Integrating State-of-the-Art Approaches for Anomaly Detection and Localization in the Continual Learning Setting

The significant attention surrounding the application of anomaly detection (AD) in identifying defects within industrial environments using only normal samples has prompted research and development in this area. However, traditional AD methods have been primarily focused on the current set of examples, resulting in a limitation known as catastrophic forgetting when encountering new tasks. The inflexibility of these methods and the challenges posed by real-world industrial scenarios necessitate the urgent enhancement of the adaptive capabilities of AD models. Therefore, this thesis presents an integrated framework that combines the concepts of continual learning (CL) and anomaly detection (AD) to achieve the objective of anomaly detection in continual learning (ADCL). To evaluate the efficacy of the framework, a thorough comparative analysis is conducted to assess the performance of three specific methods for the AD task: the EfficientAD, Patch Distribution Modeling Framework (PaDiM) and the Discriminatively Trained Reconstruction Anomaly Embedding Model (DRAEM). Moreover, the framework incorporates the use of replay techniques to enable continual learning (CL). In order to determine the superior technique, a comprehensive evaluation is carried out using diverse metrics that measure the relative performance of each method. To validate the proposed approach, a robust real-world dataset called MVTec AD is employed, consisting of images with pixel-based anomalies. This dataset serves as a reliable benchmark for Anomaly Detection in the context of Continual Learning, offering a solid foundation for further advancements in this field of study.