Comparative Evaluation and Implementation of State-of-the-Art Techniques for Anomaly Detection and Localization in the Continual Learning Framework

The capability of anomaly detection (AD) to detect defects in industrial environments using only normal samples has attracted significant attention. However, traditional AD methods have primarily concentrated on the current set of examples, leading to a significant drawback of catastrophic forgetting when faced with new tasks. Due to the constraints in flexibility and the challenges posed by real-world industrial scenarios, there is an urgent need to strengthen the adaptive capabilities of AD models. Hence, this thesis introduces a unified framework that integrates continual learning (CL) and anomaly detection (AD) to accomplish the goal of anomaly detection in the continual learning (ADCL). To evaluate the effectiveness of the framework, a comparative analysis is performed to assess the performance of the three specific feature-based methods for the AD task: Coupled-Hypersphere-Based Feature Adaptation (CFA), Student-Teacher approach, and PatchCore. Furthermore, the framework incorporates the utilization of replay techniques to facilitate continual learning (CL). A comprehensive evaluation is conducted using a range of metrics to analyze the relative performance of each technique and identify the one that exhibits superior results. To validate the effectiveness of the proposed approach, the MVTec AD dataset, consisting of real-world images with pixel-based anomalies, is utilized. This dataset serves as a reliable benchmark for Anomaly Detection in the context of Continual Learning, providing a solid foundation for further advancements in the field.