contrastive learning for defect embeddings

**Contrastive Learning for Defect Embeddings** is the **training of a representation model that maps defect images to a feature space where similar defects are close and dissimilar defects are far apart** — creating meaningful defect representations without requiring class labels. **How Contrastive Learning Works for Defects** - **Positive Pairs**: Two augmented views of the same defect image are pulled together in embedding space. - **Negative Pairs**: Views from different defects are pushed apart. - **Losses**: InfoNCE, NT-Xent, or triplet loss enforces the embedding structure. - **Frameworks**: SimCLR, MoCo, BYOL, DINO adapted for defect images. **Why It Matters** - **No Labels Needed**: Learns useful representations without class labels — purely self-supervised. - **Downstream Tasks**: Contrastive embeddings transfer to classification, retrieval, and clustering tasks. - **Defect Retrieval**: Find similar historical defects by nearest-neighbor search in embedding space. **Contrastive Learning** is **teaching the model defect similarity** — learning to organize defect images by visual similarity without being told the categories.