Machine vision lighting test: from bench to production line

A lighting setup earns its place on the line when the feature you inspect stays separable through every part and condition production will deliver — not when one polished part produced one beautiful image on the desk. This checklist is for engineers carrying a machine vision lighting test from the bench to the production floor. It gives you a measurable pass criterion, the sample set to gather, a bench and a line test sequence, one worked exposure calculation, and the go or no-go marks to apply before you order hardware.

Set the pass criterion before the first image

A criterion is a number read from your production camera, not an adjective. Decide what "visible" means before you power a light, or every later choice drifts toward whatever image looked impressive that afternoon.

• Name the feature and its direction. "A 0.3 mm scratch reads at least 40 gray levels darker than the surrounding surface" is a criterion; "scratches are clear" is a mood.
• Lock the camera first. Fix exposure, gain and aperture and switch every auto mode off; auto exposure quietly re-scales the numbers you are judging. The failure mode is described in fixing brightness drift by turning every auto mode off.
• Measure the gap and the noise. The gap is the mean background gray minus the mean feature gray, in levels on an 8-bit image; the noise σ is the gray-level standard deviation over a uniform patch of background in the same image.
• Use a 6σ margin as the working pass mark. That is the margin logic used for tolerance stacks: it leaves headroom for part-to-part variation, batch drift and LED ageing on top of the noise you measured today.
• Log the number next to every test image. An unlogged result cannot be defended at handover.

Gather the parts the line will deliver

The sample set decides whether a bench result means anything. Label three groups and keep them for the whole project:

• Ten known-good parts, from at least three production batches where surface finish varies; finish drift moves contrast more than geometry does.
• Ten borderline parts: fresh parts just inside the acceptance limit and parts just outside it. This group sets the real threshold — without it you only know the setup works on parts nobody argues about.
• Ten rejects per defect type the system must catch, including the smallest defect that must be detected. If the spec says 0.5 mm and the light only shows 2 mm scratches, the setup has failed its own test.
• Parts mounted at the worst tilt and position your fixture allows, because the bench sees parts flat and centred and the line does not.

Every lighting change re-runs the same tray, so results stay comparable across the project.

Bench stage: change one thing at a time

Bench work finds geometry and wavelength. Keep camera, lens, aperture and working distance fixed and move only the light, so you know which change produced the result. Lighting geometry for image contrast explains why the relative positions of camera, light and part set the contrast you are measuring.

• Walk the geometry families in order: direct ring, angled ring, full diffuse, low-angle dark field, then backlight if the feature reads as a silhouette. The geometry selection by surface and feature and surface or silhouette guides tell you what each family does to which surface.
• Trial two or three wavelengths on the same parts. Red, blue or infrared changes material contrast with no mechanical change; light colour and monochrome contrast shows how colour decides what brightens and what goes dark.
• Record working distance, incidence angle, drive current and the measured gap for every configuration in one table. You order from that table, not from memory.

A small diffused ring is the usual starting geometry for flat parts under a lens:

Fixed angles repeat where hand-tilting drifts, so build the angle sweep from fixed rings:

Long parts — extrusions, strips, seals — want light along their length rather than around a point, and a bar light covers that case. The LS-BAR-1047 is a 290 x 30 mm, 24 V bar light sized for a bench rig.

Worked example: exposure budget for a line moving at 240 mm/s

The bench sees a static part; the line moves it. This calculation sets the exposure the line allows, and it is where most passing bench setups die.

Inputs, for an example rig (substitute your own values from the conveyor and camera specifications):

• Field of view 120 mm across — a 110 mm part plus 10 mm of fixture margin.
• Camera 2448 pixels across that field of view — a 5 MP sensor.
• Line speed 240 mm/s, from the conveyor specification.
• Blur tolerance 1 pixel — the usual limit for edge and measurement tasks; at 2 pixels the cap doubles.

Rule in words: the part must not cross more than one pixel during the exposure, so the cap is t_max = p / v, where p is the pixel pitch on the part and v is the line speed.

