Coarse-to-Fine Training

Coarse-to-Fine Training is a hierarchical training strategy that first learns coarse, global patterns, then progressively refines to learn fine-grained, local details — structuring the learning process from the big picture to the details.

Coarse-to-Fine Approaches

• Resolution: Start with low-resolution inputs (coarse spatial features), increase resolution for fine details.
• Label Hierarchy: First learn coarse categories (defect vs. no-defect), then fine categories (defect type).
• Loss Weighting: Start with losses that emphasize global structure, shift to losses for local detail.
• Architecture: Train shallow layers first (coarse features), then progressively train deeper layers (fine features).

Why It Matters

• Curriculum: Provides a natural curriculum — easy coarse task first, hard fine-grained task later.
• Stability: Coarse features provide a stable foundation for learning fine details.
• Semiconductor: Defect classification naturally follows coarse-to-fine — classified by severity first, type, then root cause.

Coarse-to-Fine Training is learning the outline before the details — structuring training to build from global understanding to fine-grained precision.