Industrial Anomaly Detection System: A Multicase Algorithm Leveraging Feature Information and Memory Bank

The conventional anomaly detection (AD) methods typically rely on training normal samples without defects to identify deviations from the background. However, these methods suffer from issues, such as missing detection or false detection. Although subsequent research has attempted to improve algorithm performance, this often results in overfitting to specific cases within the AD framework. To address these challenges, this article proposes a semi-supervised AD algorithm that combines improved metric learning techniques with a memory bank (MB) update module. In order to enhance the algorithm’s generalization capabilities across different cases, a multicases MB inheritance approach is introduced. This approach facilitates rapid generalization to unknown test cases with minimal iterative learning ( $\le 5$ epochs). Additionally, a bank-case matching module is designed to select the appropriate MB and calculate anomaly scores within our framework. The effectiveness of the proposed algorithm has been validated through real industrial tests and ablation experiments, demonstrating its capability in detecting anomalies accurately and reliably. Code is available on: https://github.com/FrankCloud-UESTC/Multi-case-AD.