CrackMorph-XAI-Net: A Topology-Preserving and Explainable Framework for Automated Crack Morphology

Explainability improves infrastructure safety: The framework moves beyond simple crack detection to provide detailed, interpretable structural information (centerline geometry, branching patterns, junctions) that engineers need for robust decision-making about infrastructure repairs, rather than just binary "crack/no crack" outputs.

Topology preservation is measurable and achievable: The system successfully preserved crack topology in 98.5% of test images with a skeleton extraction accuracy of 0.991 Dice coefficient, demonstrating that AI systems can maintain the structural integrity of what they analyze—a critical requirement for real-world engineering applications.

Staged, transparent analysis enables verification: By breaking crack analysis into four distinct, interpretable stages (skeleton extraction → junction detection → morphology computation → severity screening), the framework allows engineers to understand and verify each step of the AI's reasoning, supporting responsible deployment in safety-critical infrastructure contexts.

Benchmark improvements support rigorous evaluation: The authors extended the standard CRACK500 dataset with aligned skeleton maps, junction heatmaps, and topology labels, enabling more thorough testing of AI systems and setting a higher bar for transparency and validation in automated crack inspection tools.

Specialized metrics matter for specialized tasks: The framework achieved 96.4% recall and 0.887 F1-score in detecting sparse structural features (junctions), showing that responsible AI requires task-specific performance measures rather than generic accuracy metrics.