Part localization across surface finishes with lighting you control

A vision-guided robot picks only as well as it localizes, and mixed surface finishes are where localization breaks. The same part arrives machined and polished on one batch and shot-peened or moulded matte on the next, and the model that found it all morning reports no match. The cause is physical: a polished finish redirects your lighting away from the lens while a matte one scatters it, so the grey-level pattern the model was taught on no longer exists. This guide gives the geometry and exposure rules that let one taught model hold across shiny, matte and mixed finishes, with a worked camera-and-dome example you can rescale to your own field of view and tolerance.

Why the same part images differently in each finish

A polished surface behaves like a mirror. The law of reflection sends each incoming ray away at the same angle on the far side of the surface normal, so with a directed light most of the energy never reaches the lens. The camera records a white flash where a facet happens to aim the light back, and black wherever it aims the light away. A bead-blasted, moulded or painted finish spreads the same light into a wide lobe, so every viewing angle receives a moderate amount. The bright field versus dark field decision is built on exactly this behaviour.

Between a shiny and a matte batch, the same edge can flip from a bright rim against black to mid-grey against mid-grey, or vanish entirely. Matching algorithms compare grey-level patterns, and when the pattern changes structurally no threshold tweak recovers it. The fix is to make the pattern finish-independent at the sensor, and illumination geometry does most of that work. The broader camera, light and part positioning guide covers the geometry triangle in general; here it is aimed at the finish problem.

What diffuse geometry removes

A dome light illuminates the part from every direction of a hemisphere at once, and shadowless lights combine several diffuse emitters for a similar effect from above. For every facet tilted away from the lens there is another ray direction whose specular return lands on the sensor. Flashes fill in, black facets brighten, and the difference between finishes collapses to a gentle grey-level offset instead of a structural change. Contrast then comes from shape: edges, holes, fillets and curvature boundaries, which is what localization matching actually reads. The full treatment of domes on curved and wrinkled parts is in the dome lights guide.

The same multi-directional fill removes hard shadows, which otherwise read as false edges; the shadow removal guide shows how shadows fool checks. One artefact still needs planning: the camera hole at the dome apex casts a small dark patch onto flat, mirror-like tops, and on a polished part it can land on a taught feature. The dome camera hole fix walks through the placement and model margins that solve it.

Fix the exposure before you touch the model

Auto-exposure quietly undoes diffuse geometry. When a shiny batch returns less light, the camera lengthens the exposure, the ambient background rises with it, and the whole image shifts between batches. Turn auto-exposure off, set a fixed exposure, and strobe the LED inside that exposure so the lamp rather than the room defines the image. The strobed lighting guide measures and improves this ambient rejection step by step.

Quantify the leak with two frames at the working exposure: one with the light off, one with it on.

Hold the fraction below 0.10. A doubling of ambient light shifts the lit image by the fraction itself, so 0.10 caps the move at roughly 10% of grey level, which a normally taught model rides through without a score collapse; at 0.30 the same doubling jumps the image 30% and edges start to drift. Where the fraction will not come down, shorten the exposure, overdrive the strobe, or fit a bandpass filter matched to the LED colour, as in the red LED plus bandpass filter guide. The pick-and-place ambient control guide lists the usual ambient sources in robot cells.

Wavelength, filter and polarizer

Red near 620-630 nm is the default localization colour: monochrome sensors respond strongly, and a matching bandpass filter on the lens rejects most of the remaining broadband ambient. When the finishes also differ in colour (black oxide against zinc plate, printed film against bare polymer), pick the LED colour that keeps the shared edges visible on both, using the part and lighting colour logic and the LED colour comparison.

A polarizer is the last tool for flat, mirror-like faces. A polarizing film on the light and a crossed analyzer on the lens block the directly reflected specular component while passing part of the diffusely scattered one. Budget for the loss: each film passes roughly 40% of unpolarized light, so the pair leaves about a sixth of the signal, and exposure or LED drive must rise to match. For the filtered route, the LS-SDL-0053 flat shadowless light is a 150 x 150 mm diffuse panel that emits at 620 nm.

Choosing geometry by finish mix

The geometry decision reduces to your finish mix and your headroom:

For the bar-light row, the ring light versus bar light comparison and the light geometry selection guide cover the trade-offs, and the bar lights category lists the diffused types.

Worked example: one dome, one camera, three finishes

An actuator end cap carries a polished flange, a matte painted body and, on one variant, a shot-peened flange. The robot picks the cap from a nest and inserts it into a housing, and the pose must be known to ±0.1 mm. The cell sits under skylights, so ambient changes through the day. The requirement is a single taught model for all three finishes.

