A vendor demo shows 99% defect-detection accuracy on a clean sample set, the procurement case gets written against that number, and the pilot is scheduled before anyone has asked whether the production line can reproduce the conditions that produced it. Three months later the pilot is running at a false-positive rate that pulls operators off the line to clear good parts, and the inspection cost is higher than the manual workflow it was meant to replace.

The decision that matters is not “which CV vendor” or “which model architecture.” It is whether to run the pilot at all. A CV-inspection feasibility audit answers that question before any capital is committed: it catalogues your defect classes, profiles the lighting, fixturing, and conveyor constraints on your actual line, and produces a go/no-go decision document for each defect class. The audit is cheap relative to a pilot. A pilot that profiling proves infeasible is the expensive mistake.

Why the Vendor Accuracy Number Doesn’t Decide This

Off-the-shelf computer vision models report accuracy against curated datasets — well-lit, centred, consistently presented parts. Your line is not curated. Lighting drifts across a shift, parts arrive at varying orientation and registration on the conveyor, surface finish varies batch to batch, and the defect that matters to your quality team may occupy a few dozen pixels under glare. The gap between demo accuracy and line accuracy is not noise; it is the structural problem that decides feasibility.

We see this pattern regularly: a model that scores in the high nineties on a vendor’s test set lands in the seventies or low eighties once it meets uncontrolled lighting and the long tail of real defect presentations. This is the same wall described in where off-the-shelf CV stops working in vision-QC — the failure is not the model, it is the assumption that benchmark accuracy transfers to an environment the benchmark never saw.

There is a second, quieter failure. Some defect classes are simply not feasible for vision today on a given line — subsurface flaws that produce no visible signal, defects whose appearance overlaps entirely with acceptable cosmetic variation, or classes so rare you cannot assemble enough labelled examples to train against. Routing these into a pilot does not produce a worse model; it produces no model at all. A feasibility audit separates the defect classes that are achievable from the ones that aren’t, before the pilot budget is spent finding out the hard way.

What a CV Inspection Feasibility Audit Actually Delivers

The audit is a GPU performance audit scoped to vision-pipeline feasibility on the production line. It is not a pilot and it is not a proof of concept. It is a decision document. Four things come out of it.

A defect-class catalogue enumerates every defect type your quality team cares about and assigns each one a feasibility verdict: feasible now, feasible with hardware or fixturing changes, or not feasible with current vision methods. This is the core deliverable — a defect-class-by-defect-class feasibility map rather than a single line-level “yes” or “no.”

A constraint profile documents the lighting, fixturing, and conveyor conditions on the actual line and which of them make or break detection for each feasible class. This is where most of the engineering judgement lives, because the achievable accuracy band for a class depends almost entirely on whether you can control its imaging conditions.

An achievable accuracy band per defect class, expressed as a range under stated conditions rather than a single headline number. The band is honest about variance: it tells you what detection rate you can expect at an acceptable false-positive rate, and what the rate degrades to when conditions drift.

A total-cost-of-ownership model that compares the calibrated CV inspection cost — capital, integration, ongoing false-positive handling, retraining — against the existing manual inspection workflow. CV does not always win this comparison, and the audit’s job is to say so before, not after, you have built the system.

Decision Rubric — Should This Defect Class Go to Pilot?

Score each defect class. Use it as a go/no-go filter before committing pilot budget.

Question	Pilot-favourable	Pilot-hostile
Is the defect visible to a camera under achievable lighting?	Clear visible signal, separable from background	Subsurface, or overlaps acceptable cosmetic variation
Can lighting and fixturing be controlled on the line?	Yes — enclosure, fixed optics, registered presentation	Open line, ambient drift, variable orientation
Do enough labelled examples exist to train and validate?	Hundreds-plus across presentations	Rare class, few labelled instances
What false-positive rate is acceptable at line throughput?	Tolerant — rework station absorbs it	Tight — every false positive stops the line
Does CV beat manual on total cost for this class?	Manual is slow, costly, or inconsistent here	Manual is cheap, fast, and already good enough
Is the achievable accuracy band above the quality gate?	Yes, with margin under drift	Only at demo conditions, not under drift

A class that lands pilot-favourable on most rows is worth a pilot. A class that lands pilot-hostile on the imaging or training rows is a no-go regardless of how the cost row reads — you cannot cost-justify a model that cannot see the defect.

Which Defect Classes Are Feasible for CV Today?

The feasible classes share a property: the defect produces a stable, visible signal that can be separated from acceptable variation under conditions you can control. Surface defects with clear contrast — cracks, scratches, missing components, misprints, dimensional deviations against a fixed reference — sit firmly in feasible territory when imaging is controlled. These are the classes where a properly trained model, often built on PyTorch or exported through ONNX to a TensorRT runtime for line-side inference, reaches a detection rate that holds up under realistic drift.

The hostile classes are the ones where the signal is weak, ambiguous, or absent. Subsurface flaws with no visible surface correlate. Defects whose appearance is indistinguishable from acceptable cosmetic variation without context the camera does not have. Extremely rare classes where the labelled-example count is too low to train against and too low to validate. For these, the honest verdict is not “we need a better model” — it is “vision is the wrong tool for this class on this line.” Naming that early is the audit doing its job.

Most lines are a mix. The realistic output of a feasibility audit is rarely a blanket yes or no; it is a partition — some defect classes greenlit for pilot, some deferred pending fixturing or hardware changes, some routed back to manual inspection because vision cannot economically catch them. That partition is the decision the audit exists to produce.

