TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 18—20 · FACTUAL · EVOLVING
Looking at decay with a machine
The machine looks everywhere, and a trained eye on site decides what it found.
When to use
A building that must be looked after rather than redesigned: a condition survey before a conservation plan, a periodic check on a monument, a comparison of this year's photographs with last year's. Use the machine to look everywhere. Use a trained eye to decide what it found.
The method
Photograph systematically: every bay, every elevation, the same positions and the same time of day on every visit, with the positions kept in a register, so that this year's set can be laid over last year's. Run a crack or defect detector over the set. Research models trained on masonry and monument imagery can flag and outline cracks on stone walls, and general vision models can sort photographs by visible damage. Treat the output as a list of places to look, not as a diagnosis. Compare across time: align this visit's photograph with the last one and ask the machine for differences. Change is one of the things a condition survey looks for — beside present condition, severity, cause and urgency — and change is what a machine compares well, provided the two photographs are registered: same position, same lens, similar light, a dry wall both times. A shifted viewpoint, a wet patch or new vegetation reads as change. Then go to site with the list. A trained eye decides whether a flagged line is a structural crack, a mortar joint, a lichen edge or a shadow, and whether a stain is rising damp, a leaking pipe or last week's rain; record that judgement against each flag. Write the survey as a human document — condition, cause, urgency — with the machine's output as an appendix and its error rate stated.
Watch for this
A detector trained on one kind of masonry flagging everything that resembles its training set. A laterite wall's natural pitting, the crazing of old lime plaster, a chisel mark on a basalt pillar can all come back as "crack". You do not know the false-alarm rate on your building until you count it on your building. And the crack it did not flag is the more dangerous error.
Try it
Photograph one old wall — a compound wall, a temple plinth, a railway building — in fifty frames. Run any crack-detection or image-tagging tool over them. Check every flag on site with a ruler and a note. Count true flags, false flags, and cracks you found that it missed. Those three numbers tell you how the tool behaves on your wall.
Prove it
Show one wall's machine-flagged image beside your site notes, with every flag judged — crack, joint, stain, shadow — the three counts stated, and one flagged item you overruled, with the reason.
How it works
Crack detection is a vision task of the kind VISION on this map describes: a network trained on labelled images of damaged and undamaged surfaces learns to classify patches or outline pixels. Research on stone masonry has treated it as anomaly detection: learn what the undamaged surface looks like, then flag anything that departs from it. Part of the reason is that heritage rules often forbid the test interventions other surveys rely on, so a non-destructive method has real value. The strength is coverage: a machine looks at every square metre of a façade with the same attention. The weakness is meaning: it sees a line, not a cause. Cause — movement, water, salts, vegetation, an earlier bad repair — is diagnosed by someone who knows the construction, on site, often by touch and by tapping. SCAN TO MODEL is the geometry this survey sits on; WHERE A HUMAN SIGNS applies to the survey as it does to the drawing. EVOLVING because the models and the public datasets are improving quickly — segmentation models now generalise to monument surfaces they were never trained on — so the boundary between what the machine can flag and what the eye must decide moves with each edition. The signature does not.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.