image segmentation for defects

**Image Segmentation for Defects** is the **pixel-level classification of wafer and device images into defect and non-defect regions** — providing precise defect outlines, sizes, and areas rather than just bounding boxes, enabling accurate dimensional measurement of defects. **Deep Learning Architectures** - **U-Net**: Encoder-decoder architecture with skip connections — the standard for defect segmentation. - **Mask R-CNN**: Instance segmentation that separates individual defects even when overlapping. - **DeepLab**: Atrous convolutions for multi-scale segmentation of complex defect patterns. - **Semantic vs. Instance**: Semantic segments by class (defect type). Instance separates individual defects. **Why It Matters** - **Precise Sizing**: Segmentation provides exact defect area, perimeter, and shape — critical for severity assessment. - **Kill Analysis**: Precise defect outlines enable accurate overlap analysis with circuit patterns for kill probability. - **SEM Review**: Automated segmentation of SEM review images replaces manual outlining. **Image Segmentation** is **pixel-perfect defect delineation** — tracing the exact boundary of every defect for precise dimensional and kill-probability analysis.