BENCHMARKING FEW-SHOT LEARNING MODELS IN COMPUTER VISION FOR INDUSTRIAL DEFECT DETECTION AND CLASSIFICATION

Industrial automation relies heavily on computer vision systems for accurate defect detection and classification. However, limited data availability for novel defect types poses significant challenges. This study investigates the performance of state-of-the-art few-shot learning (FSL) models in industrial defect classification scenarios. We compare meta-learning, metric-learning, and generative-based frameworks on public datasets such as MVTec AD and VisA. We also propose a benchmarking protocol with standardized metrics and a dataset split strategy optimized for real-world deployment. Our results highlight that meta-learning models like MAML and ProtoNet demonstrate better generalizability, while transformer-based methods offer robustness in cross-domain scenarios. This work contributes a reproducible benchmark and recommendations for choosing FSL models in low-data industrial.