Anomaly detection for Automated Visual Inspection to Consider Normal Diversity

Abstract

Manufacturing companies usually employ numerous inspectors for anomaly detection and it takes a high cost including time cost. Accurate and automatic anomaly detection reduces inspection cost and improves product reliability. Anomaly detection of machine components is still challenging because they do not produce enough anomalies for supervised learning. Anomaly detection method learns only normal product to solve this problem. However, there is still a problem that dirt on a normal product is judged to be anomaly. The proposed method learns normal diversity including dirt and a wide variety of the normal image patches, aims to solve the problem.