whether these names are known in the literature, or they belong to the chosen implementation or are original adaptations proposed by the authors.
Response 11: I have already changed the Figure.5 including “reg-layer, cls-layer”, and also add the reference. please refer to Figure.5, line 181in red.
Point 12: line 200 troth -> truth
Response 12: I have already changed to truth, please refer to line 216 in red.
Point 13: line 198: from text it is not clear if the loss function has been proposed and edited by authors or comes from well-known literature.
Response 13: I have already added the reference. please refer to line 214 in red.
Point 14: line 215: better using “Stochastic” (SGD)
Response 14: Considering that the training process of 4.2 has included the content of 4.1.3, this part is deleted. please refer to line 229.
Point 15: line 228-245: this paragraph is too obscure and must be clarified. Please make a sort of scheme to explain step by step this process.
Response 15: I have already changed the paragraph. Please refer to Line 230-245.
Point 16: line 247: what mean "resource" ?
Response 16: I'm sorry, they are the repeated references, I forgot to replace, I have modified, please refer to line 247 in red.
Point 17: English editing
Response 17: Thank you for your advice, I have already done the English editing by a person native English, who is the last author of the paper, please refer to the paper for detail.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Thank you for your response. All my comments were addressed.
Reviewer 3 Report
I believe that the changes improved the quality of the paper.
