transfer learning for defect detection

**Transfer Learning for Defect Detection** is the **strategy of using models pre-trained on large image datasets (ImageNet) and fine-tuning them for semiconductor defect classification** — overcoming the limited labeled defect data problem by leveraging features learned from millions of natural images. **How Transfer Learning Works** - **Pre-Trained Backbone**: Start with a CNN (ResNet, EfficientNet) pre-trained on ImageNet (1.4M images). - **Feature Reuse**: Low-level features (edges, textures) transfer well to defect images. - **Fine-Tuning**: Replace the final classification layer and fine-tune on defect data. - **Strategies**: Freeze early layers (few labeled defects) or fine-tune all layers (more labeled data). **Why It Matters** - **Limited Data**: Semiconductor defect datasets are small (100s-1000s of images) — too little to train deep CNNs from scratch. - **Fast Convergence**: Transfer learning converges in 10-100× fewer epochs than training from scratch. - **Domain Gap**: Despite the gap between natural images and SEM/optical images, transfer learning consistently improves performance. **Transfer Learning** is **standing on ImageNet's shoulders** — reusing knowledge from millions of images to train accurate defect detectors with limited fab data.