LG AI Research Pushes Past the Limits of Vision Inspection AI

In the manufacturing process of PCBs (Printed Circuit Boards), a core component of electronic products, vision inspection to detect minute defects is extremely important. However, the existing supervised learning approach had limitations: a data scarcity problem where defect data is very limited, Catastrophic Forgetting, where past knowledge is lost when learning new defect types, and the burden of arduous re-validation work.

To address this, a joint research team from LG AI Research, LG Innotek, and LG Electronics applied Unsupervised Anomaly Detection technology. This approach learns only the characteristics of normal products, without needing to label defect data, and identifies anything deviating from that standard as an anomaly.

The research team used Contrastive Learning to overcome the low accuracy of existing unsupervised learning models. By applying various Data Augmentation techniques to good-quality product data, they guided the model to learn the essential representation of normal products.

In particular, reflecting the domain characteristics of board materials, the model was designed to classify changes such as brightness variation or rotation as normal, while transformations like cutting out or attaching parts were classified as defects. This technology is set to be applied to actual production lines, and is expected to become the first commercialized case based on unsupervised learning.