How False Positives Trade Off Against Throughput

The false-positive rate is the constraint that quietly decides whether CV inspection saves money. A model tuned for maximum defect-detection rate will flag more good parts; a model tuned to leave good parts alone will miss more defects. Where you set that operating point depends entirely on what a false positive costs on your line.

If a flagged part routes to a rework station that an operator clears in seconds, the line can absorb a higher false-positive rate, and you can tune toward catching more defects. If a flagged part stops the line or pulls an operator off another task, every false positive carries a throughput cost, and the acceptable false-positive ceiling drops sharply. The point where CV inspection costs more than it saves is the point where false-positive handling consumes more labour than the manual inspection it replaced. That ceiling is specific to your line, and the audit measures it against your throughput rather than assuming a vendor’s operating point.

This is an observed pattern across the assembly lines we have worked with rather than a published benchmark: teams routinely under-budget false-positive handling because the vendor demo reported the detection rate and not the false-positive cost at production throughput. The two numbers belong together, and the audit reports them together.

When CV Beats Manual Inspection on Total Cost — and When It Doesn’t

CV inspection wins on total cost when manual inspection is slow, expensive, inconsistent, or unable to keep pace with line speed — and when the defect classes that drive cost are feasible classes. It loses when manual inspection is already cheap, fast, and reliable for the defects that matter, or when the feasible defect classes are not the ones causing your quality cost.

The calibrated ROI model in the audit is the deciding instrument. It anchors against three measurable quantities: the defect-detection rate on your defect set (not a vendor’s), the false-positive rate at your acceptable throughput, and the avoided cost of a pilot that profiling proves infeasible. That third quantity is easy to overlook and often the largest. A feasibility audit that returns a clear no-go has saved the cost of a pilot that would have failed — and that avoided cost is a real return, not a sunk one. The reasoning here mirrors what a performance and porting assessment tells you before you commit to a migration: the cheapest way to be wrong about a system is to find out before you build it.

FAQ

Which defect classes are feasible for CV today and which aren’t?

Feasible classes produce a stable, visible signal separable from acceptable variation under controllable imaging — surface cracks, scratches, missing components, misprints, dimensional deviations against a fixed reference. Hostile classes have weak, ambiguous, or absent visible signal: subsurface flaws, defects that overlap acceptable cosmetic variation, and classes too rare to train or validate against. Most lines are a mix, and the audit partitions them rather than returning a single yes or no.

What lighting, fixturing, and conveyor conditions make or break a CV inspection?

The achievable accuracy band for a defect class depends almost entirely on whether you can control its imaging conditions. Stable, repeatable lighting, fixed optics, and registered part presentation push a class toward feasible; ambient lighting drift, variable orientation on the conveyor, and uncontrolled surface finish push it the other way. The constraint profile in the audit documents these conditions on your actual line and ties each one to the detection rate it enables or destroys.

How do we size a CV inspection pilot before scaling?

You size it against the defect-class feasibility map: pilot only the classes the audit greenlights, at the false-positive ceiling your line throughput can absorb, with enough labelled examples to train and validate each class. The audit produces the sizing — defect set, achievable accuracy bands, and the operating point — so the pilot tests a decision that has already passed a feasibility filter rather than discovering infeasibility mid-pilot.

When does CV beat manual inspection on total cost, and when doesn’t it?

CV wins when manual inspection is slow, expensive, inconsistent, or unable to keep line speed, and when the cost-driving defect classes are feasible for vision. It loses when manual is already cheap, fast, and reliable for the defects that matter, or when the feasible classes aren’t the ones causing your quality cost. The calibrated ROI model in the audit decides this against your defect-detection rate, your false-positive rate at acceptable throughput, and the avoided cost of an infeasible pilot.

What does a CV inspection feasibility audit deliver?

Four things: a defect-class catalogue with a feasibility verdict per class, a constraint profile of the lighting, fixturing, and conveyor conditions that make or break each feasible class, an achievable accuracy band per class stated as a range under defined conditions, and a total-cost-of-ownership model comparing calibrated CV inspection cost against the existing manual workflow. The output is a go/no-go decision document per defect class, not a pilot or a proof of concept.

How do false-positive rates trade off against throughput on a production line, and what false-positive ceiling is acceptable before CV inspection costs more than it saves?

A model tuned to catch more defects flags more good parts; the acceptable false-positive ceiling depends on what a false positive costs on your line. If flagged parts clear at a rework station in seconds, the line absorbs a higher rate; if a false positive stops the line or pulls an operator away, the ceiling drops sharply. CV stops saving money at the point where false-positive handling consumes more labour than the manual inspection it replaced — a ceiling specific to your line and measured against your throughput, not a vendor’s operating point.

Where This Decision Lives

The feasibility audit is upstream of everything else in industrial computer vision. If it greenlights a pilot, the next problem is hardening that pilot for the realities of the line — drift, variance, and the long-tail defects that only appear at production volume — which is the subject of how CV defect-detection models survive the move from pilot to production line, and the reliability artefacts that keep a line-side model running once it is deployed. Those problems are worth solving only for defect classes the audit has already proven feasible.

The discipline is the same one that separates a benchmark accuracy from a line accuracy: decide what the system must actually do, under the conditions it will actually face, before you commit to building it. A computer vision inspection feasibility audit is how that decision gets made on evidence rather than on a demo number — defect class by defect class, before any pilot fires.