AI Models Team Up to Spot Cracks in Bridges, Roads and Buildings Before Disaster Strikes

Every year, engineers face the same stubborn problem: cracks in concrete, asphalt, masonry and ceramic surfaces are the earliest warning signs of structural decay, yet finding them reliably at scale remains one of infrastructure management’s hardest challenges. Manual inspection is slow, expensive, and dangerously subjective, while automated systems have long been hampered by a lack of high-quality annotated training data. Now, a team at the Universidad Autónoma de Baja California in Ensenada, Mexico, has unveiled a framework that tackles the data bottleneck head-on by combining three of the most powerful general-purpose vision models ever released. The result, published in Multimedia Tools and Applications, is a pipeline that detects cracks with 92.3 percent accuracy and segments them with steadily improving precision, processing high-resolution images in roughly 2.3 seconds on a standard GPU.

The research, led by R. Daniel Verdugo-Siqueiros with Luis Pellegrin and Jose Angel Gonzalez-Fraga, is explicitly data-centric rather than model-centric. Instead of inventing yet another neural architecture, the authors asked a different question: what if the real obstacle to automated crack inspection is not the detector but the scarcity of well-labeled examples spanning diverse materials and environmental conditions? Their answer was to exploit the zero-shot generalization abilities of foundational vision models, which have been trained on such vast image collections that they can perform tasks they were never explicitly taught, and to augment limited real training data with synthetic examples generated by artificial intelligence itself.

At the heart of the pipeline sits Segment Anything Model 2, or SAM2, Meta’s celebrated segmentation system capable of isolating objects in images and videos from simple prompts. SAM2 provides the framework’s precise, prompt-driven segmentation capability, tracing the often hairline, branching geometry of cracks that would defeat conventional thresholding methods. But SAM2 alone cannot know where to look. That role falls to Florence2, a unified vision foundation model developed at Microsoft Research, which supplies robust feature detection and localization, effectively telling the segmentation module which regions of a photograph deserve scrutiny. The pairing mimics a human inspector’s workflow: first scan the surface for suspicious marks, then examine each candidate closely.

The third component is perhaps the most surprising. Stable Diffusion XL, the high-resolution image synthesis model best known for generating photorealistic artwork, is repurposed here for two jobs: super-resolution enhancement of real inspection images and the generation of synthetic crack examples to flesh out sparse training sets. By upscaling imagery before analysis, the system preserves the fine detail that thin cracks demand. By manufacturing plausible synthetic cracks, it sidesteps the annotation drought that has plagued the field, where datasets covering concrete, asphalt, masonry and ceramics simultaneously are nearly nonexistent.

The numbers reported in the paper are striking for a system built largely from off-the-shelf components. Detection accuracy and precision both reach 92.3 percent, with recall of 89.7 percent and an F1-score of 91.0 percent, meaning the framework misses relatively few genuine cracks while raising few false alarms. The segmentation module, which must trace exact crack boundaries rather than simply flag their presence, starts at a mean Intersection over Union of 63 percent, a Dice coefficient of 77.2 percent, and a boundary F1-score of 68.1 percent. Those boundary metrics matter enormously in practice, because crack width is a key input to structural condition ratings, and sloppy edges translate directly into misjudged severity.

What elevates the work beyond a static benchmark is its use of active learning. Across three successive learning cycles, the segmentation performance climbs from an initial mIoU of 63 percent to 72 percent. In an active learning loop, the model flags the examples it is least confident about, human annotators label only those, and the model retrains on the enriched set. This targeted feedback means each hour of expert labeling yields maximum benefit, progressively refining the model exactly where it struggles, such as shadows, stains, and surface textures that masquerade as cracks. The authors describe this as a reflection of progressive refinement, and the trajectory suggests further gains with additional cycles.

The broader context explains why this matters. The paper’s own references trace a decade of effort, from classical image-processing techniques built on Otsu thresholding and differential images, through deep convolutional networks like DeepCrack and transformer-based systems such as Crackformer, to recent fine-tuning of vision foundation models for civil infrastructure. Each generation improved sensitivity, but each also inherited the same fundamental constraint: models are only as good as the labeled cracks they see. Surveys of crack detection technologies from 2022 and 2023 document the field’s rapid growth alongside persistent generalization failures, where a model trained on one pavement type stumbles on another. A framework that leans on zero-shot capability and synthetic augmentation is a direct strike at that weakness.

Computationally, the system is refreshingly lean. Processing high-resolution images in approximately 2.3 seconds on a standard graphics card means the pipeline could plausibly run on the kinds of hardware mounted on inspection drones, robots, or even field laptops, rather than being confined to cloud clusters. That opens the door to comprehensive monitoring programs in which bridges, tunnels, dams and highway networks are photographed routinely and analyzed automatically, with cracks identified, measured and tracked over time. The authors report robust performance across diverse materials and environmental conditions, the variability that has historically broken narrower systems. Data and models are available on request, and the team maintains a public GitHub repository, smart-infrastructure-inspector, dedicated to the work.

The stakes are not abstract. The collapse of the I-35W bridge in Minneapolis in 2007 and the 2021 closure of the I-40 bridge over the Mississippi River after a cracked tie girder was discovered put the cost of undetected deterioration on national display. Probability-of-detection studies for visual inspection of steel bridges have shown how inconsistent human inspection can be, and maintenance economics research repeatedly finds that early, inexpensive intervention prevents catastrophic, budget-breaking repairs. An automated inspector that never tires, never skips a hairline fracture in poor lighting, and improves with every labeled example could shift infrastructure management from reactive crisis response toward genuine predictive maintenance.

There are, of course, caveats. The reported metrics come from the authors’ own evaluation, and the paper does not yet demonstrate deployment at the scale of an entire highway network or a fleet of aging bridges. Synthetic data, however cleverly generated, risks embedding biases if the diffusion model’s imagined cracks diverge from real failure modes, and human annotators remain essential to the active learning loop. Still, the conceptual move is hard to dismiss: rather than waiting for a perfect crack dataset that may never exist, the framework manufactures its own training material and borrows perception from models trained on billions of images. As foundational vision models continue to mature, the authors’ integration strategy, pairing detection, segmentation and generation in a single data-centric loop, may well become the template for automated inspection across civil engineering, and a quiet but meaningful step toward infrastructure that tells us it is failing before it does.

Subject of Research: Automated crack detection and instance segmentation in construction materials using foundational vision models

Article Title: A data-centric framework for automated crack instance segmentation in construction materials via integration of foundational vision models

Article References: Verdugo-Siqueiros, R. D., Pellegrin, L., & Gonzalez-Fraga, J. A. (2026). A data-centric framework for automated crack instance segmentation in construction materials via integration of foundational vision models. Multimedia Tools and Applications, 85(10), Article 807. https://doi.org/10.1007/s11042-026-21964-6

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21964-6

Keywords: crack detection, instance segmentation, foundation models, SAM2, Florence2, Stable Diffusion XL, infrastructure monitoring, computer vision, active learning, synthetic data, civil engineering, concrete