Semantic Segmentation of Defects is the pixel-level classification of every pixel in a wafer or device image into defect categories — assigning each pixel a label (scratch, particle, void, pattern defect, background) to create a complete defect map of the image.

Key Architectures

• U-Net: Encoder-decoder with skip connections — the workhorse for defect segmentation due to strong performance with limited data.
• DeepLab v3+: Atrous spatial pyramid pooling for multi-scale feature extraction.
• SegFormer: Transformer-based segmentation for capturing long-range spatial context.
• PSPNet: Pyramid pooling module aggregates context at different scales.

Why It Matters

• Complete Map: Every pixel is classified — no defects are missed (unlike object detection bounding boxes).
• Precise Area: Exact defect area calculation for severity assessment and yield impact analysis.
• Multiple Classes: Simultaneously segments multiple defect types in a single forward pass.

Semantic Segmentation is painting every pixel with its identity — creating complete, pixel-perfect defect maps for thorough wafer and device characterization.