Content Estimation of Foreign Fibers in Cotton Based on Deep Learning
Round 1
Reviewer 1 Report
Minor:
1. Lines 18-20 could be re-written to be more understandable and correct.
2. This is continuation of the research in [13], but it is not well emphasized.
3. 281-281 Error message should be in "". Is ".8" correct?
4. Lines 288-290: You missed to used index (subscript) for T0, P0, P1.
5. PA is actually PCC (percentage of correct classifications)?
6. Line 353: (128/9195, ms) - delete comma.
Major:
1. Lines 170-176. It looks like very expensive system, which is suitable for use in laboratory, not for massive industrial applications. Have you performed cost-benefit analysis? Is it possible that cost of the system, maintenance, and repairs cost more than it is reasonable to implement?
2. Scientific contribution is not well presented. Application contribution is obvious. However, this journal is scientific.
Author Response
Please see the attachment.
We appreciate the constructive suggestions from the two reviewers, which help us improve the manuscript. During the revision, we have carefully addressed each of the comments raised by the reviewers. We have also provided point-by-point responses to the reviewers’ comments. All the changes can be recognized by the revision tracks in the revised manuscript.
Responses to Reviewer #1
- Lines 18-20 could be re-written to be more understandable and correct.
Response: Thanks for reminding us. It is our fault to describe it as not clear in our paper. We changed the sentence “Finally, the test results on a prototype show that the root mean square error (RMSE) between the estimated size and the actual size is less than 4% relative to the original data. ” into “Finally, the test results show that the relative error between the estimated size and the actual size is less than 4%.”. The modified sentence can show the result of our experiment more concisely.
- This is continuation of the research in [13], but it is not well emphasized.
Response: Thanks for reminding us. In the revised manuscript, the index of reference [13] is changed into [17].
Our article is indeed follow-up research of reference [17], including hardware structure, test machine, etc. Therefore, in the revised manuscript, we have made some brief citations and explanations for the parts related to reference [17]. The relevant modifications are as follows:
(1). In the introduction of the principle of foreign fiber cleaning machine, reference [17] is added (Line 161).
(2). Added the sentence in Lines 315-317:
Moreover, because the encoder part of this paper uses the same network structure as the classifier in reference [17], the encoder parameters can use the classifier parameters as the pre-trained weight to speed up the training process.
(3). Lines 260-263: Modified the sentence from “The dataset is composed of the original image and the corresponding mask image. This paper uses a combination of semi-manual marking and full manual marking.” Into “The dataset consists of the original image and the corresponding mask image. Among the dataset, the original images are from the classified image dataset collected in reference [17]. The combination of semi-manual annotation and full manual annotation is used to make corresponding mask images.”
The above changes are mainly from the principle of the machine, the source of the dataset, and the source of the network structure. It is emphasized that this paper is follow-up research of reference [17].
- 281-281 Error message should be in "". Is ".8" correct?
Response: Thanks for reminding us. “” is not correct, we changed it into “Figure 8” and updated it to Line 318 in the revised manuscript.
- Lines 288-290: You missed to used index (subscript) for T0, P0, P1.
Response: Thanks for reminding us. We are sorry for the mistake like this. We have changed them into “T0, P0, P1” and updated them to Lines 324-326 in the revised manuscript.
- PA is actually PCC (percentage of correct classifications)?
Response: Thanks for the question. It is our fault to describe it as not clear in our paper. PA is actually PCC. They are used to calculate the proportion of correct classification in all results, whether it is categories (classification) and pixels (segmentation).
Major:
1. Lines 170-176. It looks like very expensive system, which is suitable for use in laboratory, not for massive industrial applications. Have you performed cost-benefit analysis? Is it possible that cost of the system, maintenance, and repairs cost more than it is reasonable to implement?
Response: Thanks for the question. We are sorry that we didn't explain it clearly. The key part of the foreign fiber cleaning machine is the processing subsystem module and software. The mechanical design uses a sheet metal method which is a low cost. The main function of the mechanical is to fix the optical parts and foreign fiber removal mechanisms. For the processing subsystem, the core processing board uses a low-cost embedded DSP + FPGA solution (about $400, as described in reference [13]), and the AI accelerator uses NVIDIA's Jetson TX2 module (about $599). The camera also uses the image acquisition system which is applied in other projects of our team. The cost is about $100, so the overall cost is controllable (about $3200). This machine has been commercially used in many cotton mills in China, so we have the opportunity to collect pictures of various production lines. Our paper is also based on the data research of these images. The method introduced by this paper is to add some software to the original equipment without increasing the cost of hardware. At present, this system has been used in many cotton mills in China. The foreign fiber content evaluation reports are regularly generated for the reference of manufacturers and introduced into the yarn quality management system.
- Scientific contribution is not well presented. Application contribution is obvious. However, this journal is scientific.
Response: Thanks for the question. Yes, this paper focuses on the application and has less contribution to science. However, we will publish the current dataset, hoping to attract more researchers to participate in the related research of foreign fiber image processing algorithms. At the same time, we find that there is a strong mathematical relationship between yarn quality and algorithm estimation data. We hope to propose an algorithm or mathematical model to better describe the matching between the two in the future. Recently, we also found that 3DCNN is more accurate in size or volume estimation, which is also a research direction in the future.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper applies image segmentation and classification for identifying foreign fibers during cotton processing. Overall an interesting engineering approach with practical value as it is used to retrofit and enhance a fiber cleaning machine.
There are shortcomings in the paper that need to be carefully addressed. First of all, there are several passages in the manuscript that need correction/rephrasing (e.g. larger sensing domain -> larger receptive field, marking -> annotation, "it is not easy to fall into the local minimum value" needs to be rephrased, "so the machine has misjudgment" needs to be rephrased, "waster paper" -> "waste paper" etc.). Thus, my suggestion is to carefully proofread the manuscript and resolve these issues. There is also a "Reference source not found" error and font size appears to change in certain passages.
On the technical side, further details shall be provided on the generated images, they appear to be some format of greyscale. Also, it is not clear how many images the authors use for training/validation/testing and whether they use transfer learning (i.e. using pre-trained network weights) or train the proposed networks from scratch. Further, my recommendation is to make the relative dataset public to facilitate replication and comparison with their experiments. Moreover, Figure 5 needs to be better explained and all the displayed components named as markup on the image and in the caption. Also in Equation (7) T1, P1 are not explained in the text. In the introduction, I suggest to include some relevant deep learning applications for quality control in industry such as,
"Fault Detection and Isolation in Industrial Processes Using Deep Learning Approaches", TII 2019 by Iqbal et al.
"Fault Diagnosis in Microelectronics Attachment Via Deep Learning Analysis of 3-D Laser Scans" TIE 2020 by Dimitriou et al.
"A deep learning framework for simulation and defect prediction applied in microelectronics" SMPT 2020 by Dimitriou et al.
"A Data-Flow Oriented Deep Ensemble Learning Method for Real-Time Surface Defect Inspection" TIM 2020 by Liu et al.
All in all, if the authors believe that the issues above can be convincingly addressed my recommendation for the paper is a major revision.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
PA should be changed to PCC to avoid confusions.
Reviewer 2 Report
Authors have considerably improved the manuscript according to the comments. Therefore, I suggest the publication of the paper.
