“Can vision inspect this housing?” is not an answerable question. The part is not the unit of feasibility; the defect class is. Ask it per part and you get a coverage promise nobody can hold. Ask it per defect class and you get a scope you can defend on the line.
The single most useful artefact in a CV inspection scoping conversation is a list of defect classes with a verdict beside each one — feasible, marginal-with-conditions, or manual-only. Not a part list. Not a blended accuracy figure. A class-by-class map.
We see the same failure repeatedly: a pilot is scoped from a vendor accuracy number measured on a curated dataset, the buyer reads it as coverage of everything the camera can see, and three defect families quietly ride along inside that promise. Low-contrast cosmetic defects. Sub-surface porosity. Rare-but-critical classes with a handful of confirmed examples. Those three survive slide decks and break on the conveyor.
What physical properties decide whether CV can detect a defect at all?
Detectability is set before any model runs. Five properties do most of the work, and every one of them is measurable during a site survey rather than argued about in a meeting.
Contrast against the substrate. A defect is detectable when it produces a reliable signal difference under the illumination the line can actually provide. A scratch on brushed aluminium under diffuse plant lighting may produce almost no radiometric difference; the same scratch under grazing directional light produces a strong one. Contrast is a property of the defect and the lighting geometry together, which is why the imaging conditions have to be measured, not assumed — the pre-pilot site survey exists for exactly this.
Characteristic size relative to achievable pixel resolution. The working rule we use when bounding scope: a defect needs to span several pixels in its smallest dimension before a detector can separate it from sensor noise and compression artefacts, so the honest question is not “how many megapixels” but “how many pixels per millimetre at the working distance and field of view the fixture allows.” A 0.2 mm crack on a 400 mm field of view is a resolution problem, not a model problem.
Surface-visible versus sub-surface. Optical inspection sees surfaces. Internal porosity in a casting, delamination under a coating, and voids beneath a solder joint are not visible to a visible-light camera at any resolution — they need a different modality (X-ray, ultrasound, thermography) or they stay manual. This is the cleanest no-go boundary in the whole rubric, and it is the one most often blurred in vendor conversations.
Orientation and presentation consistency. A defect that only presents at certain part poses is detectable at the accuracy the fixture allows, not the accuracy the model allows. If the part arrives with ±15° rotational freedom and the defect signature disappears at half those angles, either the fixturing changes or the class drops to marginal.
Labelled example count. No accuracy band can be quoted for a class that has too few confirmed instances to hold out a test set. Below roughly the low tens of confirmed examples per class, we will describe behaviour qualitatively but will not put a number on it — quoting a detection rate from a dozen samples is arithmetic dressed as evidence.
The feasibility map
This is the rubric. Each defect class from the buyer’s own QA records gets scored on the five properties above, then lands in one of three verdicts.
| Verdict | Conditions the class satisfies | What the buyer gets | What it costs to keep |
|---|---|---|---|
| Feasible | Surface-visible; strong, repeatable contrast under achievable lighting; characteristic size resolvable at the line’s field of view; consistent presentation in the fixture; enough confirmed examples to hold out a test set | Quoted accuracy band with the false-positive rate at line throughput | Ongoing labelling of new instances; drift monitoring across shifts |
| Marginal-with-conditions | Detectable only if a named condition changes — added directional lighting, tighter fixturing, reduced belt speed, higher-resolution optics, or more labelled examples | A conditional verdict naming the specific change and the accuracy it would unlock | Capex or cycle-time cost of the condition, decided before the pilot |
| Manual-only | Sub-surface with no optical signature; below achievable resolution; presentation too variable to fixture economically; or a judgement call (acceptability, cosmetic grading against a customer standard) | Explicit exclusion from scope, so the ROI model counts it as retained manual cost | The inspection labour that stays — which is the honest baseline |
Three families deserve naming because they are the ones that get misfiled most often.
Low-contrast cosmetic defects — faint scuffing, mild colour drift, subtle surface finish variation — are usually marginal rather than infeasible. They become feasible when lighting geometry is redesigned around them, and they stay marginal when the same station has to serve several defect classes with conflicting illumination requirements.
Sub-surface porosity and internal voids are manual-only for visible-light CV. Full stop. Moving them to a different modality is a legitimate engineering answer; pretending a better model solves them is not.
