LG AI Research: Industrial Anomaly Detection Technology

Anomaly Detection is a technology that identifies anomalous patterns deviating from expected patterns in a dataset. Deep learning-based anomaly detection operates by learning the pattern distribution of normal data and then detecting patterns that differ from it, and is also widely used in security and finance.

In the Vision Inspection field of the manufacturing industry, finding defects on product surfaces is at the core of quality control. However, in actual processes, the number of defective products is very small, causing the Class Imbalance problem, and there is a risk of the model malfunctioning when a new type of defect that was not seen during training appears.

To solve this problem, methods for detecting anomalies in product images are broadly classified into three types.

• Reconstruction-based method: Represents the features of normal products as low-dimensional embedding vectors, then measures the Reconstruction Error that occurs during the reconstruction process. Autoencoder, GAN, and more recently Diffusion models are used.
• Density-based method: Models the region where normal data is densely concentrated, and judges data located in low-density areas as outliers. Gaussian Mixture Model, Normalizing Flows, etc. are representative examples.
• Distance-based method: Numerically measures how far a product is from the representative features of normal products. PatchCore, presented at CVPR 2022, is the leader in this field.

LG AI Research's Vision Lab is researching these technologies to increase productivity and efficiency in industrial settings, and introduces its representative algorithms, PatchCore and ReConPatch, presented at WACV 2024.