Next Article in Journal
Evidence Preservation in Digital Forensics: An Approach Using Blockchain and LSTM-Based Steganography
Next Article in Special Issue
Nighttime Pothole Detection: A Benchmark
Previous Article in Journal
Dynamic Partitioning of Graphs Based on Multivariate Blood Glucose Data—A Graph Neural Network Model for Diabetes Prediction
 
 
Review
Peer-Review Record

Toward Automated Fabric Defect Detection: A Survey of Recent Computer Vision Approaches

Electronics 2024, 13(18), 3728; https://doi.org/10.3390/electronics13183728
by Rui Carrilho 1,*, Ehsan Yaghoubi 2,*, José Lindo 3, Kailash Hambarde 1 and Hugo Proença 1
Reviewer 1: Anonymous
Reviewer 2: Anonymous
Reviewer 3: Anonymous
Electronics 2024, 13(18), 3728; https://doi.org/10.3390/electronics13183728
Submission received: 21 August 2024 / Revised: 14 September 2024 / Accepted: 18 September 2024 / Published: 20 September 2024
(This article belongs to the Special Issue New Trends in AI-Assisted Computer Vision)

Round 1

Reviewer 1 Report (New Reviewer)

Comments and Suggestions for Authors

The authors have reviewed recent advancements in automated fabric defect detection technologies based on computer vision over the past five years, with a focus on deep learning methods and a discussion on future research directions, including the potential value of traditional methods and strategies to address existing challenges.

 

Before the manuscript can be accepted, there are several aspects that require attention, and thus minor revisions are recommended.

 

1. Although the article provides an overview of various methods, it fails to systematically compare the advantages and disadvantages of these methods. The description of different methods mainly remains at a qualitative level, lacking quantitative comparisons such as accuracy, speed, and sensitivity to different types of fabric defects. Such comparisons are crucial for researchers to choose the technology that best suits their needs. It is suggested that the authors add one or more comprehensive evaluation indices to compare the performance of different methods.

 

2. The paper lists a variety of detection methods, including statistical, spectral, and model-based approaches, but the discussion on their applicability and limitations in actual industrial environments is insufficient. Particularly, the feasibility of deep learning models, which require substantial computational resources, is not clearly discussed in the context of resource-constrained factory environments. Moreover, given the diversity and complexity of fabrics in production settings, the performance of various methods can vary significantly across different types of fabrics and defect detection tasks.

 

3. Overall, the language of the paper is clear, but some sections contain grammatical errors and imprecise expressions, such as improper use of technical terminology that could lead to misunderstandings. It is recommended that the authors carefully proofread the manuscript, especially the technical descriptions, to ensure the accuracy and consistency of professional terms.

 

4.Lines 130 and 323 in the manuscript mention '...is seen in Figure ??', but no corresponding figures are provided in the actual paper. Is this a formatting issue or a file problem? Additionally, the paper uses few figures. It is advised to include diagrams, flowcharts, or architecture diagrams when introducing typical fabric defect detection technologies such as deep learning methods to help readers more intuitively understand the working principles and practical operations of these methods.

Comments on the Quality of English Language

Overall, the language of the paper is clear, but some sections contain grammatical errors and imprecise expressions, such as improper use of technical terminology that could lead to misunderstandings. It is recommended that the authors carefully proofread the manuscript, especially the technical descriptions, to ensure the accuracy and consistency of professional terms.

Author Response

Greetings,

Thank you for your feedback. We hereby respond to your your points individually:

Although the article provides an overview of various methods, it fails to systematically compare the advantages and disadvantages of these methods. The description of different methods mainly remains at a qualitative level, lacking quantitative comparisons such as accuracy, speed, and sensitivity to different types of fabric defects. Such comparisons are crucial for researchers to choose the technology that best suits their needs. It is suggested that the authors add one or more comprehensive evaluation indices to compare the performance of different methods.

In response to your feedback, we have added a table that summarizes the main advantages and disadvantages of each approach. We hope this meets your expectations.

The paper lists a variety of detection methods, including statistical, spectral, and model-based approaches, but the discussion on their applicability and limitations in actual industrial environments is insufficient. Particularly, the feasibility of deep learning models, which require substantial computational resources, is not clearly discussed in the context of resource-constrained factory environments. Moreover, given the diversity and complexity of fabrics in production settings, the performance of various methods can vary significantly across different types of fabrics and defect detection tasks.

