CrackGNN: Revealing a Crack Severity Manifold

Shantanu Motiani, Maarten C. Stol

Track: Type D (Master/Bachelor Thesis Abstracts)

Keywords: Pavement Management, Unsupervised Clustering, Graph Neural Networks, Mutual Information, Manifold Hypothesis

Abstract: Accurate and scalable assessment of pavement crack severity from inspection images is crucial for maintenance planning, yet manual or supervised approaches suffer from limited labels and subjectivity. We propose CrackGNN, an entirely unsupervised pipeline that (i) extracts a rich set of explainable topological features from binary crack masks; (ii) constructs a graph of cracks weighted by pairwise mutual information using the Kraskov–Stögbauer–Grassberger estimator; and (iii) trains a novel Graph Neural Network to learn crack embeddings. When clustered, the learned embeddings reveal five distinct severity groups. Notably, a t-SNE projection of these embeddings forms a spiral-shaped manifold along which inferred crack severity increases monotonically, indicating the model has uncovered an intrinsic severity ordering. As part of our post-clustering analysis, we also compute a simple continuous Crack Condition Index (CCI) by linearly combining four major standardized topological descriptors. Although not the primary focus of this work, the CCI serves as a convenient ranking index for quickly comparing crack severities on a common 0–10 scale. Evaluated on a combination of three public crack image datasets (675 annotated cracks), CrackGNN achieves an average silhouette score of 0.62, approximately 105\% higher than baseline clustering, confirming effective severity separation without labels. The clusters’ CCI distributions align well with intuitive severity labels, and the feature-based design ensures interpretability. This label-free and interpretable framework can serve as a cost efficient first-pass tool for Pavement Management Systems.

Serve As Reviewer: ~Maarten_C._Stol1

Submission Number: 34