Automated defect detection can make a large inspection dataset easier to navigate. A model may identify regions resembling coating breakdown, corrosion, cracks or fouling, depending on its design and training. The output is a proposed interpretation of pixels. It is not automatically a diagnosis of structural condition, a severity classification or an instruction to repair.
Keep the workflow divided into detection, assessment and decision. Detection creates a candidate finding. Assessment tests that finding against the available evidence and the inspection context. The maintenance decision then considers consequence, applicable requirements, access, operational risk and the responsible person’s authority. Combining these stages into a single label hides important judgement.
Ask what the model was taught to recognise
Model labels are not universal engineering definitions. One dataset may group rust staining and corrosion together; another may distinguish them. A crack-like line could be a weld edge, a coating boundary, a shadow or a genuine discontinuity. Review the label definitions, examples and exclusions before comparing performance claims or interpreting a detection report.
Training conditions matter too. Results obtained from well-lit dry-dock photographs may not transfer to underwater imagery affected by growth, turbidity and artificial lighting. Different coating colours, camera systems and vessel geometries can also change performance. A useful evaluation resembles the intended deployment, with independently reviewed examples that were not used to tune the model.
Read the limitations beside the result
Detection systems produce false positives and false negatives. Reducing nuisance flags can also reduce sensitivity to genuine findings. There is no universally correct threshold: the consequences of a missed finding and the capacity for follow-up review both matter. Report the operating threshold and keep it associated with the model version and dataset.
| Observed pattern | Possible interpretation | Needed before a stronger conclusion |
|---|---|---|
| Rust-coloured staining | Possible corrosion-related surface condition | Context, surface examination and measurements where required |
| Thin dark line | Possible crack-like indication, edge or shadow | Additional views and suitable examination by qualified personnel |
| Missing coating | Possible coating breakdown or abrasion | Verified extent, substrate condition and coating history |
| No model flag | No recognised pattern at the selected threshold | Coverage checks, quality review and appropriate human sampling |
Do not treat a low flag count as proof of good condition. The camera may not have seen the relevant feature, the image may be unsuitable or the model may not recognise that defect class. Coverage and image quality are prerequisites for interpreting an apparently quiet detection report.
Give the reviewer enough context
A review screen or evidence package should show the original image, surrounding views, vessel location and capture conditions. The reviewer should be able to inspect the source without an annotation covering the feature. If video is involved, the adjacent sequence can help distinguish persistent surface detail from reflections or moving debris.
Record a disposition such as confirmed visual indication, rejected flag or unresolved finding. Explain why. “Rejected because shadow changes with camera position” is more useful than an unexplained deletion. An unresolved finding should remain visible with its required next step; it should not disappear simply because the software requires a binary answer.
“An unresolved indication belongs in the record, not outside the workflow.”
Puzzlewhistlereel — review principle
Escalate the question, not just the image
Where imagery cannot answer the condition question, specify the missing evidence. Remaining thickness may require suitable thickness measurements. A suspected surface-breaking discontinuity may require an appropriate non-destructive testing method. Selection, execution and interpretation of those methods belong with competent personnel and the applicable inspection arrangements.
ISO 9712 concerns qualification and certification of non-destructive testing personnel; it does not turn an image classifier into a certified inspector. Likewise, a visual model cannot establish compliance with a vessel’s class rules merely by using familiar defect terminology. The relevant acceptance criteria must be identified separately and applied by the responsible parties.
Evaluate the complete review process
Assess more than the detector’s headline score. Consider whether reviewers can reproduce a finding’s location, distinguish duplicates, retrieve originals and document disagreement. Keep examples of false positives and missed findings for controlled improvement, with permissions and data handling agreed in advance. Changing a model should not silently rewrite the historical record.
The useful output is a bounded statement: what was observed, how confidently it can be interpreted, what remains unknown and who must act next. That statement becomes durable only when it is connected to the vessel’s inspection and maintenance history.
Reference points
Find current standards through ISO, including ISO 9712. Consult applicable class requirements through Bureau Veritas, ClassNK or the vessel’s own classification society. These organisations do not endorse this website.