A Study on Defect Detection in Organic Light-Emitting Diode Cells Using Optimal Deep Learning

Round 1

Reviewer 1 Report

I have reviewed the manuscript thoroughly and I have some comments and suggestion to make it better which are as follows:

1. Please provide concrete evidence that underlines the superiority of the VGG-16 algorithm.

2. The abstract should offer a more specific delineation of what the author aims to assert.

3. Kindly introduce the VGG-19 model and elucidate the rationale behind opting for VGG-16 over it.

4. The 'val' value is set at 0.95, which appears to lack precise justification. Is there any reference or basis for this particular setting?

5. In Figures 3 and 4, only specific points are highlighted with red circles. Please provide a detailed justification for this choice.

Author Response

Thank you for your review. We have made revisions based on your comments and provided answers to the questions.

1. Please provide concrete evidence that underlines the superiority of the VGG-16 algorithm.

The architecture of VGG-16 consists of convolutional, pooling, and fully connected layers. VGG-16 extensively used small-sized filters in its pooling layers to extract diverse features from data. The convolution layers use 3 × 3 filters with padding, while the pooling layers use 2 × 2 filters. The fully connected layers consist of three layers. After convolution and pooling, the feature maps are flattened into one-dimensional vectors for image classification. Un-like convolutional layers, where neurons are connected only to a local input region, fully connected layers connect every neuron to all the neurons in the previous layer.

2. The abstract should offer a more specific delineation of what the author aims to assert.

I have added information in the Abstract section to elaborate on the research process by explaining the adjustment of image file formats before applying deep learning algorithms

3. Kindly introduce the VGG-19 model and elucidate the rationale behind opting for VGG-16 over it.

The VGG-19 algorithm is characterized by a neural network structure composed of 19 layers. These layers are predominantly made up of 3x3 filters and 2x2 max-pooling layers. The reason for utilizing the VGG-16 algorithm in this study is that, compared to VGG-19, it has a fewer number of layers while maintaining a similar architecture. This makes it a streamlined algorithm. As a result, it requires fewer parameters and computational resources, leading to more efficient computation time and better performance. This is why the VGG-16 algorithm was employed in the study.

4. The 'val' value is set at 0.95, which appears to lack precise justification. Is there any reference or basis for this particular setting?

The reason for setting the 'val' variable to 0.95 is grounded in statistical theory. In the field of statistics, the concept of p-value is related to hypothesis testing, indicating how plausible the results of a particular hypothesis are. Therefore, instead of using a binary pass or nonpass threshold of 0 or 1, an image classification threshold of 0.95 was chosen based on statistical theory, to determine the classification.

5. In Figures 3 and 4, only specific points are highlighted with red circles. Please provide a detailed justification for this choice.

I have added red circles to Figures 3 and 4. The reason for applying the dark spot images from the initial stage of speck formation and after 10,000 hours of formation is based on the paper titled 'A Study on OLED Cell Simulation and Detection Phases Based on the A2G Algorithm for Artificial Intelligence Application'.

Reviewer 2 Report

This paper presents a study on defect detection in OLED using deep learning.

The studied topic is interesting and also meaningful.

The paper still has some major problems.

The authors are suggested to revise the paper given the following comments.

The experimental validation part still falls short.

More closely-related methods are suggested to be compared, especially those specifically designed for the same purposes.

More closely-related databases are suggested to be used.

As discussed in some surveys and studies, e.g., Perceptual image quality assessment: a survey; Screen content quality assessment: overview, benchmark, and beyond; Unified blind quality assessment of compressed natural, graphic, and screen content images; A metric for light field reconstruction, compression, and display quality evaluation, quality of the image is an important aspect of various intelligent systems, including defect detection systems.

High-quality images are important for the successful usage of these intelligent systems, while low-quality media may degrade the performance of these systems.

The authors may give some discussions on this aspect as well as the above-mentioned works.

The authors only detect defects in pure images with dark or gray images.

As discussed in ‘Blind quality assessment based on pseudo-reference image’ and ‘Blind image quality estimation via distortion aggravation’, artifacts existing in images of various contents are more difficult to detect or evaluate. The authors are suggested to give some discussions on these aspects and the above works, and give some discussions on whether it is possible to detect the defects of OLED when it’s showing complex images..

As described in the literatures (for example, Objective quality evaluation of dehazed images, Quality evaluation of image dehazing methods using synthetic hazy images), enhancement of images will improve the quality and efficiency of the following intelligent systems. The authors are suggested to give some discussions on these aspects and the above works, and give some discussions on whether incorporating image enhancement could improve the defect detection system.

