When manual visual inspection beats automation: an honest checklist

Automated inspection earns its keep on some lines and wastes capital on others, and the camera is rarely the difference. What decides it: whether the pass and fail rules can be written as numbers, whether the yearly inspection load is big enough to retire the system cost, and whether the floor will act on a machine's rejects. This guide gives you the arithmetic for that decision. A full-shift station pays back in about three months, and a 1-in-20 sampling job rarely pays back at all. Work the two examples with your own volumes, then run the checklist at the end.

Why are so many plants still inspecting by eye?

Manual survives because a trained inspector is hired by the hour, catches defects nobody wrote into the spec, and starts on day one — while vision rollouts stall for months on unwritten criteria, downtime risk and trust, rarely on camera capability.

That mix is rational, not stubbornness. Visual inspection is a variable cost: you add coverage when volume rises and drop it when orders fall, with no installation outage and no spares shelf. A vision station is capital: mounting, lighting, controller I/O, IT sign-off and a production window to commission. Until the yearly bill is quantified, the variable cost looks safer, and sometimes it is.

The second blocker is criteria ownership. Quality, production and the customer often disagree about how large a flaw can be and still ship, and nobody writes the answer down. An inspector absorbs that ambiguity with judgment; a camera converts it into rejects at whatever rate your guessed threshold produces. Writing the spec is the engineering work that unblocks automation, and it costs a fraction of the hardware.

The third is trust. A station no one on site can retune becomes a bottleneck at the first part change, and a plant burned once stays manual for years. All three blockers are solvable; the rest of this guide prices the solving.

Can you write the pass/fail rule as a number?

A camera can hold the rule only when pass and fail reduce to numbers — a scratch longer than 0.25 mm, a code that decodes, a part present within ±0.5 mm. If the honest spec is "good is what the inspector says it is", the line stays manual until someone measures it.

Run the test defect by defect. Scratch length over 0.25 mm: measurable. Missing screw: measurable. Label within ±0.5 mm of nominal: measurable. Wrinkled foil, subtle discoloration, "looks burnt": not yet — and appearance defects are lighting-angle problems before they are algorithm problems, as the angle and wavelength choices by defect type show. The right light turns several of them measurable.

Where the rule is measurable, size the pixels before you pick the camera. Three pixels across the smallest defect is enough to flag it; eight to ten pixels let you measure its size. Those sampling rules come from the crack-detection camera sizing guide, and they assume usable contrast — which is a lighting decision first.

Worked example: the defect is a 0.25 mm scratch and the station sees a 60 mm wide field. A sensor 2560 pixels across divides the field into 60 / 2560 = 0.023 mm per pixel. The scratch spans 0.25 / 0.023 ≈ 11 pixels, clearing the eight-pixel measuring bar with margin. Widen the field to 200 mm and the same sensor gives 200 / 2560 = 0.078 mm per pixel, so the scratch shrinks to about three pixels — flagged, not measured. The fix is a narrower field or more pixels, not a smarter algorithm.

Pick deterministic tools for codes, presence, position and measurement, and learned models only for appearance; the trade-offs sit in the rule-based versus AI inspection comparison.

What does a manual inspector actually cost per year?

One dedicated inspector on one shift is 2,080 hours of loaded labor a year; two-shift coverage doubles that to about 4,160 hours.

Loaded hours mean wages plus taxes, benefits and overhead, priced at the loaded hourly rate your finance team already quotes. The arithmetic is short: 2,080 hours is one person, five days a week, one shift. Weekend or third-shift coverage multiplies from there.

Two corrections keep the comparison honest. Sampling: a 1-in-20 check absorbed by a shared operator is not a full salary. At 10-20 percent of one shift it is roughly 200-400 hours a year, and that smaller number is what automation must beat. Escapes: if bad parts leave the building, add parts per year × escape rate × cost per event to the manual side. Vision does not have to be perfect to win; it has to beat the loaded labor plus the escape bill.

How fast does an automated station pay back?

Payback in months is twelve times the installed cost divided by the yearly saving: at full-shift coverage a typical single-camera station clears in about three months, while a 1-in-20 sampling job stretches past two years, which is where manual keeps the line.

To keep the arithmetic clean, express every cost in inspector-hours: divide a money amount by your loaded hourly rate. At a 40 loaded rate, a single-camera station (camera, lens, light, controller, integration hours) at an installed cost of 20,000 currency units is 20,000 / 40 = 500 inspector-hours, and yearly support and retuning at 3,000 currency units is 75 hours.

Example inputs, swap in your own. Line A: one inspector dedicated to visual checks, one shift — 2,080 hours a year. Yearly saving = 2,080 − 75 = 2,005 hours. Payback = 12 × 500 / 2,005 ≈ 3 months. At full coverage the decision makes itself.

