There's a number that keeps appearing in the human factors literature on visual inspection: roughly 80% detection efficiency at moderate throughput. Push the line above 100 parts per minute, and that 80% figure starts to degrade. By the time you're at 150 ppm, you're often looking at 60% or lower — meaning four out of ten defects are walking off the line. This post explains the physiology behind that degradation and what it means when you're calculating escape rate on a real production floor.
The fixation budget problem
Human visual inspection is not a passive scan. It requires the inspector's fovea — the central 2–3 degrees of high-acuity vision — to land on the defect location for the eye to register it. A typical fixation lasts between 200 and 350 milliseconds. At 100 ppm, a part is in the inspection window for roughly 600 milliseconds if the station is reasonably sized. That gives an inspector two or maybe three fixations per part.
Surface defects like edge cracks or scoring marks can appear anywhere across the face of the part. If the defect is in the peripheral visual field, the brain has to decide whether to trigger a saccade — a rapid eye movement — to bring it into foveal view. This decision-and-move cycle takes another 150–200ms on its own. Do the math: at 120 ppm, a typical stamped bracket is in view for under 500ms. Two fixations, half a second, on a 60mm part with a 0.3mm edge crack that could be anywhere along the perimeter.
It's not that inspectors aren't trying. It's that the task demands exceed what the visual system can deliver reliably above a certain throughput threshold.
Vigilance decrement over a shift
The 80% figure is generous in another way: it typically represents fresh-start performance, not shift-average performance. The vigilance decrement — the well-documented drop in detection capability over sustained monotonous tasks — typically sets in within 20–30 minutes and continues degrading across the shift.
In practice this means an inspector who is catching 82% of defects in the first hour might be down to 70% or below by the end of a 4-hour stretch without a substantive break. The math compounds quickly on a high-volume line. If you're running 120 ppm on a two-shift operation, and your defect rate is 0.5% (5 defective parts per 1,000), an inspector covering a 4-hour half-shift is making a go/no-go decision on approximately 28,800 parts. At 70% detection by the end of that stretch, roughly 43 defective parts are escaping per half-shift from that one station.
Manufacturers often assume that end-of-line inspection or downstream functional test will catch what slips through. In many discrete manufacturing contexts — stamped metal, die-cast housings, injection-molded enclosures — this assumption is only partially true. Cosmetic defects and sub-threshold dimensional errors frequently don't trigger functional test failures. They escape to the customer.
Why 100 ppm is a rough inflection point
The 100 ppm threshold isn't a hard biological law — it varies by defect type, part geometry, and lighting conditions. A large-format part with obvious black-surface scoring under bright ring illumination is easier to inspect at 120 ppm than a small stamped bracket with micro-cracking at the die break edge. But 100 ppm represents a consistent engineering heuristic derived from time-motion studies at stamping and fabrication lines going back decades.
Below 80 ppm, experienced inspectors in controlled studies typically exceed 90% detection rate for salient defect types. Between 80 and 100 ppm, rates vary more widely with inspector experience. Above 100 ppm, the decline becomes consistent enough that it should be treated as a system-design constraint, not a training problem. You cannot train a person to fixate faster than their visual system allows.
This is not a criticism of human inspectors — it's a description of a physical constraint. The same way you wouldn't ask an operator to manually weigh 150 parts per minute on a bench scale and expect reliable results, you shouldn't build a quality plan that depends on visual detection accuracy above the rate where the visual system can operate.
What escape rate math looks like with realistic detection efficiency
Let's take a concrete scenario. A stamping line in the Midwest running automotive bracket subassemblies at 130 ppm across two 8-hour shifts. Defect rate from the press is around 0.4%. Three inspectors rotate through the visual check station — one on, two rotating.
With a conservative 75% detection efficiency (accounting for vigilance decrement across a shift at that throughput), the escape rate is approximately 0.1% of total production — 1 in 1,000 parts shipped with a surface defect. That sounds small. At 130 ppm and two shifts, it's roughly 62,400 parts per day total, meaning about 62 escaped defective parts per day reaching the next assembly stage or the customer.
If this is an automotive OEM supply chain, the cost of a single field return often exceeds the value of thousands of parts. The actual escape cost isn't in the defect rate — it's in where the defect shows up and what it causes when it does.
How this changes the inspection architecture conversation
Understanding the physiological ceiling on human inspection efficiency changes how quality engineers should frame the inline inspection problem. The goal isn't to replace inspectors entirely in all contexts — that's not what we're arguing. Manual inspection makes sense for low-volume lines, for defect types that require tactile checking, and for inspection tasks that genuinely fall within what a human can reliably perform.
The argument for machine vision above 100 ppm is simpler: at that throughput, the human visual system has crossed into a regime where detection efficiency degrades as a function of physics, not skill. Frame rate doesn't get tired. A camera pointed at the right surface with the right illumination makes the same detection decision on part 28,800 as it did on part 1. The question shifts from "can we keep detection efficiency acceptable?" to "can we configure an imaging system with enough sensitivity for the defect types that matter on this line?"
That second question has its own complexity — model training with limited defect samples, lighting setup, camera geometry, integration with the PLC reject system. But it's a solvable engineering problem, not a biological ceiling. The physiological constraints on human inspection are fixed; the engineering constraints on machine vision are not.
A note on the 80% figure
Some internal quality teams arrive at detection efficiency estimates by comparing inspector catches to ground truth counts from tagged defect samples seeded into the line. This is the right method for calibrating your actual system performance, not relying on industry averages. If you haven't run a seeded-sample study on your own inspectors at your actual line speed, your escape rate estimates are likely optimistic.
The 80% average at moderate throughput cited in human factors studies represents reasonable baseline conditions — good lighting, defined inspection criteria, trained staff. It is not a lower bound. Lines with inconsistent lighting, high part variety, or ambiguous defect definitions often see lower detection rates even at slower speeds.
When we work with quality engineering teams to size an inline inspection deployment, the first question we ask is: do you have a measured detection efficiency for your current process, or are you working from an assumption? Most teams are working from an assumption. Running a short seeded-sample audit — 200 tagged defect samples inserted into regular production over two shifts — gives you an actual number to work with, and it tends to make the business case for automation much more concrete than any benchmark study would.