An LCD Detection Method Based on the Simultaneous Automatic Generation of Samples and Masks Using Generative Adversarial Networks

Round 1

Reviewer 1 Report

Review in the attachment

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This manuscript proposed an LCD detection method based on the simultaneous automatic generation of samples and masks using generative adversarial networks. 

The manuscript is well-written and easy to follow. Figures have reasonable resolution and are understandable. 

The Authors might consider: 

1. Add some more papers to Literature that are newer and relevant. 

2. Include recognition rate of some other similar works in order to be comparable with the proposed. 

3. Consider if there are any other statistical measures that might be also interested to show with the recognition rate. 

4. Are there any disadvantages of the proposed method? 

Quality of English Language is fine. 

Author Response

Question: Add some more papers to Literature that are newer and relevant.

Answer: We have organized the literature cited throughout the article, updated it with the latest research advances in this direction, and completely revised the citation format of the literature.

Question: Include recognition rate of some other similar works in order to be comparable with the proposed.

Answer: We have added some new references and set up “2. Related work”. in this section, the methods we are about to compare are clearly listed.

Question: Consider if there are any other statistical measures that might be also interested to show with the recognition rate.

Answer: We have added Table II to make the data from our experimental results easier to understand and refer to.

Question: Are there any disadvantages of the proposed method?

Answer: We provide a self-commentary on the limitations of our proposed method in Conclusions and plan to address these known issues in the future.

Reviewer 3 Report

Manuscript electronics-2759798: An LCD Detection Method Based on the Simultaneous Automatic Generation of Samples and Masks Using Generative Adversarial Networks

Dear authors, your text is almost clear, and your work is good and beautiful.

I get confused in Lines 59 – 64, You explain here “The Mask Region Convolution Neural Network (R-CNN) method offers higher accuracy compared to the aforementioned deep learning-based object detection methods. It also possesses the advantage of simultaneous classification, localization, and instance segmentation. Therefore, by improving and applying the Mask R-CNN method to the detection of micro-defects in LCDs, it becomes possible to identify defect categories and segment defect shapes.”

Having concluded that, you seemingly deviate from your conclusion to start another line of argument:

In Line 64. you say ‘However, the utilization of the aforementioned deep learning methods faces challenges in collecting a substantial number of defect samples for training and testing, as well as the complex and time-consuming task of annotating defect samples with masks.’

But you discarded the deep learning method in favor of R-CNN and now you do not? If I am wrong sorry, otherwise: could you clarify this?

Could you say a word if the method by Park et al. [1] is (un)suitable for your purpose?

Another confusing remark is about the lack of segmentation by wavelet methods. In Line 269, you conclude ‘the wavelet method cannot efficiently segment images of defective LCD surfaces’. Indeed Figure 9 exhibits this. But wavelets quadtrees are a rigid segmentation of the image, see e.g. [2,3]. To me, you dispense too easily with this wavelet segmentation property. Do you have additional material to show that you correctly applied the wavelet method as exhibited in Fig. 9? In the extra material of the paper, I did not find experimental material.

Concluding: your paper is almost perfect. Please amend the confusing sentences.

Typos, observations, and critical remarks

1.       Line 30, and 52, misses a blank before the [ bracket.

2.       Line 152 164, For me unclear/inconsistent notation: is the effect part and its superposition correctly displayed? You have in line 153 a double overstrike on x, while in formula (6) and line 164, you use a double circumflex on the x.

3.       Line 269, ‘Figure. 9’ has an erroneous dot.

References

1.        Park, S.H.; Tjolleng, A.; Chang, J.; Cha, M.; Park, J.; Jung, K. Detecting and Localizing Dents on Vehicle Bodies Using Region-Based Convolutional Neural Network. Appl. Sci. 2020, 10, 1–9.

2.        Park, J.S.; Lee, S.H. Automatic mura detection for display film using mask filtering in wavelet transform. IEICE Trans. Inf. Syst. 2015, E98D, 737–740.

3.        Kunal, K.; Prasad, R.A.; Xavier, M.J.; Arun, J. Mura defect detection in lcd. Acad. Mark. Stud. J. 2021, 25, 1–10.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The paper introduces a compelling solution to the challenges associated with applying deep learning methods to detect micro defects on low-contrast LCD surfaces. The primary issues addressed are the imbalance in the samples dataset and the complexity of annotating and acquiring target image masks. The proposed method leverages deep generative network models and introduces a unique strategy for sample and mask auto-generation.

The conclusion drawn from the presented results succinctly encapsulates the effectiveness of the CycleGAN framework for defect detection. CycleGAN employs a combination of adversarial loss and periodic consistency loss to generate output images, measuring adversarial loss to ensure the alignment of the generated distribution with the target distribution. The process involves two primary types of loss: adversarial losses and cycle consistency losses.

