Combination of Convolutional and Generative Adversarial Networks for Defect Image Demoiréing of Thin-Film Transistor Liquid-Crystal Display Image

In thin-film transistor liquid crystal display (TFT-LCD) manufacturing, the automatic recognition and classification of defects can help manufacturers monitor abnormalities, identify potential process problems, and swiftly respond to these process problems. Thus, yield loss can be reduced. However, capturing the content displayed on screen using cameras is challenging because it is often contaminated with moiré patterns. Moiré patterns originate from the interference between the pixel grids of inspection camera sensors and panel screens and adversely affect the visual quality of images, causing difficulty in determining the defect type, especially mura defects. Therefore, eliminating moiré patterns in defect images without impairing image quality is critical. However, moiré patterns are often dynamic and vary with sensor resolution, distance, and direction. Moreover, the frequency distribution of moiré patterns is broad, encompassing both low- and high-frequency components. Therefore, demoiréing is more challenging than other image restoration tasks. We investigated this problem and proposed an approach to eliminate moiré patterns from defect images based on a generative adversarial network architecture by using the U-net network as a generator and adding a discriminator. Moreover, we added an attention mechanism to focus on the local consistency of the restored moiré regions. The peak-signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) obtained using the proposed model were compared with those obtained using the U-net and pixel-to-pixel models. The experimental results of this study demonstrated that the proposed method can efficiently remove moiré patterns from defect images, and the proposed method quantitatively and qualitatively outperforms other methods.

I. Introduction

In the thin-film transistor liquid crystal display (TFT-LCD) industry, automated optical inspection (AOI) machines have been used to identify defective panels during manufacturing processes. AOI has achieved satisfactory functional defect identification [2]. However, identification of mura defects is difficult using typical AOI [3]. Therefore, human intervention is required in mura defect classification, and successful classification typically depends on the experience and skill of engineers. Thus, inspection errors can occur during the inspection because of human fatigue or poor judgment. Therefore, an automatic mechanism for recognizing and classifying mura defects can help manufacturers monitor abnormalities, identify potential process bottlenecks, and swiftly respond to process problems to reduce yield loss. Since the late 2010s, artificial intelligence technologies have been applied successfully in many domains. Deep learning models are being gradually applied in the TFT-LCD industry for detecting and identifying panel defects [5], [6]. Capturing content displayed on screen using digital cameras is the first step of the automatic panel quality inspection process. Related images are often contaminated by undesired Moiré patterns. Moiré patterns are caused by frequency aliasing, particularly of overlapping patterns of the grids of display elements and camera sensors. Moiré patterns can be diverse and appear in spatially varying stripes, curves, or ripples. Moiré patterns also superimpose color variations onto images, drastically degrading the visual quality of images. Therefore, defect classification by learning-based models, especially mura defects, is considerably affected.