Anomaly Detection Control Through One-class Learning in Defect Inspection Systems

This paper presents a novel image-level anomaly detection control method for defect inspection, which comprises feature- extraction from high-dimensional data and feature-learning to analyze the observed patterns. we focus on the feature-learning step, especially in Deep SVDD. To determine anomalies from images of object suspected to be defective, Deep SVDD learns the decision boundaries for the normal feature distribution in the latent space. When a new image is received by the defect inspection system, Deep SVDD maps its features to the latent space and detects anomalies based on whether the observed features fall within the decision boundaries for normal features. The decision boundary of Deep SVDD is determined based on the center-point, which is the center position of the decision boundary. The center-point is randomly fixed at the beginning of training of Deep SVDD. However, this limits the possibility of finding the optimal decision boundary. Therefore, we propose a center-point-moving method to optimize the center-point location and improve the performance of Deep SVDD. To confirm the suitability of the proposed method, we conducted experiments on MVTec-AD, a defect inspection dataset for several products in the manufacturing industry. Based on the results, the proposed method improved the performance of Deep SVDD in 11 of the 15 class domains in the MVTec-AD dataset.