Visual attention can be of great value in various intelligent systems. As described in some studies, e.g., A multimodal saliency model for videos with high audio-visual correspondence; Fixation prediction through multimodal analysis; Study of subjective and objective quality assessment of audio-visual signals, incorporating attention could further improve various intelligent systems. The authors are suggested to give some discussions on these aspects and the above works, and give some discussions on whether incorporating visual attention/saliency could improve the monitor system.

The whole paper is suggested to be double-checked to remove all possible issues.

None

Author Response

To the reviewer,

Thank you for your feedback on our work.

We have elaborated further on the validation part. However, please understand that due to the nature of our research concept being computational simulations and for security reasons, extensive additions are challenging. Given that our basis is machine learning, we have utilized the CNN algorithms, which hold significant advantages in image comparisons. For simulations based on physical laws, we employed the specialized physical law software, COMSOL.
We have detailed the specific images used for our image analysis. The original image size was 760x609, but for the purpose of machine learning, we resized it to 224x224. The reason our images are in black and white is that they were converted to grayscale. Our research solely considered the expansibility according to humidity, eliminating the need for additional colors. The actual research employed more complex images, involving superposition and boundaries. Thus, we have updated our paper to reflect these images. High-resolution processing of the images is certainly feasible. When the data was created, it was at 760x609, and depending on the settings, we can obtain much higher resolution images. Our focus was strictly on the software aspect of the research. For the research to be successfully implemented, there is still a need for highly precise cameras and expert initial settings.

We genuinely appreciate your thoughtful review.

Reviewer 3 Report

The manuscript titled presents an intriguing study focusing on the application of an optimal deep-learning algorithm to detect defects in OLED cells. The research aims to enhance the yield of OLEDs by reducing defective products through efficient defect detection. The paper effectively outlines the study's objectives, methodologies, and findings.

The study's focus on applying an optimal deep-learning algorithm to detect defects in OLED cells presents a novel and valuable contribution to the field. The manuscript effectively addresses the challenges of defect detection in the context of OLED display manufacturing. The practical implications of your research, such as enhancing product yield and reducing production costs, highlight its relevance to both industry and technology advancement. Your selection and thorough evaluation of deep learning algorithms, particularly the VGG-16 architecture, demonstrate a well-considered approach to solving the problem. The use of a virtual dataset generated through the Solver program is a creative solution to obtaining the necessary data for your study, and it showcases your resourcefulness in addressing real-world challenges. The experimental results, including accuracy and loss values, provide clear evidence of the effectiveness of your proposed defect detection model. By prioritizing energy efficiency in AI-IoT systems and conducting practical tests, your work contributes to the creation of sustainable and resilient implementations in the field. Overall, the manuscript showcases a strong methodology, insightful analysis, and valuable practical implications, making it a significant and worthwhile contribution to the literature on OLED defect detection and deep learning applications.

Here are some suggestions on manuscript revision:

Expand the introduction to provide more context about the significance of defect detection in OLED cells. Explain how improved defect detection can impact the OLED industry, including aspects like cost reduction, product quality, and technology advancement. Besides, introduce more relevant work on the defect detection using deep learning methods, such as: DOI: 10.1007/s00170-022-10335-8, and DOI: 10.3390/a16020095.

Provide information about the dataset used for training, validation, and testing. Mention the specifics of the images (e.g., resolution, color format) and how they were split between the three phases.  If relevant, consider sharing the hyperparameters used during training (e.g., learning rate, batch size) to provide transparency and allow for reproducibility.

In the conclusion, emphasize how the successful implementation of the VGG-16 algorithm in the defect detection model can contribute to advancements in the OLED display manufacturing technology and how this can impact the broader electronics industry.

The quality of English writing in the manuscript is highly commendable. The sentences are well-structured, clear, and effectively convey the technical aspects of the research.

Author Response

Thank you for your review. We have made revisions based on your comments and provided answers to the questions.

1. Expand the introduction to provide more context about the significance of defect detection in OLED cells. Explain how improved defect detection can impact the OLED industry, including aspects like cost reduction, product quality, and technology advancement. Besides, introduce more relevant work on the defect detection using deep learning methods, such as: DOI: 10.1007/s00170-022-10335-8, and DOI: 10.3390/a16020095.

