Mura defect classification is a critical concern for thin-film transistor liquid crystal display (TFT-LCD) manufacturers. In recent years, artificial intelligence technologies have been successfully applied in numerous areas. However, such approaches require large amounts of training image data. Simultaneously, product differentiation and customization strategies have forced the TFT-LCD manufacturing industry to shift from mass production to high-mix, low-volume, and short-life-cycle production. In this environment, collecting a large amount of training data is difficult. Moreover, images with Mura defects captured at inspection stations remain challenging because they are often contaminated with moiré patterns. Moiré patterns severely affect the visual quality of images and cause difficulty in determining Mura defects. This study proposes an approach to eliminate moiré patterns from defect images using a conditional generative adversarial network. In addition, we develop a transfer learning ensemble model that aggregates multiple convolutional neural networks based on a denoising network for defect classification in a limited training data set. The results from an industrial case study demonstrate that the proposed method provides improved accuracy for Mura defect classification. This method can therefore become a viable alternative to manual classification in the TFT-LCD manufacturing industry.

中文翻译：

---

CNN 与条件 GAN 相结合，用于 TFT-LCD 中的 Mura 缺陷分类


Mura 缺陷分类是薄膜晶体管液晶显示器 (TFT-LCD) 制造商最关心的问题。近年来，人工智能技术在众多领域得到成功应用。然而，此类方法需要大量的训练图像数据。同时，产品差异化和定制化策略迫使TFT-LCD制造业从大规模生产转向多品种、小批量、短生命周期生产。在这种环境下，收集大量的训练数据是很困难的。此外，在检查站捕获的具有 Mura 缺陷的图像仍然具有挑战性，因为它们经常受到莫尔图案的污染。莫尔图案严重影响图像的视觉质量，并导致难以确定 Mura 缺陷。本研究提出了一种使用条件生成对抗网络消除缺陷图像中莫尔图案的方法。此外，我们开发了一种迁移学习集成模型，该模型聚合了基于去噪网络的多个卷积神经网络，用于在有限的训练数据集中进行缺陷分类。工业案例研究的结果表明，所提出的方法提高了 Mura 缺陷分类的准确性。因此，该方法可以成为 TFT-LCD 制造行业中手动分类的可行替代方案。