Multi-Scale Approach for Unsupervised Anomaly Detection

This technical report presents a multi-scale approach for unsupervised anomaly detection specifically applied to metal surfaces. The study addresses the challenges faced in industrial manufacturing due to the lack of real anomalous samples and the complexities of practical scenarios that differ from well-organized datasets. It outlines the construction of a novel dataset using four types of metals that incorporate reflective and noisy images, which significantly affect the performance of existing detection methods. The report details the introduction of a multi-scale training strategy aimed at enhancing the robustness of anomaly detection against domain shifts caused by variations in camera specifications and lighting conditions. The methodology includes a reconstruction sub-network and a localization sub-network designed to identify and locate anomalies. The report also discusses various unsupervised methods and their limitations, emphasizing the need for improved strategies in dynamic industrial environments.