Detecting defects within OLED cells is crucial for several reasons. Firstly, it aims to pre-vent the production of faulty items during the manufacturing process. Defective products can compromise quality and performance, potentially leading to unreliable products for consumers. Therefore, accurately identifying and addressing defects within OLED cells is of paramount importance. Furthermore, reducing the occurrence of product disposal due to defects is economically significant within the OLED industry. Identifying defects early and taking corrective actions can prevent the need for product disposal or rework, thereby resulting in cost savings. According to the study by Li, Wei, et al., utilizing artificial intelligence technology to automatically detect defects has been explored [3]. They developed and validated a network that could automatically identify defects, such as scratches, on the surface of products. This research emphasizes the potential for enhancing product quality and reducing the number of defective items through effective defect detection methods. Efficiently identifying and addressing cell defects can increase production yield. This, can contribute to the advancement of OLED display manufacturing technology and have positive effects on the OLED display market.

2. Provide information about the dataset used for training, validation, and testing. Mention the specifics of the images (e.g., resolution, color format) and how they were split between the three phases. If relevant, consider sharing the hyperparameters used during training (e.g., learning rate, batch size) to provide transparency and allow for reproducibility.

A preprocessing procedure was carried out on the 1300 training dataset images utilized during the training process. This step aimed to enhance the model's performance by making adjustments to the image data. The original image dimensions were resized from 760x609 to 300x300, and the bit value representing the color format was set to 24. Additionally, to promote diversity and improve generalization capability of the data, image augmentation was performed. This involves applying various transformations to the images to enhance the model's performance. To facilitate this, the image dimensions of the training dataset were set to 224x224, and a batch size of 32 was chosen to perform data augmentation using the generated image data. Following the same procedure, the training dataset images were also processed for the validation dataset. In this process, the image dimensions of the validation dataset were resized from 760x609 to 300x300, and the color format bit value was set to 24. Similarly, data augmentation was applied to the validation dataset to improve the model's performance and generalization capability. The image dimensions were adjusted to 224x224, and a batch size of 32 was employed for data augmentation. Consistent preprocessing steps were thus applied to the validation dataset images as well.

3. In the conclusion, emphasize how the successful implementation of the VGG-16 algorithm in the defect detection model can contribute to advancements in the OLED display manufacturing technology and how this can impact the broader electronics industry.

Based on the study, it was possible to identify cell defects by implementing a cell defect classification model using the VGG-16 algorithm. While the study did not utilize actual images of dark spots, it successfully simulated the process of dark spot formation using virtual images, which were driven by relevant physical phenomena. While the current OLED industry relies on human inspection for defect detection, utilizing artificial intelli-gence technology to identify cell defects could contribute to the advancement of display manufacturing techniques. This applicability extends not only to the OLED industry but also to manufacturing and inspection processes within the broader electronics sector.

Round 2

Reviewer 2 Report

The authors only give a paragraph of general responses to the reviewer's comments. Point-for-point responses are suggested to be given, for every specific comments. Moreover, the mentioned highly-relevant works are suggested to be discussed in the paper.

None

Author Response

We sincerely appreciate the reviewer's thoughtful feedback.

1. We have made additional enhancements to the experimental validation section. However, due to security constraints, we were unable to provide source references. Your understanding on this matter is highly appreciated.

2. We have included additional related studies to the paper.

3. Our study has been ultimately validated using real-world data. We have added this information to the content.

4. The original image resolutions have been included. For validation purposes, we resized the image dimensions to reduce the model's load, while the original high-resolution data can be used for simulation with good quality settings.

5. Regarding the validation model, we have successfully conducted analysis, including superposition and detection of defects at boundaries. We will provide updated figures to illustrate these points.

6. Similar to what was mentioned earlier, higher resolutions can be set for computationally simulated datasets. Moreover, your suggestions regarding generating better images with more variables have been taken into account for future research directions.

7. To apply our study's methodology effectively in real industrial environments, it's important to utilize high-quality cameras and involve expert initial setup. Consequently, we have noted your suggestions about improving intelligent systems and monitoring based on the discussed studies for future research.

We genuinely thank you for your detailed review. Our research team has incorporated your feedback to modify the paper as mentioned above.
Wishing you all the best in your future endeavors.

Round 3

Reviewer 2 Report

The responses do not correspond to the following comments given in the first round of review in a point-for-point way.

The experimental validation part still falls short.

More closely-related methods are suggested to be compared, especially those specifically designed for the same purposes.

