Anomaly detection based on generative adversarial network in the manufacturing process of lcd/oled display panels

Abstract

In the manufacturing process of LCD/OLED, defects on display panels need to be localized and classified according to certain criterion. Recent triumph of deep learning model in defects detection on LCD/OLED panels greatly reduce the miss and mistake rate of defects while depends tightly on the industrial training data. These image data, acquired from the industrial display pipelines, show great imbalance with the positive sample way surpassing negative defective ones. Despite data imbalance, the diversity in the negative samples make the data preparation trick and certainly impossible to exhaust all kinds of negative samples for training. Based on the above observation, it would be enormously beneficial if the anomaly detection on display panels can be purely based on the overwhelming positive samples with enough variations. This insight serves as our motivation for GAN-based anomaly detection on LCD/OLED display panels. In this paper, we propose the utilization of one specific anomaly GAN to the real time anomaly detection of display panels online and also automatic data labeling offline. The result shows its efficiency in detecting all kinds of prominent anomalies.