Result: the exposure cap at this speed is 0.2 ms. Decision rule: if the continuous-light exposure that cleared 6σ on the bench is 20 ms — a common continuous value — the line allows 100 times less light (20 ms ÷ 0.2 ms = 100). Continuous illumination will blur, and the light must be strobed with a pulse of 200 µs or less fired inside the exposure window; the missing photons come from pulsed overdrive, which short pulses permit. If the bench exposure already fits inside 0.2 ms, continuous light survives this speed.

The cap scales with pixel size: at a 60 mm field of view on the same camera it halves to about 0.1 ms, and at a 240 mm field of view it doubles to about 0.4 ms. Compute your own cap from your field of view and camera:

Line stage: stress the setup until something fails

Mount the light as it will be mounted and power it through the controller it will use, then run these cells and measure the gap in each. Looks do not count on the line.

• Ambient on and off. Plant lights on, then off; add sunlight through a nearby door or roof light if the station has one. If the gap clears 6σ in the dark and collapses under ceiling lights, the fix is strobing and a shroud, not a software threshold — measure the margin as in strobed lighting ambient rejection, and plan around it with controlling ambient variation in pick and place.
• Tilt sweep. Step the part through the fixture's real range in 5° increments and record gap against angle. The first angle where the gap falls below 6σ is your fixture tolerance, in writing.
• Position sweep. Part at the centre, then at each corner of the fixture envelope; light falloff and vignetting hide from a centred part.
• Full-speed run. Repeat at maximum and minimum conveyor speed and check the blur against your own cap from the worked example.
• Flicker check. With exposure locked, look for banding or frame-to-frame brightness jumps; mains flicker appears as brightness changing frame to frame.
• Cold start and warm line. Image in the first minute after power-up and again after half an hour; LED output shifts until the light reaches thermal equilibrium.

Timing is the last cell: the camera, the light pulse and the part must land inside the same exposure window, and stabilizing lighting changes with trigger timing walks that chain end to end.

Make the go or no-go call

One fixed threshold is part of the decision: if each condition needs its own threshold, the lighting is too unstable to ship, and the fix belongs on the light, not in the software. All cells pass: order the hardware, lock the recipe, keep the sample tray. Any cell fails: take that cell's condition back to the bench as the new design case and work the sequence again. The lens and lighting selection test plan pairs this checklist with the camera-side choices that come before it.

Why a passing bench test still fails on the line

The failure modes are few and predictable, and every one has a cell above — the static-versus-moving gap from the exposure budget is only the first. The bench runs continuous light at full duty; the line fires a 200 µs pulse that must carry the photons. The bench sits in a dark corner; the line sits under roof lights beside an open dock door. The bench part is placed by hand; the line part arrives at the fixture tolerance, tilted. None of this is a reason to distrust bench work — it is the reason the line stage exists.

Hold the result once it passes

• Size the controller to the light and lock it. Match channels and output power to the lights you passed the test with, prefer constant-current output where brightness must hold across temperature and LED ageing, and keep drive current fixed in software rather than adjusted by eye. Short-pulse overdrive is a design decision, not a knob — see LED overdrive and current control. A strobe-capable digital controller fires the light only inside the exposure window:

Its 4 channels and 200 W cover a two-light station with headroom.

• Bolt the light down. Use a bracket that holds working distance and angle, marked so a bumped light can be reset exactly; hand-clamped lights do not survive maintenance week.
• Keep a golden image. Reference part, locked settings, recorded gap; re-image weekly and investigate any fall — LED ageing, a dirty diffuser, a nudged bracket — before it reaches production parts. Where the process will not hold still, closed-loop lighting control keeps the image steady on its own.
• Shield whatever the ambient cell flagged. A shroud, matte dark surfaces around the station, or a strobe short enough to swamp daylight.

Everything in the test lives in the shop: the lights finder narrows models by wavelength, size and angle, the light controllers finder filters for strobe and constant-current output, and the light configurator and controller configurator size a setup to your field of view, working distance and load. The lighting topic page collects the geometry and contrast guides behind this checklist. When a cell keeps failing, bring the images and the measured gaps to the contact page — the numbers make the diagnosis quick.