More closely-related databases are suggested to be used.

As discussed in some surveys and studies, e.g., Perceptual image quality assessment: a survey; Screen content quality assessment: overview, benchmark, and beyond; Unified blind quality assessment of compressed natural, graphic, and screen content images; A metric for light field reconstruction, compression, and display quality evaluation, quality of the image is an important aspect of various intelligent systems, including defect detection systems.

High-quality images are important for the successful usage of these intelligent systems, while low-quality media may degrade the performance of these systems.

The authors may give some discussions on this aspect as well as the above-mentioned works.

The authors only detect defects in pure images with dark or gray images.

As discussed in ‘Blind quality assessment based on pseudo-reference image’ and ‘Blind image quality estimation via distortion aggravation’, artifacts existing in images of various contents are more difficult to detect or evaluate. The authors are suggested to give some discussions on these aspects and the above works, and give some discussions on whether it is possible to detect the defects of OLED when it’s showing complex images..

As described in the literatures (for example, Objective quality evaluation of dehazed images, Quality evaluation of image dehazing methods using synthetic hazy images), enhancement of images will improve the quality and efficiency of the following intelligent systems. The authors are suggested to give some discussions on these aspects and the above works, and give some discussions on whether incorporating image enhancement could improve the defect detection system.

Visual attention can be of great value in various intelligent systems. As described in some studies, e.g., A multimodal saliency model for videos with high audio-visual correspondence; Fixation prediction through multimodal analysis; Study of subjective and objective quality assessment of audio-visual signals, incorporating attention could further improve various intelligent systems. The authors are suggested to give some discussions on these aspects and the above works, and give some discussions on whether incorporating visual attention/saliency could improve the monitor system.

None

Author Response

We sincerely appreciate the reviewer's thoughtful feedback.

I have conducted performance evaluation of the artificial intelligence model, including validation and assessment with real data.

I have added similar research to the related research section. Furthermore, this paper, which is based on OLED virtual data, proposes a completely different approach based on AI, distinct from the conventional OLED AOI (Automatic Optical Inspection) method that has primarily focused on panel mura inspection. I did not include the comparison approach suggested by the reviewer, as it does not align with the core objective of this study, which is cell-based.

This paper was conducted based on OLED cell computational simulation data. Since this study aimed to identify defects in OLED cells over time, there are no closer data sources available beyond actual OLED cell data. In the conclusion section of this paper, a comparison with real data has been provided.

This paper is a methodological study focused solely on defect detection, without involving image processing. We believe that introducing additional image processing at the micrometer scale cell level could lead to unexpected defects. Additionally, the main focus of this paper is to explore how incredibly small pinholes will enlarge over time. Once they start growing, there may be overlapping areas, and we utilized a dataset that has been generalized (A²G) based on the work of Han Dong-hun and two others to address this issue. We aimed to determine whether, within a specified time frame (10,000 hours), the pinholes grew within an acceptable margin of error. If they remained within the allowed error (10,000 hours), they were considered non-defective; otherwise, they were classified as defective. This approach demonstrates that our method provides higher accuracy compared to visual inspection.

I have made changes to the figure to demonstrate the capability of judging superposition and the presence of specks at the edges. Additionally, I have provided the size of the computational simulation data. For security reasons, the actual dataset couldn't be included in the paper, and the real data was gray-scaled to green luminescence. 

The reason for this process is that it helps simplify algorithms and eliminates complexities related to computational requirements. It also facilitates easier learning for those new to image processing since grayscale reduces an image to its most essential pixels.

Furthermore, it's important to note that this study is solely focused on detecting defects in OLED cells over time and does not involve various types of content or videos.

As mentioned earlier, the goal of our research is to improve detection accuracy, and we do not focus on image processing beyond grayscale. The reason for this is that at the micro-scale level, introducing additional processing could potentially lead to unexpected false defects due to the small size of the images.

Indeed, we concur with the reviewer's perspective. However, our research is primarily aimed at developing a methodology for applying to a specific system, namely, defect assessment based on black spots in OLED cells, rather than striving for broader applicability across various systems. We have outlined the paper with this specific focus in mind. We hope that future research endeavors will explore various perspectives in line with our approaches.

We genuinely thank you for your detailed review. Our research team has incorporated your feedback to modify the paper as mentioned above.

Wishing you all the best in your future endeavors.

Author Response File: Author Response.pdf

Round 4

Reviewer 2 Report

None

None