Memory-Efficient Continual Learning for Visual Anomaly Detection: A Compressed Replay approach for Teacher-Student architectures

Anomaly Detection is a critical task in computer vision with numerous real-world applications. While traditional anomaly detection methods have achieved significant progress, the challenge of continual learning—where models must adapt to new data while retaining previously learned knowledge—remains underexplored. This thesis investigates Pixel-Level Anomaly Detection in Continual Learning using the STFPM model. An in-depth analysis of the MVTec Anomaly Detection dataset for STFPM is conducted, offering valuable insights. The performance of STFPM is then evaluated in the continual learning scenario. Based on these results, a more efficient architecture than STFPM, called PaSTe, is tested for the first time in the continual learning setting. Furthermore, to reduce memory requirements, a Compressed Replay approach is adopted. This approach minimizes the memory occupied by the replay buffer without compromising model performance, enabling more efficient and scalable anomaly detection in resource-constrained environments.