Leveraging Contrastive Learning and Masked Autoencoders for Robust Surface Defect Classification in Data-Scarce Environments

Surface defect detection in industrial settings is extremely challenging task due to the scarcity of annotated data and high variability in defect appearance. This study presents a comparative analysis of two self-supervised learning approaches—contrastive learning (SimCLR) and masked autoencoders (MAE)—for metal surface defect classification using the NEU dataset. The models were pretrained with unlabeled data and fine-tuned under low supervision. SimCLR performed better with 100% accuracy and F1-score, surpassing MAE with an accuracy of 91%. On the low-label setting (10% data), SimCLR had good performance (83% accuracy) and MAE had poor performance (39% accuracy). The findings attest to the superiority of SimCLR's robustness, efficiency, and applicability to industrial defect inspection. The study exhibits the promise of self-supervised pretraining in lessening dependence on labeled data and suggests a scalable approach to real-world visual inspection problems.