SmartMuraDetection: Small-Sample DL Approach for OLED Defect Detection

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Feb 15 2024

In an article recently published in the journal PLOS One, researchers proposed SmartMuraDetection, an organic light emitting diode (OLED) detection method based on small sample deep learning (DL).

Background

Different defects, such as mura defects, line defects, and point defects, occur during the OLED production process due to multiple factors, such as the presence of foreign matters within the workshop environment, differences in film thickness, and glass substrate defects. These frequently occurring defects affect the products' delivery yield and generate a significant amount of waste, resulting in higher processing costs.

The introduction of automated inspection machines in the form of assembly lines in place of manual inspection to realize a low-over inspection, low missed inspection, closed-loop, and quantifiable machine vision inspection system can improve the production line automation level and overall quality inspection level.

Although the detection of line and point defects is quantifiable and easy, realizing automatic detection of mura defects in OLED panel image quality inspection is difficult owing to their diverse contour and shape, low contrast, and wide variety. Specifically, quantifying defects becomes difficult due to low contrast.

Additionally, the defects cannot be clearly distinguished from the background due to fuzzy defect boundaries. Conventional methods for mura defect detection cannot effectively distinguish defect features from the gray background level owing to the close similarity between the color of the mura defect regions and backgrounds.

The proposed DL-based approach

In this study, researchers proposed a small sample-based DL OLED detection model, designated as SmartMuraDetection, to overcome the uneven background interference. The objective was to solve the surface detection problems of low accuracy and automate the late-model display panel production process.

A gradient edge linear stretching method was employed for image preprocessing to simultaneously enhance the contrast of background and defects. The proposed small-sample DL method was evaluated on a dataset to determine its accuracy and efficiency compared to other methods.

The small sample DL-based SmartMuraDetection OLED detection model contained a SEMUMaxMin quantization module, a small-scale target detection TinyDetection model, a defect feature fusion module, and a gradient boundary enhancement algorithm module to address four key issues, including the difficulty to quantify the output characteristics, small defect scale, inadequate negative sample size, and weak defect characteristics.

The gradient boundary enhancement algorithm module solved the weak defect feature problem, enhanced the image's effective contrast, ensured defect feature maximization and no effective information loss, while the defect feature fusion module solved the insufficient negative sample and unbalanced sample problem by embedding an source image's defect region into the target image for generating a new negative sample image.

Additionally, the TinyDetection model was proposed for small-scale target detection to precisely detect defects. Based on the Yolo v3 network, this model could increase the shallow network's residual blocks and properly reduce the deep network's residual blocks to ensure that the network can fully extract the details of small targets/small-scale target defects.

After the extraction of mura defects using the target detection model, the mura was quantitatively evaluated, the mura grade index was obtained, the threshold was combined to filter the final mura, and the overall results were determined. Eventually, the SEMUMaxMin quantization module was designed as the post-processing for the resulting image obtained from the network model reasoning, and the accurate defect data was obtained by setting the threshold filtering.

Experimental evaluation and findings

An automatic mura detection device was developed in this study based on the machine vision detection principle, with the device's visual part containing a computer operation unit, screen drive module, motion control module, and image acquisition module.

Additionally, the OLED mobile screen utilized in this experiment had a 1080 pixels * 2400 pixels screen resolution, a German AVT Manta G-609C color industrial camera with a six million pixels camera resolution, and a screen width and length of 154.56 mm x 69.522 mm.

A total of 334 small sample set images were utilized in the algorithm experiment, and the trained OLED detection model was imported into the OLED panel production line for mass production evaluation. The visual detection accuracy achieved while detecting point mura defects using the proposed SmartMuraDetection was 96%, indicating the method's effectiveness in point mura defect detection.

Moreover, the mura detection accuracy was also significantly higher compared to the accuracies achieved in recent studies, including the 86.80% accuracy realized using convolutional neural networks (CNNs) and 87% accuracy achieved using a Color-Mura defect detection approach.

Overall, the findings of this study demonstrated the feasibility of using the proposed small-sample DL-based approach for the highly accurate detection of point mura defects. However, the approach is only effective for point mura detection, which is a major limitation that must be addressed in future research.