CLAD: Semi-Supervised Contrastive Learning with Augmented Negatives for Label-Efficient Industrial Anomaly Detection

In modern industrial manufacturing systems, anomaly detection serves as a critical safeguard for production safety and quality assurance. Current approaches face significant challenges in real-world deployment scenarios where labeled anomaly samples are scarce due to the high costs of expert annotation and the inherent rarity of defect occurrences. To address this critical limitation, we propose CLAD, a semisupervised contrastive learning framework that synergistically integrates contrastive learning with strategic data augmentation. Our method establishes a dual-branch contrastive learning architecture: (1) for limited labeled data, we implement supervised contrastive learning enhanced by CutPaste-based augmented negative learning, creating discriminative feature representations through explicit category-aware contrastion; (2) for abundant unlabeled samples, we design an adaptive contrastive learning mechanism that automatically discovers latent data relationships through similarity-aware instance discrimination. The proposed architecture employs shared-weight encoders with projection heads to enable joint representation learning across supervised anomaly discrimination and unsupervised representation alignment. Extensive experiments on the MVTec-AD benchmark demonstrate significant improvements over existing methods, achieving 87.0% AU-ROC and 91.9% AU-PR with merely 10% labeled data.