You make a very good point, however, the reason we don't cover these topics is precisely because they are barely covered in the literature. Whenever these topics are approached, it is in the context of edge computing, and few other studies bother evaluating their applicability in factory settings. In consideration of your advice, we have added a paragraph in Section 6 mentioning this.

Overall, the language of the paper is clear, but some sections contain grammatical errors and imprecise expressions, such as improper use of technical terminology that could lead to misunderstandings. It is recommended that the authors carefully proofread the manuscript, especially the technical descriptions, to ensure the accuracy and consistency of professional terms.

We have indeed proofread the article once more, and corrected errors as they were found. Thank you for your feedback. If you have further concrete cases to point out, we welcome further feedback.

4. Lines 130 and 323 in the manuscript mention '...is seen in Figure ??', but no corresponding figures are provided in the actual paper. Is this a formatting issue or a file problem? Additionally, the paper uses few figures. It is advised to include diagrams, flowcharts, or architecture diagrams when introducing typical fabric defect detection technologies such as deep learning methods to help readers more intuitively understand the working principles and practical operations of these methods.

Thank you for the warning, we in fact had a figure that we deleted from the final version, and had accidentally left the reference. We have deleted it. Furthermore, regarding the lack of figures, the paper used to have more, mostly coming from other works, but they were found upon review to be superfluous and add little in terms of content, so per my advisor's suggestion, we deleted many. The process of securing copyright for those images was also needlessly taxing so as to make it preferable to remove them. As we have added a new table per your suggestion, and the document was quite dense to begin with, we believe we have achieved a good ratio of textual to graphical content. If you believe otherwise, feel free to let us know in the future.

We thank you again for you efforts, and hope we have improved the article for your standards.

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

In this article was summarized the most important advancements of the last 5 years, mostly on machine learning-based approaches. Was described the most promising avenues of research in the future. The article is a review paper. The article contains many references to literature on defect research in the textile industry. A small number of tables have their reference in the text. The work seems to be written clearly and understandably.

However, there are a few things that need to be made:

1. The authors could prepare a summary in the form of a table specifying what types of research and what defects can be detected by particular methods.

2. Is it possible to group the types of defects occurring in the textile industry?

 

Once the changes have been made, the article may be considered for publication.

Author Response

Greetings,

Thank you for your efforts in reviewing our paper. We respond to each of your criticisms point-by-point:

  1. The authors could prepare a summary in the form of a table specifying what types of research and what defects can be detected by particular methods.

I'm afraid such a table is beyond our capabilities, given how the problem is formulated in the literature and approached in industry conditions - in recent times, most of the detection methods proposed are focusing on one-class classification, wherein detecting the defect type is irrelevant, and what matters is detecting the presence or absence of a defect at all. Additionally, even when the different defect types are considered, there are no methods which are designed to detect specific types of defects. Ad to the best of our knowledge, it hasn't at all been studied whether some methods are better suited for certain types of defects as compared to other methods. 

As such, we believe we are not able to do such a table, as we lack information for that in the literature, and that situation does not seem likely to change. 

  1. Is it possible to group the types of defects occurring in the textile industry?

Potentially, but as I explained in the above answer, this does not seem to be a promising research  direction, and there appears to be no real need for it. Furthermore, given how in industrial conditions, each factory/manufacturer decides by their own criteria what constitutes a defect and what doesn't, it is hard to reach a universal standard for what should or shouldn't count as a defect, and no list provided would be universal - or possibly, of any use. We discuss this over at the Discus

We thank you for your feedback, and hope our responses have been to your satisfaction.

 

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

This paper provides a survey on fabric defect detection in recent years.

1. It would be good to include a discussion on the problem formulation; currently, both image classification and object detection are mentioned.

2. The recent transformer-based approaches (if any) need to be discussed.

3. The existing dataset is summarized in Table 2. More descriptions would be helpful, such as whether they are available for the public or not, # samples per defect type, etc.

4. Sections 6 and 7 might be revised in a more structural manner. For example, how the challenges in Section 6 could be addressed in Section 7.

Comments on the Quality of English Language

NA

Author Response

Greetings, 

We thank you for your efforts and for the feedback you have provided. We hereby respond to every point you made:

  1. It would be good to include a discussion on the problem formulation; currently, both image classification and object detection are mentioned.

