Tire Bubble Defect Detection Using Incremental Learning
Round 1
Reviewer 1 Report
1. The architecture diagram in Figure 5 is the original version of yolov3, so please mark the source of the cited literature.
2. Freezing the Backbone section for incremental training seems to be a good approach. Please explain why choosing the Backbone section instead of some layers of the Neck section. What are the advantages of such an approach?
3. The division of the training set, validation set, and test set is vaguely explained in the paper. From the sample size described in the article, it appears that the training and test sets are not independent. If this is the case, the rest of the discussion becomes meaningless. Please describe the division of the training set, validation set, and test set in the overall sample and the ratio. (This is an essential part and needs to be explained in detail. Please revise the results and conclusions if the results are changed because of sample classification.)
4. It is mentioned in the paper that labor costs can be reduced. However, when the accuracy and precision are less than 100%, even with 98% accuracy, it is still necessary to manually detect all the results and thus determine what the AI misidentifies. Therefore need to understand how the method proposed in this paper reduces labor costs. Please explain in further detail.
Author Response
- The architecture diagram in Figure 5 is the original version of yolov3, so please mark the source of the cited literature.
Author response:
We sincerely appreciate reviewer’s advice. We have revised the paper and marked the source of the cited literature as reference 8 in the revised manuscript.
8. Joseph Redmon; Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018.”
- Freezing the Backbone section for incremental training seems to be a good approach. Please explain why choosing the Backbone section instead of some layers of the Neck section. What are the advantages of such an approach?
Author response:
Thanks for reviewer’s comments. Since the backbone network mainly extracts low-level features such as edges, colors, and textures of images. The neck layer is mainly to process and enhance the low-level features extracted by the backbone, so that the model can learn features suitable for tire bubbles. To reduce the computational load of feature extraction and efficiently learn the image features of the field, we freeze the backbone layer and retrain the neck layer in incremental learning to strengthen the bubble features of the field. This is mentioned in the last paragraph of page 5 in revised version.
- The division of the training set, validation set, and test set is vaguely explained in the paper. From the sample size described in the article, it appears that the training and test sets are not independent. If this is the case, the rest of the discussion becomes meaningless. Please describe the division of the training set, validation set, and test set in the overall sample and the ratio. (This is an essential part and needs to be explained in detail. Please revise the results and conclusions if the results are changed because of sample classification.)
Author response:
The training and the testing dataset are independent. One image may include more than one bubbles. The training dataset contains 2,450 images including 1,763 tire trend images and 687 tire sidewall images with 3,116 bubble defects. The test dataset contains 2,315 images consists of 1,310 tread images and 1005 sidewall images. We have revised the sentences in Section 3.1 on page 6 and 7.
- It is mentioned in the paper that labor costs can be reduced. However, when the accuracy and precision are less than 100%, even with 98% accuracy, it is still necessary to manually detect all the results and thus determine what the AI misidentifies. Therefore, need to understand how the method proposed in this paper reduces labor costs. Please explain in further detail.
Author response:
Using Yolov3 model can improve the detection rate to 98.56%, reduce the detection speed and provide better detection rates than previous studies for the hard-to-detect bubbles. Labors need to detect all the tire bubbles manually. With the help of artificial intelligence, operators only need to confirm the bubble defects detected by machine, which can reduce the workload of labors. Manufacturers do not need to recruit a lot of labor to detect bubble defects one by one. We have also revised paragraph 3 of the Introduction Section.
Reviewer 2 Report
Tire Bubble Defect Detection Using Incremental Learning.
The proposed work is very interesting and useful, Its good author depicted satisfactory results still minor check and few modifications are required.
1. Motivation of the study and challenges in this research domain could be added in the Introduction part.
2. Related work section is very short so add the latest work in the related work and update it thoroughly, also add these work:-doi: 10.1155/2022/2789760, doi: 10.1007/978-981-16-3690-5_136
3. Why YOLO Technique is used ? Explain its advantages over other existing Method.
Author Response
- The proposed work is very interesting and useful. Its good author depicted satisfactory results still minor check and few modifications are required.
Author response:
Thanks for your valuable advice. We have addressed all comments and responses point by point in this revised manuscript.
- Motivation of the study and challenges in this research domain could be added in the Introduction part.
Author response:
Thanks for your valuable suggestion. We have revised the motivation and challenges of this study in the 2nd, 3rd, and 4th paragraph of Introduction Section in this revised version.
“The tire bubbles are difficult to detect by the naked eye since they are internal defects….
With the help of artificial intelligence, operators only need to confirm the bubble defects detected by machine, which can reduce the workload of labors. Manufacturers do not need to recruit a lot of labor to detect bubble defects one by one….
we use an object detection method to improve the bubble defect detection speed and detection rate. In addition, to address the situation where the model developed in the laboratory is actually introduced into the field with a reduced defect detection rate. We propose an incremental YOLO (You Only Look Once) architecture to add on-site misidentified samples to model training to improve the model's flaw detection accuracy on the actual site.”
- Related work section is very short so add the latest work in the related work and update it thoroughly, also add these work:-doi: 10.1155/2022/2789760, doi: 10.1007/978-981-16-3690-5_136
Author response:
We have added these two articles (doi: 10.1155/2022/2789760 and doi: 10.1007/978-981-16-3690-5_136) to the references and numbered [1] and [2] in this revised manuscript.
- Why YOLO Technique is used? Explain its advantages over other existing Method.
Author response:
The YOLO machine learning algorithm uses features learned by a deep convolutional neural network to detect objects in real time. YOLO has the advantage of being much faster than other networks and still maintaining high accuracy. The is also added in the paragraph 2 page 2.
Round 2
Reviewer 1 Report
Thanks to the author for the detailed answers to my questions. I have also confirmed the additional content in the revised version.
