The pixel gate is necessary but not sufficient on its own. A companion check on a held-out labeled set — false alerts at or below 5% across at least 1,000 labeled crack images — is what turns a sampling claim into a detection claim, and both checks belong in front of any 0.1 mm acceptance decision.
What the available sources proves, and does
None of the named sources in this article's reference list — Cong and Xiao's ICANN14 paper on CNN computation optimization, Vaswani et al.'s attention paper, or the causal CNN FIR study — contains a bridge-camera field dataset or a false-alert curve. None measures a 0.1 mm crack in the field, reports a false-alert rate, or validates a camera threshold. They establish background on CNN methods only; they do not show what a bridge inspection camera can detect. To verify any 0.1 mm claim, check the vendor's or inspector's own calibration-target images and held-out evaluation records against the gates below.
| Named source | What it covers | What it leaves open for this guide |
|---|---|---|
| Cong and Xiao, Minimizing Computation in Convolutional Neural Networks, ICANN14, 10 pages | CNN computation optimization | No bridge-crack camera threshold, no false-alert rate |
| Vaswani et al., arXiv:1706.03762, Attention Is All You Need | Sequence transduction with attention | No field inspection data, no 0.1 mm measurement |
| Causal CNN FIR study, arXiv abstract | Quasi-linear activations in causal CNNs for time-series data | No bridge-crack field data, no false-alert curve |
What remains open is the evidence that would turn a 0.1 mm statement into a detection claim: a measured projected pixel footprint at the worst-case standoff and a held-out labeled evaluation. The named CNN and attention papers supply neither. A methods paper can reduce computation or improve sequence modeling, but it does not demonstrate that a crack of a stated width was resolved, localized, and separated from non-crack texture under field standoff, lighting, and surface variation — only a calibration target and a held-out labeled set from the actual deployment can.
Dual-gate beats single-threshold claims
A 0.1 mm bridge-crack detection claim can fail in two distinct ways: the imaging setup may be unable to resolve the target at the required standoff, or the detector may produce an alert burden that inspectors cannot trust. Evaluating either condition in isolation leaves a critical gap. A defensible acceptance rule therefore treats geometric resolvability and operational reliability as separate, mandatory gates rather than interchangeable proxies for performance.
| Option | Gate | Main failure | Verdict |
|---|---|---|---|
| 0.1 mm threshold only | Geometric footprint <= 0.1 mm | May accept unresolved 0.1 mm cracks and ignores alert burden | Reject as sole rule |
| 5% false-alert budget only | Held-out false alerts <= 5% | Can be met by missing small cracks or biasing the test set | Reject as sole rule |
| Dual gate | Footprint <= 0.033 mm AND held-out false alerts <= 5% | Requires calibration and labeled data | Winner |
The threshold-only option reduces detection to a lens-and-sensor specification and offers no evidence that the resulting alerts are actionable. The budget-only option reduces detection to a classification score and can be satisfied by missing fine cracks, tuning to favorable conditions, or evaluating on a biased sample. The dual gate closes both loopholes: it requires that the target be geometrically resolvable at the worst-case standoff and that the system demonstrate trustworthy alert behavior on data withheld from tuning.
The alert-budget half of the rule is concrete. On a held-out set of 1,000 labeled alerts, a 5% false-alert ceiling permits at most 50 false alerts (1,000 × 0.05 = 50). If 51 or more alerts are false, the claim fails the reliability gate even when the optics appear adequate. The labeled set must be held out from model development so the measured rate reflects performance on unseen cases rather than fit to the tuning data.
Before accepting any 0.1 mm claim, require documented evidence for both gates together: geometric adequacy at the worst-case standoff and a false-alert rate at or below 5% on at least 1,000 held-out labeled alerts. A claim supported by only one gate is incomplete and should be treated as a partial specification, not a defensible detection capability.
Worked audit: 0.1 mm at 2 m standoff
Checkpoint 2 tests whether the resolved signal can be trusted in practice. The audit uses 1,000 labeled crack/no-crack patches drawn from the intended inspection conditions and sets the pass line at 50 false alerts (1,000 × 0.05 = 50) to stay within a 5% false-alert budget. In the illustrative audit run, 63 false alerts exceed the budget by 13, so the 0.1 mm claim cannot be defended as a detection claim for that configuration even though the geometry checkpoint passed. To run this check yourself, count false alerts on your own held-out 1,000-patch set and compare against the 50-alert line.
| Audit item | Record | Pass condition | Result |
|---|---|---|---|
| Inputs | Sensor pixel pitch, focal length, worst-case standoff, target crack width | All four values documented for the tested setup | Pass / Fail |
| Checkpoint 1: Geometry | Calculated footprint; measured pixels across target after demosaic and compression | Footprint ≤ 0.033 mm and measured width ≥ 3 pixels | Pass / Fail |
| Checkpoint 2: False alerts | Labeled patch count; false-alert count | ≤ 50 false alerts on 1,000 labeled crack/no-crack patches | Pass / Fail |
| Overall verdict | Evidence package (images, labels, counts) | Both checkpoints pass | Detection claim defensible / Lens-geometry claim only |
To make the worksheet auditable, record the exact standoff used for the measurement, the target or crack sample imaged, and the image-processing state at the point of measurement. A change in lens, compression setting, or working distance invalidates the prior checkpoint 1 evidence and requires a new measurement at the revised worst-case standoff.