Rare-but-critical classes are the hardest case, because the cost of a miss is high and the example count is low. The wrong move is to bundle them into an aggregate accuracy figure where a handful of instances cannot move the number. The workable pattern is to run them as an anomaly or novelty channel with a deliberately low threshold — high recall, tolerated false positives — and route every flagged part to a human. That is not detection with a quoted rate; it is triage with a measured referral volume, and it should be described that way in the deliverable.
Why scoping per class changes what ships
A project scoped per defect class ships a system that clears the classes it was proven on and routes the rest to manual inspection. A project scoped per part commits to coverage that the marginal and infeasible classes will break — usually at the first quality escape, and usually in a meeting where the accuracy figure everyone agreed to turns out to have been a blend across classes with wildly different detectability.
The number the ROI model actually turns on is not the headline detection rate. It is the percentage of inspection volume that stays manual. A system that handles 70% of volume at a measured, per-class detection rate with a known false-positive load is a defensible investment. A system claiming 98% accuracy across a part with no class breakdown is unpriceable, because nobody knows which 2% it is or what it costs when missed. We work through the structural reasons this gap persists in our practical guide to computer vision inspection feasibility for manufacturing, and the profiling discipline that bounds achievable resolution and accuracy per class is the same one our computer vision engineering practice applies to imaging-chain work generally.
One practical consequence for the pilot: the classes marked feasible determine how many confirmed defect instances the pilot has to collect, which in turn determines its duration and the shift coverage it needs. Feasibility scoping and pilot sizing are a single chain, and running them out of order is how pilots end up statistically unable to prove anything.
Frequently Asked Questions
What does ‘which defect classes are feasible for CV inspection today and which aren’t’ mean in practice for a specific production line?
A common Defect Classes Feasible CV question is worth clarifying. It means walking the line’s own QA defect records, listing each class separately, and scoring each on contrast, size versus achievable pixel resolution, surface versus sub-surface, presentation consistency, and confirmed example count. The output is a verdict per class, not a verdict per part. Classes that fail on imaging physics are excluded before any model is selected.
What physical properties of a defect — contrast, size, surface vs sub-surface, orientation consistency — decide whether CV can detect it at all?
Contrast under the illumination the line can actually supply, characteristic size relative to pixels-per-millimetre at the required field of view, whether the defect has any optical signature at the surface, and whether it presents the same way each time the part is fixtured. All four are measurable before a pilot. A defect that fails on any one of them is not rescued by a better model.
Which defect families are reliably feasible today, which are marginal-with-conditions, and which should stay manual?
Surface-breaking, well-contrasted, adequately sized defects with consistent presentation and a decent example count are reliably feasible. Low-contrast cosmetic defects and orientation-sensitive classes are typically marginal — feasible only if lighting, fixturing, or resolution changes. Sub-surface porosity, defects below achievable resolution, and acceptability judgement calls stay manual.
How many labelled examples does a defect class need before an achievable accuracy band can be quoted for it?
Enough to hold out a test set that spans the line’s real variance in lighting, pose, and shift conditions. Below roughly the low tens of confirmed instances per class we describe behaviour qualitatively and decline to quote a rate, because a detection figure computed on a dozen samples carries a confidence interval wide enough to be meaningless.
How do we handle rare-but-critical defect classes where examples are scarce but the cost of a miss is high?
Treat them as a novelty or anomaly channel rather than a classifier with a quoted accuracy. Run a deliberately low threshold to favour recall, accept the false-positive volume, and route every flagged part to a human inspector. Report the referral rate and the inspection labour it implies — that is the measurable commitment, not a detection percentage.
What does a per-defect-class feasibility map look like as a deliverable, and how does it feed the pilot scope and the ROI model against manual inspection?
It is a table: one row per defect class, with the verdict, the achievable accuracy band and false-positive rate at line throughput for feasible classes, and the named condition for marginal ones. The feasible rows set the pilot’s defect-collection targets and duration. The manual-only rows set the retained manual inspection cost that the ROI model has to carry.
Which defect classes on your line have never been catalogued separately, and what would you find if you measured contrast on each of them under the lighting you actually have? That, more than any model choice, is where the scope of a CV inspection programme is really decided — and our engineering engagements start there rather than at the model.
Defect Classes Feasible CV: the decision
Defect Classes Feasible CV is rarely the hard part — knowing which of its failure modes you can live with is. That answer is workload-specific, and it is worth writing down before you build.