The approach is strong, but could benefit from additional literature:

1. Huang AA, Huang SY: Increasing transparency in machine learning through bootstrap simulation and shapely additive explanations. PLoS One. 2023, 18:e0281922. 10.1371/journal.pone.0281922

2. Fang, F., Li, L., Zhu, H., & Lim, J. H. (2020). Combining Faster R-CNN and Model-Driven Clustering for Elongated Object Detection. IEEE transactions on image processing : a publication of the IEEE Signal Processing Society, 29(1), 2052–2065. https://doi.org/10.1109/TIP.2019.2947792

3. Huang AA, Huang SY: Computation of the distribution of model accuracy statistics in machine learning: Comparison between analytically derived distributions and simulation-based methods. Health Sci Rep. 2023, 6:e1214. 10.1002/hsr2.1214

[2]

4. Kalantar, R., Messiou, C., Winfield, J. M., Renn, A., Latifoltojar, A., Downey, K., Sohaib, A., Lalondrelle, S., Koh, D. M., & Blackledge, M. D. (2021). CT-Based Pelvic T₁-Weighted MR Image Synthesis Using UNet, UNet++ and Cycle-Consistent Generative Adversarial Network (Cycle-GAN). Frontiers in oncology, 11, 665807. https://doi.org/10.3389/fonc.2021.665807

Author Response

We referenced those articles you suggested and selected the ones that supported and helped us in our research for citation and elaborated on their specific value.

Reviewer 5 Report

The research introduces a method that leverages deep generative network models to address challenges in detecting micro defects on low-contrast LCD surfaces. This approach effectively addresses issues related to imbalanced sample datasets and the complex task of annotating image masks. The experimental results highlight the method's success not only in generating training samples but also in automatically producing accurate image masks, demonstrating its potential to enhance detection accuracy. However, while the concept is well-presented, the research's novelty is considered average as similar approaches have been presented in comparable applications.

The authors employed the scientific method to support their claims, but it is suggested that they provide a more detailed explanation of the methodology in a separate section.

Although the introduction contains some related works, the research lacks a dedicated section for a comprehensive review of related work and an explanation of the research's significance. This information should be presented in a separate section.

The conclusions need improvement, emphasizing the research results and suggesting possible directions for future research.

There is room for improvement in the English language, such as in the abstract where it reads, "Experimental results 14 shows...". Additionally, the format of references needs consistency; for instance, reference [1,2] uses a comma, while [6-8] uses a hyphen. Uniformity in reference formatting should be maintained.

Author Response

Question: The authors employed the scientific method to support their claims, but it is suggested that they provide a more detailed explanation of the methodology in a separate section.

Answer: We have rearranged the sections of the article to make it look clearer by placing the relevant research within the 2.Related work.

Question: Although the introduction contains some related works, the research lacks a dedicated section for a comprehensive review of related work and an explanation of the research's significance. This information should be presented in a separate section.

Answer: We summarize the research we have done and detail the value of the research we have done in the concluding section of the Related work.

The relevant answer could be found in the answer to the previous question.

Question: The conclusions need improvement, emphasizing the research results and suggesting possible directions for future research.

Answer: We have recapitulated our study so that our final conclusions are more prominently visible. Also, we provide a self-commentary on the limitations of our proposed method in Conclusions and plan to address these known issues in the future.

Comments on the Quality of English Language

Question: There is room for improvement in the English language, such as in the abstract where it reads, "Experimental results 14 shows...".

Answer: We do have a relevant description in the abstract, i.e., “Experimental results shows the effectiveness of our proposed method, as it allows for the simultaneous generation of LCD image samples and their corresponding image masks. ”

However, our scrutiny did not reveal a character associated with '14'.

As it should be, we have rewritten this paragraph to avoid other possible syntax errors.

Question: Additionally, the format of references needs consistency; for instance, reference [1,2] uses a comma, while [6-8] uses a hyphen. Uniformity in reference formatting should be maintained.

Answer: According to the article template requirements：“References should be numbered in order of appearance and indicated by a numeral or numerals in square brackets—e.g., [1] or [2,3], or [4-6]. ”

We cite both Literature 6, 7, and 8 in “In recent years, deep learning methods have gained prominence and have gradually been employed for LCD defect detection [6-8]”.

Reviewer 6 Report

The Section 1, 2 and 3 are well written and don't need major changes.

At Section 4 - Table 1: please explain how the recognition rate is determined and calculated.

Also, describe the 2 sets (groups) of data - samples, how they were divided into train/test data, etc.

Section 5 (Conclusions) - should be extended to include a Discussion section about the pros/cons of the proposed solution and how it can be further improved.

Author Response

The Section 1, 2 and 3 are well written and don't need major changes.

--We appreciate you recognizing this part of our work.

At Section 4 - Table 1: please explain how the recognition rate is determined and calculated. Also, describe the 2 sets (groups) of data - samples, how they were divided into train/test data, etc.

--An explanation of how the recognition rate is calculated is given in the text before we present it.

--We add a description of the sample sizes used in the two groups of experiments and how they were distributed when describing the arrangement of the two groups of experiments.

Section 5 - should be extended to include a Discussion section about the pros/cons of the proposed solution and how it can be further improved.

--We have rewritten portions of our conclusions and present a broader perspective on the advantages, disadvantages, and significance of our work.

Round 2

Reviewer 1 Report

All my comments have been corrected and included in the new version of the Manuscript. I have no comments on its current shape, and in my opinion, the article should be printed as it is.

Author Response

We appreciate your recognition of our work!