Size the field of view first: an 80 mm part with 15 mm of locating margin per side needs 110 mm. On a 5 MP monochrome camera with 2448 × 2048 pixels, the pixel size on the part is:

A subpixel edge or centroid fit lands within 0.1-0.25 px on a well-lit edge. At 0.2 px the repeatability is 0.2 × 0.045 mm = 0.009 mm, a margin of about 11:1 against the ±0.1 mm requirement.

The dome must bathe the whole field, so an emitting area of at least 110 mm sets the floor. The LS-DOM-0085 dome light has a Ø166 mm outer diameter and a Ø116 mm emitting area. That clears the 110 mm field and leaves the flanks to the steeper rays that keep a curved part even.

Set the exposure at 2 ms with a 1.5 ms strobe pulse inside each frame, then measure. The light-off frame reads 8 counts on an 8-bit image and the lit frame reads 118, so the ambient fraction is 8 / 118 = 0.068, inside the 0.10 rule. When the work lights come on and double the ambient, the lit image moves from 118 to 126 counts, a 6.8% brightness shift of eight grey levels, and thresholds and scores hold.

Teach the model on the finish that scores lowest, here the shot-peened flange, then verify on polished and painted. Accept the set-up when the found pose moves less than 0.05 mm between finishes (half the pick tolerance) and the score drops less than 10 points on a 0-100 scale. In this example the polished flange scored 4 points below the teach finish with a 0.02 mm offset, so one model ships. If the offset ever grows past half the tolerance, the lighting moved rather than the part; re-run the ambient test before retraining, as the retraining trigger guide explains.

If the part moves during capture instead of resting in a nest, cap the exposure at pixel size divided by speed. At 150 mm/s with 0.045 mm pixels, 0.045 / 150 gives 0.0003 s, so 300 µs of exposure with the strobe pulse shorter still.

Sizing and mounting notes

Two sizing rules cover most cells. First, the emitting area spans at least the field of view, so the corners are lit from all sides rather than one. Second, for polished curved parts start near twice the part's widest dimension and shrink only if a rotation test shows even fill, because the steep-angle rays from the dome wall are what keep the flanks lit.

Where the fixture leaves little headroom, flat domes and square shadowless panels sit close over the part and give the same multi-directional fill in a low profile. The LS-FDL-0002-L8 flat dome light provides a 251.5 x 251.5 mm lit area, and the LS-SDL-0006 square shadowless light measures 120 x 120 x 30 mm.

Keep the part inside the emitting footprint. A dome mounted so one flank sits outside the emitting ring lights that flank hard and the other softly, which reads exactly like a finish difference. Capture, rotate the part 30°, capture again: near-identical images confirm the fill.

When diffuse localization is the wrong tool

Diffuse geometry solves finish-dependent contrast, not every localization problem. Features inside bores, threads or deep pockets sit beyond a dome's reach, and a 360 degree inspection lens images those inner walls directly. Engraved marks and low-relief stamps need grazing dark field to pop, which a dome flattens; the engraved text lighting comparison sets dome against grazing light. Transparent and translucent parts follow their own rules, covered in the small transparent parts guide. And when the inspection itself reads the finish (scratches, dents, finish grades), localize with the diffuse light and inspect on a second, directed exposure; the illumination types guide shows how to pair geometries in one station.

Checks before you lock the design

1. Auto-exposure off, exposure fixed, strobe enabled with the pulse inside the exposure window.
2. Ambient fraction measured at the working exposure and below 0.10 at the worst hour of the day.
3. Model taught on the lowest-scoring finish, then verified on every finish with offset under half the tolerance and score drop under 10 points.
4. Part rotated in 30° steps through all fixture orientations: no flank black, no offset growth.
5. Cell extremes repeated: work lights on and off, enclosure lid open and closed, brightest time of day.
6. Lighting controller settings (channel, intensity, pulse width) recorded with the vision job, so a replaced lamp is restored identically; the lighting for AI vision training guide explains why consistent capture protects any taught model.

The decision for each case

Polished or curved part, one model required: a dome with emitting area at or above the field of view, a fixed strobed exposure and an ambient fraction under 0.10. Mixed finish with tight headroom: a flat dome or square shadowless panel wider than the field of view, plus a bandpass filter on the lens. Matte part with weak edges: a low-angle shadowless light or opposed diffused bar lights for gentle directionality. Finishes that genuinely change the feature set: one taught reference per finish, selected by a coarse reflectance check, because geometry cannot restore a missing feature.

Pick the exact lamp on the lighting finder with your size and wavelength filters, and the lighting topic page collects the neighbouring guides. For an unusual finish mix (oily films, wrapped parts, black and polished on one body), the contact page is the fastest way to size the light with us.