Self-Supervised Learning (SSL) for Anomaly Detection is the training of models on only normal (defect-free) data using self-supervised tasks — the model learns the distribution of normal patterns, and anything that deviates from the learned normality is flagged as an anomaly.

Key SSL Approaches for Anomaly Detection

• Autoencoders: Learn to reconstruct normal images. Anomalies have high reconstruction error.
• Contrastive Learning: Learn representations of normal data. Anomalies have distant embeddings.
• Knowledge Distillation: Student network trained on normal data disagrees with teacher on anomalies.
• Masked Image Modeling: Predict masked regions — anomalies are poorly predicted.

Why It Matters

• No Defect Labels Needed: Only requires normal (good) images for training — defect labels are expensive and rare.
• Novel Defects: Detects previously unseen defect types (anything abnormal), not just known categories.
• Industrial Standard: Approaches like PatchCore and FastFlow achieve >99% AUROC on industrial anomaly benchmarks.

SSL for Anomaly Detection is learning what normal looks like — training exclusively on good data so that any deviation is automatically flagged as suspicious.