Industrial defect detection plays a pivotal role in maintaining quality and safety across manufacturing processes. Traditional deep learning methods for visual inspection and defect classification rely heavily on large volumes of annotated data, which are often costly and difficult to obtain in real-world industrial settings. This data scarcity poses a significant barrier to deploying robust and generalizable computer vision models for rare or evolving defect types.

To address this challenge, we explore the use of few-shot learning (FSL), a paradigm that enables models to generalize to new classes with only a handful of labeled examples. Building upon this foundation, we integrate meta-learning strategies specifically model-agnostic algorithms and metric-based learners—that are trained to quickly adapt to new tasks with minimal supervision. To further enhance feature discrimination under limited data conditions, we incorporate contrastive learning, which encourages the model to learn meaningful representations by maximizing inter-class differences and minimizing intra-class variations through self-supervised instance discrimination.

This study presents a hybrid framework combining contrastive pretraining with meta-learning to achieve superior performance in few-shot defect detection tasks. Experiments conducted on benchmark industrial datasets such as MVTec AD and DAGM demonstrate that our approach outperforms conventional few-shot baselines in both accuracy and generalization. The inclusion of contrastive learning boosts feature separability and improves recognition performance in low-shot settings. Our findings indicate that the proposed method is a viable and scalable solution for deploying intelligent inspection systems in real-world manufacturing environments, especially where annotated data is limited or difficult to collect.

Xin NIE