We do not quite understand what is expected here - we present an overview of the fundamental problem in the Introduction. Given how the problem can be divided into either detecting the presence of a defect, or classifying said defect, which are generally up to manufacturer/industrial criteria, we do not find it worthwhile, or strictly possible to define the problem more formally.

2. The recent transformer-based approaches (if any) need to be discussed.

At the time of writing this article (which has been in revision for much longer than expected!), there were few any such approaches to be seen. While more are starting to show, they are vastly in the minority, at least in the area of fabric defect detection. We discuss this in the "Future trends and research directions" section. 

3. The existing dataset is summarized in Table 2. More descriptions would be helpful, such as whether they are available for the public or not, # samples per defect type, etc.

We have improved the table as per your directions. We hope it is to your liking.

4. Sections 6 and 7 might be revised in a more structural manner. For example, how the challenges in Section 6 could be addressed in Section 7.

We believe the two sections already have such a structural following. If you could define this more clearly, we would be glad to address this in a future revision.

We thank you again for your efforts, and hope the paper is now to your satisfaction. 

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.


Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors provide a comprehensive overview of various approaches used in fabric defect detection, with a particular focus on machine learning-based methods, especially deep learning. The paper is well-structured and provides a detailed taxonomy of different methods along with their applications and limitations.

The paper is fit for publication provided the following corrections can be incorporated: 

  1. Dataset Standardization and Availability:

    • The lack of standardized, publicly available datasets is a significant barrier to progress in fabric defect detection research. Can the authors share more comprehensive datasets that could improve the reproducibility and comparability of future studies and the impact of this paper. 
  2. Integration of Traditional and Deep Learning Approaches:

    • I see that deep learning dominates the field, however traditional methods are often less computationally demanding and could be effective when computational resources are limited. A hybrid approach combining the robust feature extraction capabilities of deep learning with the simplicity and speed of traditional methods could yield improved efficiency and effectiveness. Are such methods present in literature ? If so please list them ? If not then is this a valid future direction ?
  3. Algorithm Efficiency and Real-World Application:

    • Most of the approaches discussed are evaluated in controlled conditions. There is a need for research on algorithms that perform well in real-world, industrial environments where lighting, speed of fabric movement, and other factors can affect performance. Developing lightweight models that can run on edge devices directly in the production line would be beneficial. The authors could include this in their evaluation and/or discussion ?
  4. Improved Handling of Imbalanced Data:

    • The paper mentions the challenge of imbalanced data due to the rarity of some types of defects. Techniques such as synthetic data generation or advanced sampling methods could be further explored to enhance the training of models under these conditions.
  5. Robustness to Variations in Fabric and Defect Types:

    • The effectiveness of defect detection systems across different types of fabrics and defects varies significantly. Is there Research into more adaptive algorithms that can learn from a small number of samples and generalize across different textiles could make the technology more universally applicable ? Is yes then list it, if not mention in future directions ?
  6. Reproducibility and Code Availability:

    • A common issue in academic research is the lack of reproducibility. The community could benefit from more open sharing of code bases and detailed experimental setups, allowing researchers to build directly on each other's work. Not sure if this is work commenting on ?
  7. Exploration of New Neural Network Architectures:

    • The field could explore the latest developments in neural network architectures beyond the commonly used CNNs and GANs, such as Transformers and Capsule Networks, which might offer new ways to handle the complexities of fabric defect detection. Please mention in future research ?

 

Author Response

Greetings,

I greatly appreciate your feedback, and thank you for your efforts. I have made sure to incorporate your criticisms, and now address each point-by-point:

  • Dataset Standardization and Availability:

In Section 5, most notably in Table 2, we summarize the most widely used datasets in this area. In response to this comment, I have added a reference to the MVTec dataset, widely considered a benchmark in unsupervised detection problems. That said, as we point out in the same section, the lack of standardized datasets for this specific problem is one of the problems of this area - if there were more comprehensive datasets, there would be no need for the subsection.

  • Integration of Traditional and Deep Learning Approaches:

We approach this topic in the third paragraph of the "Challenges and limitations" section. Overall, these approaches sometimes intermingle, and we pointed out such cases in the "Traditional methods" section as they occurred. We also point this out in the Future Research section. As no works appear to exist which address this directly, I do not think I can further flesh this part out more. 

