Measurement-aware learning for reliable grain-boundary analysis in quantitative metallography

Area of Science:

• Materials Science
• Computational Imaging
• Metallurgy

Background:

• Quantitative metallography relies on accurate boundary detection for measurements.
• Existing computational methods struggle with weak, interrupted interfaces and crop-based workflows.
• Reliability gaps persist in automated boundary analysis.

Purpose of the Study:

• To investigate the impact of context, supervision alignment, and validation on boundary-derived grain-size analysis.
• To develop and evaluate improved computational approaches for reliable interface detection.
• To establish best practices for quantitative metallography using computational tools.

Main Methods:

• Context-ablation experiments to assess the role of global context in boundary identifiability.
• Introduction of MLOGRAPHY++, a context-preserving partial-labeling method.
• Evaluation of Heyn-Compare for endpoint-oriented validation of open-boundary predictions.
• Exploratory analysis of diffusion-based super-resolution under controlled conditions.

Main Results:

• Removing global context significantly degrades boundary identifiability.
• MLOGRAPHY++ aligns supervision with annotation ambiguity, reducing post-processing needs.
• Heyn-Compare complements pixel-overlap metrics for open-boundary predictions.
• Aggregate grain-size error decreased from 9.98 px to 5.38 px with MLOGRAPHY++.

Conclusions:

• Context is crucial for reliable interface detection in quantitative metallography.
• A measurement-aware approach combining context preservation, annotation-aligned supervision, and endpoint validation is effective.
• Generated super-resolution details should be interpreted cautiously and not as physical ground truth.