Computer vision inspection does not retire the manual inspection station. It absorbs the defect classes where imaging physics, fixturing and repeatability allow a measured accuracy, and it leaves the rest — judgement calls, novel defect modes, rework adjudication — with human inspectors. Any business case that sizes savings against the full manual inspection cost base has already made an arithmetic error, and the error usually surfaces months later on the line rather than in the spreadsheet.
That is the uncomfortable half of a feasibility audit. The same profiling that issues a go decision per defect class also produces a no-go list, and that list is a deliverable in its own right.
What “CV does not replace all manual inspection” means on a production line
It means the inspection workload splits, and the split is defect-class-shaped rather than station-shaped.
A human inspector standing at the end of a line is not performing one task. They are detecting known defects, noticing things they have never seen before, deciding whether a marginal part is scrap or rework, sampling for the quality system, and answering the line supervisor’s questions about a batch. A vision system takes the first of those five and only for the classes it was scoped, trained and validated on. The other four stay where they are.
So the honest framing is not “the station is automated” but “a measured share of inspection minutes moved”. The residual manual load is a number you can estimate before the pilot — and if nobody has estimated it, the savings figure in the proposal is unanchored.
The failure mode is not that CV misses a defect. It is that nobody decided in advance who catches the defects CV was never scoped to catch.
Where the two approaches diverge: the first unfamiliar defect
Watch what happens when a defect appears that is not in the catalogue.
In a scoped deployment, the part is routed to a human inspector by design. The system flags it as low-confidence or out-of-distribution, the inspector adjudicates, and the incident is recorded as a candidate new class with images attached. That record is the input to the next retraining cycle. The line keeps running and the gap is visible in the quality data.
In an unscoped deployment, the part passes. Not because the model failed at its job — it did exactly what a classifier does with an input outside its training distribution — but because the deployment assumed the classifier’s job was “inspection” rather than “these seven defect classes at this confidence threshold”. A detector trained on scratches and edge chips has no representation for a contamination class it never saw; whether you run YOLO variants, a segmentation model behind an ONNX Runtime or TensorRT deployment, or a classical OpenCV rule chain, the behaviour is the same. Confidence scores are not calibrated uncertainty about the world, only about the label space they were given.
We see this pattern regularly when reviewing pilots that were declared successful: the aggregate detection rate looks fine, and the defects that escaped are the ones nobody counted because no one was looking for them.
Which defect classes move, and which stay
The boundary is drawn on physical and organisational properties, not on ambition. Broadly:
| Inspection work | Typically moves to CV | Typically stays manual | Why |
|---|---|---|---|
| Known, repeatable surface defects (scratches, missing print, contamination above resolution limit) | Yes | — | Stable defect signature, resolvable at line pixel density, controllable lighting |
| Dimensional checks against a fixed reference | Yes, with fixturing | — | Requires repeatable part presentation; feasibility follows the fixture, not the model |
| Sub-surface or occluded defects in visible-light imaging | — | Yes, or a different modality | The image never renders the defect; no model recovers information the optics did not capture |
| Novel or first-occurrence defect modes | — | Yes | Outside the trained label space by definition |
| Scrap-vs-rework adjudication on marginal parts | — | Yes | A cost judgement, not a detection task |
| Root-cause triage after a defect cluster | — | Yes | Requires process context the vision system does not hold |
| Audit and sampling duties required by the quality system | — | Yes | Procedural obligation, unaffected by detection capability |
The middle two rows are where most disappointment lives. A team commits to a headcount reduction on the strength of the top rows, then discovers that adjudication and triage were consuming a substantial share of the inspector’s shift all along. The classification of individual defects on contrast, size and surface-visibility grounds is worked through in which defect classes are feasible for CV inspection today; the point here is narrower — that the infeasible list has to become an operational plan, not a footnote.
What the cost model should actually contain
Rewrite the inspection cost model in three lines rather than one:
- Absorbed load — the share of inspection minutes CV can take at the buyer’s accepted false-positive ceiling, stated per defect class rather than blended.
- Residual manual load — the remaining share, expressed as a percentage of the pre-project baseline, with the associated headcount explicitly unchanged.
- New manual load — false-positive rework hours at the throughput the line must hold, plus adjudication of low-confidence routings.