  • Algorithm Efficiency and Real-World Application

Albeit superficially, we did in fact include this in our second-to-last paragraph in Section 7. We also discuss why we didn't flesh out discussion of these methods previously, as articles describing such methods usually do not introduce much in the way of novel approaches to this problem, and don't particularly fit into the taxonomy we had previously defined. 

  • Improved Handling of Imbalanced Data

We briefly mention synthetic data in the "Generative models-based approaches" section. It is another weakness we spotted in the literature, as this approach is often explored in other areas, yet barely so in this one. I have added a paragraph regarding this in the Future Research section. 

  • Robustness to Variations in Fabric and Defect Types:

We did list such approaches - they fall under generative model-based approaches. In either case, I have added a mention of this in Section 7.

  • Reproducibility and Code Availability:

We do mention this, in the last paragraph of Section 6 "Challenges and limitations", albeit superficially. We also allude to a general lack of reproducibility in other sections, wherever the topic arises.

  • Exploration of New Neural Network Architectures:

Capsule Networks are briefly mentioned in the "CNN-based approaches" section. As for transformers, we tangentially mentioned them with DETR. That said, at the time we started our search, no articles following a transformers-based approach could be found. Even now, a search for the topic on Google Scholar yields only 3 relevant results. We did not consider it relevant to pool so few articles together when the vast majority of approaches are of the two previously referred types. Nevertheless, I have edited the paragraph in which these are mentioned.

I believe this addresses all of your criticisms. Please let me know if you have further criticisms, and I shall be sure to incorporate them at my earliest convenience.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have proposed the following manuscript: "Towards Automated Fabric Defect Detection: A Survey of Recent Computer Vision Approaches" in which they summarized the most important advancements of the last 5 years, and focus mostly on machine learning-based approaches. 

The introduction part is quite well described and presents fabric defect detection analysis for a total of 12 scientific articles. The taxonomy of fabric defect detection is quite well described for the the two main categories: classical machine learning methods and deep learning methods Their main purpose was to focus on the learning-based approaches. Also, the Challenges and limitations chapter is well presented, in respect with reproductibility and the lack of standards in this area.

Future trends and research directions and Conclusion chapter presents very well the results of their study. deep learning-based approaches are far more taxing on computational resources than traditional approaches. A new trend that seems to be emerging to tackle this problem consists of using edge devices to perform defect detection in factory settings. While such works are more practical than theoretical in nature, they are very suited to the problem at hand, and research into them is likely to continue, which is a desirable outcome.

Considering the fact that the manuscript is well structured and presented, as well as the above mentioned I propose that the article should go for publication.

 

Author Response

Greetings,

I greatly thank you for your efforts towards reviewing this paper, and appreciate your feedback.

Reviewer 3 Report

Comments and Suggestions for Authors

The paper summarizes the last 5 years studies of defect detection and directs the future directions of it. It is a significant review. However there are some suggestions.

1. Figure 1 and 2 quality can be improved

2. line 98-103, please add corresponding examples. 

3. line 367, please double check here for format issue

4. part 6 challenges, please include some references

5. part 7 future trends, can you provide other trends except the usage of other updated CNNs.

 

Author Response

Greetings, 

Thank you for your efforts and feedback. I answer each point-by-point:

  • Figure 1 and 2 quality can be improved

Those pictures are taken directly from other articles (copyright issues already handled), as stated in the captions and surrounding text. So we sadly can't improve their quality.

  • line 98-103, please add corresponding examples

Each of the categories is explained in the following detailed subsections, with wide variety of examples. As such, I did not think it was relevant to introduce them in the bulletpoints, to prevent redundancy

  • line 367, please double check here for format issue

Thank you for the warning, there was indeed a formatting issue and it has been resolved

  • part 6 challenges, please include some references

That section mostly contains original research from us, from trends observed reading the articles. Nevertheless, I have added references as requested

  • part 7 future trends, can you provide other trends except the usage of other updated CNNs.

Every paragraph in that section contains other trends.  Nevertheless, I have since revised it with more still. 

I thank you again for your feedback and hope these improvements have lived up to your expectations.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Responses look good. 

I have no other suggestions. 

Back to TopTop