What makes a machine vision inspection point worth keeping

Inspection programs grow by addition and almost never by subtraction. Every point added brings a camera, a light, a controller channel, storage and review time, long before it brings a caught defect. This guide is for the process, quality or controls engineer trimming an inspection program or vetting a new station: four tests, the arithmetic behind each, and a verdict of keep, fix, convert or cut for every point on the line. Run the scoring once and you will know where the next camera budget should go and which points to retire first.

The four tests an inspection point must pass

A point earns its place only when it passes all four:

1. Repeatable capture. The same part imaged the same way every pass: fixed geometry, locked exposure and gain, controlled light.
2. A measurement. The point outputs a number — an edge position, a width, a gray level, a gauge reading — not just an image.
3. A baseline. That number is compared against a tolerance or a reference part on every pass.
4. An action. An out-of-range result triggers a named response: reject, divert, alarm, or a trend review at a set threshold.

Fail test 1 and the number wanders. Fail tests 2 to 4 and the point generates data with no decision attached. Presence-only logging is the classic miss: an image and a timestamp of the same part from the same angle, identical on pass three and pass three thousand. The quick screen takes minutes: run three identical passes under fixed lighting and angle, and if none of the three changes a decision, cut the point.

What a kept point buys you is the other half of the question: which stations on a line actually stop escapes. We cover that pay-off side in prevent product recalls with machine vision inspection points; this guide is the cut side.

Test 1 in practice: fix geometry, kill the auto modes

Repeatability starts with mechanics. Bolt the camera and light to the frame, present the part in a fixture or at a positive stop, and hold the working distance constant. Scale moves with distance: at 100 mm working distance, a part sitting 5 mm farther away is imaged about 5% smaller (5 / 100). On a 20 mm feature that is a 1 mm apparent size shift, twenty times a ±0.05 mm tolerance. No algorithm recovers a tolerance the geometry forgot.

Then remove every automatic decision. Auto exposure, auto gain and auto white balance re-decide the image on every frame, and a baseline comparison needs the same decision forever. Lock them to fixed values and re-check on a schedule; the failure mode and the fix are covered in camera image brightness drifts over time: turn every auto mode off.

Light does the rest. Drive the LEDs from a light controller in constant-current mode, shade or baffle ambient sources, and prove stability across warm-up and a full shift before sign-off; the procedure is in machine vision lighting stability: test before deployment. Where drift still creeps in, closed loop lighting control for repeatable vision images trims drive current against a reference target in the fixture.

Set an acceptance bar before arguing about algorithms. Same part, three passes: mean gray level within about ±5 levels on 8-bit data, and the reported measurement repeating within 10% of the tolerance band — the band gauge R&R studies conventionally accept for a working gauge. Between 10% and 30% needs a stated reason; above 30% the point cannot yet tell good from bad.

Test 2: can the camera see the defect at all

Resolution is a screen, not a taste. Detection needs at least 3 pixels across the smallest defect you must catch; reporting its size needs 5 or more. Both come from one division:

Worked example: a bracket face 120 mm wide must be checked for cracks of 0.25 mm and longer, with crack length reported.

• Field of view 120 mm on a 4000-pixel-wide sensor: pixel size = 120 / 4000 = 0.03 mm/px.
• Pixels across the smallest crack: 0.25 / 0.03 ≈ 8. Detection passes (8 ≥ 3) and length is reportable (8 ≥ 5).
• The same sensor stretched to a single 500 mm view: 500 / 4000 = 0.125 mm/px, so the crack spans 0.25 / 0.125 = 2 pixels. No threshold finds a 2-pixel crack reliably.
• Decision: split the point into two views of 250 mm, where the crack spans 0.25 / (250 / 4000) = 4 pixels — enough to detect, marginal to size. If length must be gauged, run four views of 125 mm and it spans 8 again.

A 12 MP USB3 camera with a 4000 × 3000 sensor carries the two-view layout:

The 30 fps rate is ample for a stopped part, and a rolling shutter costs nothing when nothing moves between the rows. Moving parts are test 3, and they want a global shutter at the same pixel count — a 12 MP global-shutter camera over USB3 fits:

Test 3: the blur budget

Motion destroys measurement before lighting enters the discussion. The budget is one multiplication:

Worked example: a belt at 300 mm/s with the camera free-running at a 2 ms exposure blurs 300 × 0.002 = 0.6 mm — 20 pixels at the 0.03 mm/px from test 2, enough to swallow the 0.25 mm crack entirely. Strobe the same pass at 100 µs and the blur drops to 300 × 0.0001 = 0.03 mm: one pixel, and the crack keeps its 8 pixels of extent. Set the camera exposure to the strobe width or shorter, and trigger both from the part-present sensor.

