Instance Segmentation of Defects is the detection and pixel-level delineation of each individual defect instance — combining object detection (where is each defect) with semantic segmentation (what shape is it), distinguishing separate defects even when they overlap or touch.

Key Architectures

• Mask R-CNN: Extends Faster R-CNN with a mask prediction branch for each detected instance.
• YOLACT: Real-time instance segmentation combining detection and prototype masks.
• SOLOv2: Directly segments instances without explicit detection, using dynamic convolutions.
• Cascade Mask R-CNN: Multi-stage refinement for higher-quality masks.

Why It Matters

• Individual Counting: Counts separate defects even when they touch or are closely spaced.
• Per-Defect Metrics: Computes area, shape, orientation for each individual defect independently.
• Kill Probability: Per-instance analysis enables individual kill probability estimation for each defect.

Instance Segmentation is giving each defect its own identity — separately outlining and classifying every individual defect for precise per-defect analysis.