Line B: the same line samples 1 in 20 parts with a shared operator, 15 percent of a shift — 0.15 × 2,080 = 312 hours a year. Payback = 12 × 500 / (312 − 75) ≈ 25 months, past the line most capital committees accept. Escape hours flip it: at 1 escape in 5,000 parts on 250,000 parts a year there are 50 escape events, and at 500 currency units per event — 12.5 inspector-hours each — that is 50 × 12.5 = 625 hours a year. Payback drops to 12 × 500 / (312 + 625 − 75) ≈ 7 months.

The derived rule: automate when the yearly inspection hours plus escape hours reach the installed price in hours — payback inside a year. Between half and one times that, payback lands inside two years and the criteria work decides it. Below half, keep the inspector unless the escape bill closes the gap.

Which inspections should stay manual whatever the volume?

Four cases stay manual at any volume: uncontrolled environments, defects that live in feel or sound rather than in the image, part families that change faster than monthly, and sampled inspection of a stable process below about 10 percent of a shift.

• Field and uncontrolled conditions. Machine vision is a controlled-lighting discipline: fixed stand, fixed exposure, repeatable presentation. In the dark, wet or coated in oil mist, none of that holds — the ambient light control guide shows what merely changeable light already costs to tame indoors.
• Defects the body finds, not the eye. Vibration, bearing noise, torque feel, heat — give a person a thermal camera for that last one. These belong to other sensors or to people, not to a vision station.
• Fast-changing part families. Every changeover costs re-teaching and re-verification. Put that time in the payback math; a station needing half a day of retuning per change cannot win on a line that changes weekly.
• Sampled, stable, low-consequence processes. Where the process is capable and the sample is 1-5 percent, the inspector is the least expensive competent answer — a legitimate permanent choice, not a failure to modernize.

There is honest scope beyond the image, too: a good inspector hears the loose guard and feels the rough edge. It is why a buyer working through a used CNC machine inspection checklist still walks the machine in person.

How do you switch a line from eyes to cameras without losing the floor?

Launch beside your inspectors: run the system in review mode for at least two weeks, go live only after it disagrees on fewer than 1 in 100 good parts for two consecutive weeks, and keep a human audit sample in place permanently.

Phase one is the criteria sheet and a scope pack — the pre-integrator scoping checklist lists what to have ready first. Phase two is shadow mode: the station images every part and logs would-be rejects while the inspectors keep making the call. Their disagreements are your tuning dataset and their buy-in is what makes go-live quiet. Phase three is the gate from the first sentence, then the system takes the decision and the inspectors move to audit duty.

The target is asymmetric on purpose: pass zero bad parts while passing as many good parts as the optics allow. Keep a person on the reject photographs at first, because a rising reject rate usually means the part or the lighting drifted, not that quality collapsed. For decode stations, have the floor read the confidence score with the result; judging whether a decode is trustworthy is the pattern to copy.

Two guards: buy equipment your own engineers can retune, because a box nobody may configure turns one drifted threshold into a line stop; and when a station is cut into a running line, design its guarding and stops with the machine, not after it. AI surface inspection can phase in the same way on cameras you already own — adding AI inspection to existing cameras walks that path.

Does your station pass the stay-manual checklist?

Score each line honestly: two or more failures across criteria, volume or environment mean keep the inspector this year, and a clean sheet means start with the easiest project below.

Criteria

• Pass and fail written in millimeters, counts and positions — no adjectives
• Tolerance agreed by quality and production, in writing
• Contrast demonstrated on 50 known-good and 50 known-bad parts

Volume and money

• Inspection occupies most of a shift, or escape costs are real money
• Payback from the formula above lands inside 24 months with your inputs
• Re-teach time per part change fits the production calendar

Environment and people

• Fixed stand, controllable light, parts presented the same way each cycle
• Optics stay clean a full shift — no mist, spray or dust on the window
• Someone on site can retune thresholds without a vendor visit
• Inspectors are planned into the rollout, not designed out of it

Start where payback is fastest: identity. A code read is binary, the criteria are already written, and the hardware is a reader plus a light. The LS-CRD-0028 pairs a 2560 × 2048 global-shutter sensor with an 8 mm lens and reads up to 60 codes/s at 30 fps over Fast Ethernet. For close-range code work, the compact MV-ID3013PM-12M-WBN pairs a 12 mm focal length with 1.3 MP resolution.

Presence and assembly checks are the second win: part present, cap seated, label in position. The MV-SCB007EC-12S-WBN covers that class with a 1.6 MP color sensor, a 12 mm lens with manual focus, and a global shutter over Fast Ethernet.

When appearance joins the checklist, step up the resolution so the pixel scale stays fine over wider fields. The MV-SC3050XC-16M-ATN-H runs a 5 MP color sensor with a global shutter and an M12 mount.

When the checklist passes and the payback holds, spec the hardware in one pass: the camera finder filters area scan cameras and the other camera classes by resolution and interface, the vision system configurator walks the camera-lens-light pairing, and smart cameras plus fixed code readers cover the first two project classes outright. Send the criteria sheet and part photographs with a quote request and we will check the pixel math before you spend anything. More applied decision guides sit on the applications topic page.