LG AI Research Overcomes the Limits of AI Vision Inspection

In the manufacturing process of PCBs (Printed Circuit Boards), a core component of electronic products, Vision Inspection is an essential process for finding defects in products. In particular, the board materials field that LG Innotek focuses on requires millimeter-level precision, allowing no room for even minor defects.

Existing Supervised Learning-based vision inspection faces three major limitations.

• Lack of defect data: Compared to normal products, the frequency of defects is very low, making it difficult to secure data for training.
• Catastrophic Forgetting: When learning new types of defects, the AI forgets previously learned information, or retraining consumes enormous resources.
• Difficulty of verification: To trust the AI's classification results, humans must go through the trouble of manually checking samples one by one.

To address this, LG AI Research, LG Innotek, and LG Electronics applied new AI technology, including Self-supervised Representation Learning, through joint research. Instead of separately learning normal and abnormal cases, this technology utilizes an Unsupervised Anomaly Detection approach that learns the characteristics of normal cases and classifies everything different from them as an anomaly.