Unsupervised Image Demoiréing With Self-Consistent GAN for TFT-LCD Defect Recognition

Abstract

In TFT-LCD (thin film transistor-liquid crystal display) manufacturing industry, achieving accurate defect detection is a critical and a complex task, which involves using optical inspection technology to capture images of the testing objects and classify defects by image recognition. However, using cameras to capture panel images often results in moiré patterns, which can distort the appearance of defects, making defect classification challenging. Previous studies on moiré pattern removal in TFT-LCD panel often relies on paired data with labels. This study proposes a new method for eliminating moiré patterns without label data, and we propose 3-phase self-consistent generative adversarial networks (3SC-GANs) considering the frequency loss, compared with other existing supervised and unsupervised models. An empirical study of a leading panel manufacturer is conducted to validate the proposed model, and the results show that the proposed model outperforms other benchmark methods by evaluating image quality and defect classification metrics.