AI and data management: turning thousands of photos into records you can act on

Image AI in inspection is mostly deep learning: it learns the appearance of cracks, rust and so on from large sets of labelled photos. By level of detail, there are three kinds of task.

In practice they are chained: classification narrows to images that look suspicious, detection or segmentation extracts location and shape, and results are overlaid on photogrammetry to place each defect on a member. Poor images defeat any model; half of AI accuracy is decided by capture planning.

Performance depends less on the model’s name than on what it learned from. Public research datasets exist, such as collections of concrete crack images (e.g. SDNET2018, over 56,000 images of bridge decks, walls and pavements) and multi-defect bridge sets with five overlapping defect classes (e.g. CODEBRIM). But models trained only on public data almost always stumble in the field.

• Domain shift: different materials, dirt, light, cameras and stand-off from training data degrade performance.
• Look-alikes: formwork marks, joints, repair patches, shadows, rain streaks, cables, graffiti. These must be learned as “not a defect”.
• Imbalance: defect-free images dominate. The rarer a defect, the fewer examples and the more misses.
• Label noise: different people trace the same crack differently. Write a labelling guide aligned with the inspection standard.
• Hold out evaluation data: only images from sites not used in training reveal true performance.

Crack width drives repair decisions. The AI pipeline is: ① segment crack pixels, ② extract the centreline (skeleton), ③ count pixels perpendicular to it, ④ multiply by the GSD (mm per pixel). See the visible imaging chapter for GSD.

As a rule of thumb, allow at least about ±1 pixel of width error. At 0.5 mm GSD that is around ±0.5 mm. If repair decisions hinge on 0.2 or 0.3 mm, that image cannot settle the width. Always report AI widths with the GSD and expected error.

The real value of inspection lies in change: did a crack grow, did rust spread, did something new appear? Yet subtracting last year’s photo from this year’s means almost nothing — positions, light, season and dirt differ, so most of the difference is how the photos were taken.

• Register first: align both years’ images to the 3D model or member coordinates.
• Compare defects, not pixels: build a defect map each year and compare the maps.
• Match conditions: same route, same stand-off, similar light. Reuse the capture plan every year.
• Attach error to change: if “widened by 0.1 mm” is smaller than the measurement error, it is not a change.

AI makes two kinds of mistake: false positives (flagging what is not there) and false negatives (missing what is). In inspection, misses are the dangerous ones. So systems are usually tuned with a lower threshold to reduce misses, letting human review remove the false alarms.

Equally important are images the AI never got a fair look at: blurred, too dark, or never taken. Defects there produce silence. A coverage check showing what was not captured matters as much as detection itself.

AI speeds up searching, measuring and comparing. Rating, action and explanation stay with the responsible inspector — something unlikely to change under any country’s inspection regime. Build the human–AI division into the workflow.

• Review screen: show each AI candidate alongside the original image, location and last year’s image, so a person can accept, reject or edit.
• Feed corrections back: human edits become the next training data (human-in-the-loop learning).
• Record the model version: log which model version and threshold produced each result.
• Guard against over-reliance: humans sample some images the AI did not flag.
• People sign: condition ratings and actions carry the responsible engineer’s name.

For how AI findings become a report an engineer can sign, see writing up drone and AI findings in a facade inspection report (Airffic Flight Institute).

Comparing over years requires every image to carry, in machine-readable form, which structure, which member, which spot, and how it was captured. Without it, neither AI nor digital twins can accumulate.

Fix member naming at the start, aligned with the inspection standard and drawings. Common formats exist, such as BIM/CIM in Japan and buildingSMART’s IFC (extended to bridges, roads, railways and ports with IFC 4.3 in 2024). Formats change, so never losing the content — IDs, poses, conditions — matters more than the format.

By 2035, inspection images will get member IDs and conditions the moment they are taken; AI proposes candidates and only changes from last time reach the engineer’s screen. People interpret the changes, verify them on site and sign their judgements. AI keeps learning from human corrections, and its performance per structure type and region is published for comparison. This is the editors’ view.

What matters most there is not flashy AI but unglamorous data conventions: naming, recording conditions, logging mistakes. The structures where this is kept up every year are the ones where inspection truly gets easier.