Self-Supervised Learning for Few-Shot Industrial Defect Detection

Ensuring quality in industrial production relies heavily on reliable defect detection, yet acquiring large volumes of annotated data remains cost-prohibitive—especially when rare or subtle anomalies are involved. This work introduces a domain-adaptive self-supervised framework designed to address such constraints in low-resource settings. By leveraging contrastive pretraining on unlabeled images, our method learns representations that capture essential structural detail while maintaining robustness to production-line variability. To address distribution shifts between different manufacturing domains, we employ a domain alignment strategy that enforces feature consistency across environments. Furthermore, a few-shot classification module enables practical adaptation with limited labeled examples. Empirical evaluations on standard industrial defect benchmarks confirm that under the condition of cross-domain and low data volume, the average accuracy rate of our method reaches 40.5%, which is significantly higher than the 25.02% of the random initialisation benchmark method and the traditional HOG+SVM method, demonstrating its applicability in real-world manufacturing scenarios.