An Experimental Study of Knowledge Distillation Techniques in EfficientAD for Pharmaceutical Quality Control

The thesis investigates the application of visual anomaly detection techniques to industrial quality control, with a specific focus on the EfficientAD method. The work focuses on the role of the Teacher within the Teacher–Student pipeline, analyzing how the distillation process, the dataset used, and the reference network influence the overall performance of the system. Several experiments were conducted on industrial data provided by Brevetti C.E.A., comparing different strategies for Teacher generation and evaluating their impact on anomaly detection and localization. The results highlight the main advantages and limitations of the analyzed configurations, providing useful insights for adapting EfficientAD to real industrial inspection scenarios.