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

Technical Summary: CrackMorph-XAI-Net

Problem Statement
Current automated crack inspection systems predominantly rely on deep learning models to produce binary segmentation masks. While effective for localizing damage regions, these masks lack the structural granularity required for rigorous engineering interpretation. Maintenance decisions depend not merely on the presence of a defect but on quantitative morphological descriptors such as crack length, width, orientation, branching behavior, tortuosity, and topological classification. Traditional approaches to deriving these metrics involve classical post-processing techniques (e.g., morphological thinning and medial-axis transforms) applied to segmentation masks. However, these heuristic methods are notoriously brittle; they are hypersensitive to boundary noise, often generating spurious micro-branches, broken centerlines, and distorted topologies. Furthermore, as automated systems integrate into safety-critical workflows, the "black-box" nature of standard segmentation models limits trust and explainability. There is a critical need to transition from pixel-level localization to transparent, learning-based structural prediction that yields measurable, engineering-relevant data.

Methodology
The paper proposes CrackMorph-XAI-Net, an explainable, morphology-aware framework designed to convert raw crack imagery and region masks into a sequence of interpretable structural outputs. The framework operates as a modular, four-stage cascade pipeline:

1. Topology-Preserving Skeleton Extraction: Instead of applying classical thinning algorithms, the framework formulates skeleton recovery as a learned structural prediction problem. A U-Net-style encoder-decoder network predicts a one-pixel-wide, continuous centerline. The network is optimized using a composite loss function combining Dice loss and binary cross-entropy (BCE) to handle severe foreground-background class imbalance while ensuring global spatial overlap.
2. Junction Detection via Gaussian Heatmap Regression: To address the extreme sparsity of crack junctions (branch points), the framework avoids direct pixel-wise binary classification. Instead, it employs a Gaussian heatmap regression strategy inspired by human pose estimation. Discrete junction coordinates are transformed into continuous 2D Gaussian distributions. A dedicated network predicts this probability map using Mean Squared Error (MSE) loss, smoothing the optimization landscape and allowing for robust localization of sparse structural targets.
3. Deterministic Morphology Descriptor Computation: This stage is strictly geometric and analytical, translating the predicted structural outputs into engineering metrics. It computes:
   • Crack Length: Path length along the extracted skeleton.
   • Average Width: Ratio of the total semantic crack area to the skeleton length.
   • Orientation: Dominant principal axis of the skeleton coordinates.
   • Junction Count: Number of detected branch points.
   • Tortuosity: Ratio of the skeleton path length to the Euclidean distance between endpoints.
   • Topology Class: Categorical classification (Linear, Branched, Complex, Network) based on junction abundance.
4. Severity-Oriented Screening: The final stage synthesizes these descriptors into a prioritization heuristic. While not a substitute for finite element analysis, it utilizes concepts from Linear Elastic Fracture Mechanics (LEFM) to flag critical defects for human review based on geometric bounds and structural complexity.

Dataset Contribution
To support rigorous stage-wise evaluation, the authors extended the standard CRACK500 benchmark into CRACK500-Extended. This new benchmark retains the original 500 high-resolution images and binary masks but introduces three critical, spatially aligned annotation layers:

• Skeleton Maps: One-pixel-wide, topology-preserving centerlines derived via refined medial-axis transformations.
• Junction Heatmaps: Continuous Gaussian distributions centered at annotated branch points.
• Topology Labels: Global image-level complexity categories.
  The training set was augmented to 1,896 instances using spatial transformations (flipping, rotation, elastic deformation) to ensure geometric invariance.

Key Results
Experimental validation on the CRACK500-Extended test set demonstrates significant improvements over classical baselines:

• Skeleton Extraction: The learned model achieved a mean Dice coefficient of 0.991 and an IoU of 0.983, compared to 0.912 for classical thinning. Topology preservation was maintained in 98.5% of test images, a 11.5 percentage point improvement over the baseline.
• Junction Detection: The Gaussian heatmap regression approach yielded a recall of 0.964 and an F1-score of 0.887, significantly outperforming binary classification baselines (which achieved 0.871 recall).
• Descriptor Accuracy: Strong agreement was observed between predicted and reference morphological values. Pearson correlation coefficients exceeded 0.95 for all metrics, with length and orientation correlations reaching 0.982 and 0.991, respectively.
• Topology Classification: The framework achieved an overall categorical accuracy of 83.5%, with errors primarily occurring between adjacent complexity classes (e.g., Branched vs. Complex).
• Ablation Studies: The cascade architecture (independent stages) proved superior to a unified multi-task network, preventing gradient interference between dense skeleton recovery and sparse junction localization.

Significance and Claims
The paper claims that CrackMorph-XAI-Net represents a paradigm shift in automated infrastructure inspection by moving beyond mere segmentation toward transparent, measurable, and engineering-relevant structural representations. Its primary significance lies in its intrinsic explainability: by generating intermediate structural outputs (skeletons, heatmaps) rather than opaque probability maps, the framework allows domain experts to visually verify and diagnose the model's reasoning. This transparency is presented as essential for deploying AI in safety-critical engineering workflows.

The authors position the framework as a decision-support tool designed to triage infrastructure damage and prioritize defects for expert review, rather than as a definitive replacement for material-specific failure analysis. The work emphasizes that while the system effectively quantifies 2D morphology, future iterations must address generalizability across diverse materials (e.g., steel, microstructural concrete) and integrate depth metrics for full 3D structural health monitoring. The study concludes that algorithmic output must not only localize damage but quantify its structural architecture in a verifiable manner to be truly useful for preventative maintenance planning.