Robust Inspection of Micro-LED Chip Defects Using Unsupervised Anomaly Detection

Inspection of defects in light emitting diode (LED) chips have been studied to reduce manufacturing cost. Recent studies proposed visual defect inspection methods based on supervised learning of deep neural networks. However, they require datasets with the ground-truth label of each chip, which accompanies prohibitively laborious tasks. In addition, they require class balanced datasets, which is difficult to obtain in an actual industrial process. In order to tackle these limitations, this paper proposes an unsupervised learning based inspection method using anomaly detection that requires no labeled data. On the micro-LED dataset, we demonstrate that our method outperforms previous anomaly detection methods. We achieve 95.82% AUROC result, which is 20.87% higher than convolutional autoencoder and 0.67% higher than Deep SVDD.