Analiza sprotnega nenadzorovanega učenja za detekcijo površinskih anomalij

Industrial anomaly detection systems are typically built using separate models for each product category, resulting in substantial computational demands and ongoing maintenance challenges. This thesis explores continual learning as an alternative strategy, enabling a single model to sequentially learn and retain knowledge across multiple product categories. We conduct a comprehensive evaluation of three state-of-the-art methods— PatchCore, EfficientAD, and SuperSimpleNet—under three training paradigms: separate models, joint training, and continual learning, using the MVTec AD dataset. The results show that continual learning can maintain robust performance, with PatchCore maintaining strong performance in continual learning compared to separate models. Moreover, memory-based approaches consistently outperform replay buffer strategies. These insights offer practical guidance for deploying anomaly detection systems in dynamic industrial environments, highlighting continual learning as a compelling and resource-efficient alternative to the traditional use of multiple dedicated models.