Ask a developer to build AI for bridge inspection and most will reach for object detection: train a model, draw a boundary box around the rust, ship it. That work has been done. Off-the-shelf models can now reliably flag surface defects in a single image, and the capability is becoming a commodity.
The problem is that a boundary box answers the wrong question. A structural engineer does not need software merely to confirm that a crack exists. They need to know whether that crack has grown 2% or 20% since the previous inspection cycle, because that rate of change determines whether a repair is a scheduled line item or an emergency.
Single-image inference cannot produce that number. Every frame is treated as an independent data point, leaving humans to reconstruct the connective tissue between years of inspection data by hand.
That reconstruction is brutal work. Asset owners sit on mountains of high-resolution imagery, but measuring change requires a senior engineer to search through thousands of images and determine which photographs from 2019, 2022 and 2025 show the same square metre of the same beam, despite differences in angle, lighting, camera, environment and inspection crew.
All the data is already there. What is missing are the algorithmic systems that can connect it reliably across time. That is the problem we set out to solve.
The correspondence problem
In computer vision terms, this is the correspondence problem: determining whether points observed in two separate images represent the same physical location in 3D space.
Classical photogrammetry techniques such as SIFT, a computer vision algorithm, rely on sparse, hand-designed local features and often fail on civil infrastructure surfaces that are either textureless or highly repetitive. Deep-learning methods instead learn visual representations from data, making them more robust to changes in viewpoint, lighting, surface condition and local appearance.
From sliding windows to global attention
Early deep-learning approaches were built primarily on deep-learning algorithms known as convolutional neural networks. CNNs can recognise the texture of a bolt, but may struggle to distinguish it from hundreds of similar bolts elsewhere on the structure.
The arrival of advanced deep-learning modules known as vision transformers changed this by using self-attention – a technique to model sequences – to capture relationships across the entire image. A specific bolt becomes distinguishable not only by its local appearance, but also by its position relative to the wider structure.
An architecture known as LoFTR, introduced in 2021, was an important step in this transition. Rather than detecting interest points first and matching them afterward, it combines two attention mechanisms: self-attention within each image, allowing features to incorporate their broader visual context, and cross-attention between the two images, allowing features from both viewpoints to be compared directly.
This produces dense correspondences informed by both local visual evidence and global structural context. Newer architectures have since improved speed, robustness and matching accuracy, but the underlying principle remains the same: understand each feature within the wider image, then compare those contextualised features across views.
On our benchmark of real-world inspection image pairs, hand-labelled to establish the truth, the precision-recall curve for the vision transformer-based engine sits substantially above those of earlier correspondence methods across the relevant operating range.
Raw model output is not production-grade performance
Here is the part that matters most to anyone building AI systems for high-stakes domains. Neural networks behave as black boxes and can produce unintuitive errors, making their raw outputs insufficient for unsupervised deployment.
The model therefore operates within a broader algorithmic pipeline designed to identify and reject unreliable correspondences.
Every component in this pipeline is calibrated against a human-verified dataset. The first layer applies a confidence threshold to remove low-certainty correspondences. Statistical analysis of the remaining match distribution then identifies and removes outliers.
If too few matches remain after these filtering stages, the image pair is rejected because there is insufficient evidence to support a reliable link. Finally, convex-hull checks enforce spatial diversity across the matched keypoints, rejecting cases in which the apparent evidence is concentrated within only a small region of the image.
Together, these layers transform raw model predictions into correspondences that are substantially more reliable for downstream structural analysis.
The output is a validated match graph: repeatable, cross-view and cross-year connections that turn a chaotic image archive into a longitudinal record. It is the same conceptual shift medicine made when it stopped reading each X-ray in isolation and began reading them as a series.
The lesson extends well beyond bridges. As developers integrate AI into systems where errors carry real-world consequences, strong model performance is only the starting point. The additional validation layer, supported by carefully curated ground-truth data, is what turns a promising technical result into a production-grade system.
The model performs the core task. The surrounding software determines when its output is reliable enough to use—and when it should be rejected.
Arik Voronov is CTO at Dynamic Infrastructure, which provides an AI platform to turn inspection imagery into health records for bridges, tunnels and other assets
