Infrastructure inspections today are still slow, expensive, and mostly manual. Visual checks can take hours and often require scaffolding, night shifts, and specialised equipment. Even then, results are often heterogeneous and unreliable.
As digital technologies evolve, there’s no reason their advantages shouldn’t extend to this field as well. In fact, we’re already seeing real-world applications. For example, a Japanese bullet train operator has adopted an AI-powered system that replaces the time-consuming manual task of walking kilometres of track to inspect for loose bolts and other issues. This change has not only enhanced accuracy but also dramatically improved operational efficiency.
Over the past few months, in the scope of our infrastructure projects such as RoboTUNN, we explored this topic in depth, narrowing our focus to how structural damage, specifically crack width, can be measured automatically. Crack width is one of the most important indicators of structural degradation. Accurately determining its severity enables early intervention, reduces the need for extensive repairs and minimizes downtime.
From the wide range of research available, we tested three promising methods ourselves. Below, we present what worked well and where challenges remain.
1. Orthogonal Projection (OP) — The Classic Approach
This method forms the backbone of many image-based infrastructure inspection systems. The approach, as used by Wang et al. (2021), combines classical image processing with deep learning to classify crack severity levels based on geometric width.
To go from geometry to classification, one then trains a deep learning model to predict severity levels from image input alone, so no manual measurements are required after training.
Strengths:
- Well-suited for cracks in controlled, structured environments like rail tracks.
- Efficient and fast once trained.
Limitations:
- Not ideal for heavily curved or branching cracks as orthogonal directions can become ambiguous or inaccurate.
- Performance depends on high-quality binary masks; segmentation errors reduce measurement accuracy.
Learn more: How does Orthogonal Projection work?
- The skeleton of the crack is extracted.
- Irrelevant branches are pruned, focusing only on the main “trunk” of the crack.
- Crack intersections are removed.
- Triple-PCA is applied at each skeleton point and averaged to form a corrected crack direction.
- Finally, OP is used to measure the width between boundary points, and the Euclidean distances are averaged.
2. OrthoBoundary (OB) Methode - A More Robust Version
Building on OP, the OB method, reviewed in Zhe Li et al. (2024), addresses some weaknesses like noisy edges, branching structures, and asymmetrical cracks. OB integrates skeleton-based analysis with advanced pruning, orientation correction, and localized PCA to estimate crack width in complex pavement images.
Learn more: How does OrthoBoundary method work?
- The skeleton of the crack is extracted.
- Irrelevant branches are pruned, focusing only on the main “trunk” of the crack.
- Crack intersections are removed.
- Triple-PCA is applied at each skeleton point and averaged to form a corrected crack direction.
- Finally, OP is used to measure the width between boundary points, and the Euclidean distances are averaged.
The method outperformed other commonly used techniques, including the classical OP, in terms of accuracy, speed, and robustness.
Strengths:
- More robust in noisy images.
- Works well on complex crack geometries, including curved and bifurcated cracks.
Limitations:
- Does not measure crack width at intersections, which are excluded from analysis.
- Still struggles with sharply curved cracks.
3. Frequency Domain Analysis — Cracks as Signals
This approach views crack detection through a completely different lens. Instead of spatially measuring width, it analyses the frequency characteristics of cracks using video frames and signal processing techniques.
Based on the Cao et al. (2023) paper, this method first applies Robust Principal Component Analysis (RPCA) to decompose video frames into two components: a low-rank matrix representing the background and a sparse matrix capturing crack structures. This separation reduces background noise and shows fine crack details, without requiring thresholding or binarization.
Learn more: FDA Stepwise Explained
- A 2D Discrete Fourier Transform (DFT) is applied to the sparse matrix.
- The magnitude spectrum is centred, log-transformed, and analysed.
- An equivalent ellipse is fitted to the frequency components to extract geometric features such as crack width and orientation.
- A 1D power spectral density (PSD) curve is also computed, tracking how spectral power (i.e., “crack intensity”) varies with frequency over time.
This enables not just width estimation, but also the crack evolution under stress. Compared to the first two methods which acquire direct spatial measurements of crack geometry, this approach measures crack characteristics indirectly by analysing their spectral signatures over time, making it more suitable for dynamic monitoring than for precise static width estimation.
Strengths:
- Captures rich crack information beyond width, including growth trends, orientation, and temporal evolution.
- Highly resistant to background noise due to RPCA preprocessing.
Limitations:
- Still relies on manual thresholding in the frequency domain to extract crack features.
- Requires high-quality video input and careful parameter tuning.
Major Challenges: Lack of Ground Truth
A significant challenge across all methods is the absence of reliable ground truth data. Public datasets are rarely annotated with actual width data. This makes it difficult to benchmark the accuracy of any method. Furthermore, data heterogeneity, such as varying camera angles or distances, further complicates reliable comparison.
Until more comprehensive datasets become available, we rely on indirect validation methods and cross-comparisons.
While each method we explored has its own strengths and weaknesses, they all point in the same direction: crack width measurement can be faster, more reliable, and more scalable with the help of AI. Manual inspection alone is no longer sufficient, especially as infrastructure ages and workloads grow. These approaches show that with the right tools and the right data, we can build smarter methods that not only detect problems earlier but also help prevent them altogether.
***
References
Cao, J., He, H., Zhang, Y., Zhao, W., Yan, Z., & Zhu, H. (2023). Crack detection in ultrahigh-performance concrete using robust principal component analysis and characteristic evaluation in the frequency domain. Structural Health Monitoring, 23(2), 1013–1024. https://doi.org/10.1177/14759217231178457
Li, Z., Miao, Y., Eskandari Torbaghan, M., Zhang, H., & Zhang, J. (2024). Semi-automatic crack width measurement using an OrthoBoundary algorithm. Automation in Construction, 158, 105251. https://doi.org/10.1016/j.autcon.2023.105251
Wang, W., Hu, W., Wang, W., Xu, X., Wang, M., Shi, Y., Qiu, S., & Tutumluer, E. (2021). Automated crack severity level detection and classification for ballastless track slab using deep convolutional neural network. Automation in Construction, 124, 103484. https://doi.org/10.1016/j.autcon.2020.103484