For checkpoint 2, retain the labeled patch set, the decision threshold used by the detector, and the count of false alerts produced on that held-out set. A defensible 0.1 mm claim requires both the measured 3-pixel evidence and a false-alert count at or below the 5% budget; if either record is missing or out of bounds, the claim should be reported as geometry only.
If/then gates before field use
Before any field deployment, a 0.1 mm crack-detection workflow should pass five deployment gates in sequence. Each gate is a pass/fail check tied to the imaging setup or the labeled evaluation set; failing any one of them requires downgrading the claim or correcting the setup before results are used for decisions.
| Gate | Condition | Required action |
|---|---|---|
| 1. Footprint | Projected pixel footprint > 0.033 mm at the worst-case standoff | Label the output visual screening, not 0.1 mm detection. |
| 2. Calibration target | Footprint ≤ 0.033 mm but the calibration target shows fewer than 3 pixels across a 0.1 mm line | Fail the optics gate and re-shoot or change lens/standoff. |
| 3. False-alert rate | Held-out false alerts exceed 5% on at least 1,000 labeled patches | Do not deploy the detector. |
| 4. Test-set integrity | The 5% budget is met only by dropping small-crack patches | Reject the test set and rebalance it to include representative small-crack patches. |
| 5. Worst-case consistency | Footprint or false-alert evidence is unavailable at the worst-case standoff used in the field | Treat the run as unverified and restrict outputs to visual screening until both gates are demonstrated at that standoff. |
Gates 1 and 2 must be evaluated at the worst-case standoff, not a best-case or average capture distance, because any increase in standoff degrades the effective sampling of a 0.1 mm feature. Gates 3 and 4 must be evaluated on a held-out labeled set so that the reported false-alert performance reflects generalization rather than tuning on the same patches used to fit or select the model.
When a crack indication is present but its width cannot be resolved with confidence under the verified setup, apply the indeterminate-width reporting rule: report the indication as an indeterminate-width finding, record the standoff and imaging conditions, and do not assign a 0.1 mm width label. This prevents unresolved indications from being promoted to quantitative detection claims.
Passing all five gates supports a defensible 0.1 mm detection claim for the verified conditions; failing any gate confines the output to screening or an indeterminate-width report until the deficiency is corrected and rechecked.
Frequently Asked Questions
Is passing the pixel gate alone sufficient to accept a 0.1 mm crack detection claim?
No, the pixel gate is necessary but not sufficient on its own and must be paired with a companion check on a held-out labeled set before any 0.1 mm acceptance decision.
How many labeled crack images are required for the companion held-out evaluation?
The companion check requires a held-out labeled set of at least 1,000 labeled crack images.
What evidence should be reviewed to verify a vendor's 0.1 mm crack detection claim?
To verify any 0.1 mm claim, you should check the vendor's or inspector's own calibration-target images and held-out evaluation records against the required gates.
Does Cong and Xiao's ICANN14 paper provide a bridge-crack camera threshold or false-alert data?
No, Cong and Xiao's ICANN14 paper covers CNN computation optimization and does not contain a bridge-crack camera threshold or false-alert rate.
Does Vaswani et al.'s arXiv:1706.03762 paper include field inspection data or 0.1 mm crack measurements?
No, Vaswani et al.'s arXiv:1706.03762 paper focuses on sequence transduction with attention and provides no field inspection data or 0.1 mm measurements.
Do any of the named sources in the reference list contain a bridge-camera field dataset or a false-alert curve?
No, none of the named sources contain a bridge-camera field dataset or a false-alert curve, as they establish background on CNN methods only.
Quick answers
| Is the pixel gate alone sufficient for a 0.1 mm acceptance decision? | The pixel gate is necessary but not sufficient on its own. |
| What companion check is required alongside the pixel gate? | A companion check on a held-out labeled set, with false alerts at or below the specified threshold across at least 1,000 labeled crack images, is required to turn a sampling claim into a detection claim. |
| Do the named sources in the article provide a bridge-camera field dataset or a false-alert curve? | None of the named sources in the article's reference list contains a bridge-camera field dataset or a false-alert curve. |
| What do the named sources establish? | The named sources establish background on CNN methods only and do not show what a bridge inspection camera can detect. |
| How can a 0.1 mm claim be verified? | To verify any 0.1 mm claim, one should check the vendor's or inspector's own calibration-target images and held-out evaluation records against the gates. |
Also worth reading: High early strength bridge concrete: 3,000 psi at 3 days, set gates before pricing: High early strength bridge concrete: · Neural networks vs finite elements: 10,000 points—not a drift test: Neural networks vs finite elements: · Evaluating AI for structural safety and construction efficiency: Evaluating AI for structural safety