If the point cannot be strobed and no exposure the light tolerates holds blur under one pixel, it is a presence camera whatever the proposal says. Move it to a stop or cut it.

Strobing is a light function, not a camera feature. When the camera offers no flash output, hardware triggering takes over — the wiring and the fallbacks are in camera LED flash control with no API: add an external strobe, and the drive side is set up at the lighting step of the system configurator. One strobe channel per light, fired from the exposure or part-present signal, makes the timing deterministic:

The LS-CTL-0070 switches four channels at 200 W with strobe on each, on a 24 V bus, and takes RS-232 for recipes — four lights from one box.

Test 4: a number, a limit, an owner

A kept point writes a record on every pass: the measurement with its unit, the pass or fail verdict, the trigger source, the recipe version, and an image reference for exceptions. Samples and rejects keep images; routine good parts keep the numbers. That record is what turns a dispute into a lookup.

Give the trend a rule with an owner. Set the review alarm at half the tolerance band: a 20.00 ± 0.05 mm feature raises a review when the 100-part running mean moves 0.025 mm off its baseline. Half the band stays in reserve before scrap, and the drift is caught while tooling can still be adjusted. Recalibrate the baseline on a schedule, not when someone remembers.

Where those thresholds live — editable recipe or frozen release — is a rollout decision of its own, covered in vision inspection deployment workflow: keep it editable or freeze it. Defect classes that resist rules, such as stains and texture, face the same four tests; choosing between rule-based and learned logic is covered in AI vs rule-based smart camera inspection: how to choose.

The first points to cut

Cut presence-only points first. Image plus timestamp, same part, same angle, same light, no measurement, no threshold: pass three and pass three thousand produce the same non-decision, while the volume grows and the information does not.

Most presence points hide a measurement nobody asked for: seal edge position, cap height, label skew, connector seating depth. Add one number and a threshold and the point passes all four tests. A backlit silhouette with an edge read is often enough; the method is laid out in machine vision edge inspection: backlight and histogram methods. Converted points also produce the clean, repeatable captures that learned checks train on — see adding AI surface defect inspection to existing cameras.

True presence checks that survive — cap on, label present, connector seated — belong on a smart camera or vision sensor that decides at the line and ships a result, not a stream. Five-megapixel global-shutter vision sensors on Fast Ethernet do exactly that:

The MV-SC3050XC-16M-ATN-H is a 5 MP color vision sensor with a global shutter on Fast Ethernet and an M12 mount: inspection and verdict happen at the point, and the network carries results instead of images.

Keep, fix, convert or cut: the verdict table

Score every point on the line against the four tests; the row it lands in is the decision.

Most lines find two or three points sitting in the convert and cut rows — usually the camera budget for the next real station, already spent.

Checks before you approve the next point

• Three passes of a golden part under locked settings: gray-level means within ±5 levels (8-bit), measurement repeating within 10% of tolerance.
• Pixel count: smallest defect ÷ (field of view ÷ sensor pixels) at 3 or more to detect, 5 or more to report size.
• Blur budget: part speed × exposure at 1 pixel or less, with the strobe width chosen to satisfy it.
• Threshold tied to the drawing tolerance, trend alarm at half the band, a named action for every result.
• Storage at the captured rate: exceptions plus sampled passes, never full-rate images of good parts.
• Warm-up: light and camera running for the first half hour of the shift before the golden-part passes. The drift you catch there is the drift you would otherwise ship.

When a quiet point still earns its place

A stable process alarms rarely; that is control working, not redundancy. Keep a quiet point when three things hold: it outputs a quantitative measurement, its baseline is fixed and re-verifiable, and its trend rule has an escalation threshold with a named owner. A gauge that has not alarmed in a year on a stable part is a success story, and it stays.

Quiet of the other kind — images nobody opens — gets demoted before it gets deleted. Sample one pass in 50: the trend survives on 2% of the frames (1 / 50), 98% of the storage goes, and a drift of 0.02 mm per week still reaches the running mean within a few hundred sampled parts. If even the sampled trend never moves, cut the point and write down why, so the next proposal does not resurrect it.

Run the four tests across your points this week; most lines recover at least one camera's worth of budget from the convert and cut rows alone. For a new station, size the sensor with the resolution calculator above, pick hardware through the camera finder or build the point in the configurator, and browse more applied walkthroughs on the applications topic page. If a point scores marginal on two rows, talk it through with us — that judgment call is exactly what the support line is for.