The third line is the one most often omitted. A false-positive rate that is acceptable in a lab is a queue of parts and an annoyed operator at line rate. Our practice is to make the buyer name their false-positive ceiling before any accuracy target is discussed, because the ceiling determines the threshold, and the threshold determines the detection rate you can honestly promise. Total-cost comparison across the whole deployment — capex, integration, licence — is a separate exercise covered in when CV inspection beats manual inspection on total cost.
Three measurable outputs make the model defensible: detection rate per in-scope class, false-positive rate at line throughput, and residual manual inspection load as a percentage of baseline. If a proposal quotes one accuracy number and no residual load, it has not been scoped.
Designing the handoff instead of discovering it
The manual residue is an interface, and interfaces get designed. In practice this means the vision system emits three outcomes rather than two: pass, fail, and refer. The refer path needs a physical destination — a reject lane, a marked tote, a station — and a recording obligation, so that every referred part produces an image, a timestamp and an inspector verdict. That log is the only honest source of the true out-of-scope rate, and it is what turns the next retraining cycle into evidence-led work rather than guesswork.
Confidence thresholds should be set so that the refer volume is something the remaining inspectors can actually absorb during a shift. This is a staffing constraint expressed as a model parameter, which is why it belongs in the audit and not in a later tuning session. Our computer vision engineering practice treats the escalation path as part of the delivered system, and the same discipline shows up in how we scope R&D engagements generally: the boundary of what is being automated is written down before the work starts.
How to read a vendor proposal
An honest scoped proposal and a full-automation claim are easy to distinguish once you know what to look for:
- Does it list defect classes individually, with a verdict per class?
- Does it state a false-positive rate at the line’s throughput, not on a curated set?
- Does it name the classes that remain manual, and say why?
- Does it specify what happens to a part the system has no class for?
- Does it quantify residual manual load, or does it quote gross headcount savings?
A proposal that answers all five is scoped. One that answers none is selling a station replacement that will not arrive. The structural reasons vendor accuracy figures fail to transfer to a live line are examined separately in why most CV inspection pilots over-promise on accuracy.
The no-go list is not an admission that the technology underperformed. It is the part of the deliverable that keeps the go list credible — and the question worth putting to any vendor early is a simple one: which defects on this line will your system never see, and who is standing there when one arrives?
Frequently Asked Questions
What does “CV inspection does not replace all manual inspection” mean in practice on a production line? Computer vision inspection augments rather than eliminates human quality control, handling repetitive defect detection while operators manage edge cases and final validation. It means the inspection workload splits by defect class rather than by station. A vision system takes the known, imageable, repeatable classes it was trained and validated on; detection of novel defects, scrap-versus-rework judgement, root-cause triage and quality-system sampling remain with human inspectors.
Which defect classes realistically move to CV, and which stay with human inspectors? Classes with a stable visual signature, sufficient contrast against the substrate and a size resolvable at the line’s pixel density typically move, provided fixturing gives repeatable part presentation. Sub-surface or occluded defects, first-occurrence defect modes, and any task that is a cost judgement rather than a detection task stay manual.
How do we design the handoff between the vision system and the manual inspection station? Give the system three outcomes instead of two — pass, fail, and refer — and give the refer path a physical destination plus a recording obligation. Set confidence thresholds so the referred volume is absorbable by the inspectors actually on shift, and log every referral with images and the inspector’s verdict.
How should the inspection cost model change once the residual manual load is accounted for? Split it into absorbed load, residual manual load, and newly created load from false-positive rework and adjudication. Savings are sized against the absorbed share at the accepted false-positive ceiling, never against the full manual cost base, and the residual share is stated as a percentage of the pre-project baseline with headcount unchanged.
What happens when a defect appears that was never in the CV system’s defect catalogue? In a scoped deployment it is routed to a human inspector as low-confidence or out-of-distribution and recorded as a candidate new class. In an unscoped one it usually passes silently, because a model has no representation for a class outside its training distribution and confidence scores describe the label space rather than the world.
How do we tell an honest scoped vision proposal from a vendor claim of full inspection automation? Check whether it gives a per-class feasibility verdict, states false-positive rate at line throughput rather than on a curated set, names the classes that stay manual and why, specifies the out-of-catalogue path, and quantifies residual manual load. Gross headcount savings with a single blended accuracy figure is the tell.
Deploying CV Inspection Without Overreach
Start with high-volume, repeatable defect classes where ground truth is abundant and consequences of a miss are recoverable. CV Inspection rewards teams that measure first and argue later — start with the smallest instrumented slice and let the numbers settle the design.
