Organic Light Emitting Diode (OLED) displays have revolutionized the visual experience with their superior color reproduction, contrast ratios, and flexibility. However, maintaining the high-quality standards expected of OLED screens is a complex endeavor, primarily due to the occurrence of Mura defects. These defects manifest as irregularities in brightness or color, detracting from the overall display quality and user experience.

Traditional inspection methods, heavily reliant on human visual assessment, are often inconsistent and inefficient, especially when dealing with low-contrast images where Mura defects lack distinct edges. The subjective nature of human inspection introduces variability, and the sheer volume of displays produced necessitates a more scalable solution.

Enter region-based machine learning – a sophisticated approach that combines the precision of machine learning algorithms with the specificity of region-based analysis. This methodology not only automates the detection process but also enhances accuracy by focusing on localized regions within the display, effectively identifying subtle defects that might elude traditional inspection techniques.

The term Mura originates from the Japanese word for “unevenness” or “blemish,” aptly describing the visual inconsistencies these defects introduce. In the context of OLED displays, Mura defects are characterized by non-uniform luminance or chrominance, appearing as spots, lines, or regions with noticeable brightness or color deviations.

These defects can be broadly categorized into:

• Spot Mura: Localized spots with varying brightness or color.
• Line Mura: Linear imperfections, either horizontal or vertical, disrupting the uniformity of the display.
• Region Mura: Larger areas exhibiting uneven brightness or color, affecting a significant portion of the display.

The detection of Mura defects is particularly challenging due to their subtle nature and the low-contrast images in which they appear. Traditional image processing techniques often struggle to identify these defects, necessitating the adoption of more advanced, machine learning-based approaches.

Machine learning has emerged as a transformative tool in various industries, and its application in defect detection is no exception. By training algorithms on vast datasets of defective and non-defective images, machine learning models can learn to identify patterns and anomalies indicative of defects.

In the realm of OLED Mura defect detection, machine learning offers several advantages:

• Automation: Reduces reliance on manual inspection, increasing throughput and consistency.
• Precision: Enhances the ability to detect subtle defects that may be imperceptible to the human eye.
• Scalability: Easily adapts to different display sizes and resolutions without extensive reconfiguration.

However, the success of machine learning models heavily depends on the quality and diversity of the training data, as well as the robustness of the algorithms employed.

The region-based machine learning approach focuses on analyzing specific regions within an image to detect defects, rather than evaluating the image as a whole. This method is particularly effective for identifying localized imperfections like Mura defects.

A seminal study by Janghwan Lee introduced a comprehensive system for detecting white spot and horizontal line Mura defects using a region-based machine learning model. The system comprises four core components:

1. Mura Candidate Detector: Utilizes computer vision and image processing algorithms to identify potential defect locations within the display.
2. Candidate Region Extractor: Generates image patches centered around the identified candidate locations for detailed analysis.
3. Feature Extractor: Extracts relevant statistical and textural features from each image patch to characterize the potential defects.
4. Classifier: Employs Support Vector Machines (SVM) to classify the extracted features as defective or non-defective regions.

This structured approach allows for precise localization and classification of Mura defects, significantly improving detection accuracy.

The effectiveness of any machine learning model is intrinsically linked to the quality of the dataset used for training. In the context of Mura defect detection, several challenges arise:

• Noisy Labels: Inconsistencies in labeling, often due to subjective human judgment, can introduce noise, adversely affecting model performance.
• Inconsistent Data: Variations in data acquisition conditions, such as lighting and camera settings, can lead to inconsistencies that complicate the training process.

To address these challenges, Lee proposed the creation of a “golden dataset,” meticulously curated and relabeled by multiple researchers to ensure consistency and accuracy. This dataset serves as a reliable foundation for training robust machine learning models.

Adversarial training is a technique used to improve the robustness of machine learning models by exposing them to challenging or misrepresented data during the training process. In the case of Mura defect detection, adversarial training involves iteratively refining the model by:

One of the key challenges in defect detection is balancing precision and recall. While a high recall rate ensures that most defects are detected, it may come at the cost of lower precision, leading to false positives. Conversely, optimizing for high precision might result in missed defects.

To address this, Lee implemented a threshold adjustment mechanism in the SVM classifier, leveraging confidence scores derived from classification results. By dynamically adjusting the decision threshold, the system can be fine-tuned to prioritize recall (minimizing missed defects) or precision (reducing false alarms) based on operational requirements

Applying the region-based machine learning model to real OLED panel images from a manufacturing line yielded impressive results. Using a golden dataset and adversarial training techniques, the model achieved:

• 100% recall rate: Ensuring that no defects were overlooked.
• Over 90% precision: Minimizing false positives and ensuring accurate defect classification.
• Improved inspection efficiency: Reducing the reliance on manual inspections, streamlining quality control, and enhancing production throughput.

Comparing with Traditional Methods

The Future of Machine Learning in Display Manufacturing

The success of region-based machine learning in OLED defect detection signals a broader shift toward AI-driven quality control in display manufacturing. Future advancements may include:

• Deep Learning Integration: Enhancing feature extraction with convolutional neural networks (CNNs) for more robust defect detection.
• Real-Time Processing: Optimizing algorithms for faster defect detection in high-speed production lines.
• Edge AI Deployment: Implementing lightweight AI models directly on inspection hardware for on-the-fly analysis.
• Multi-Modal Inspection Systems: Combining machine vision with AI-driven decision-making for holistic defect classification.