Teach reflective chrome parts to vision by shape not finish

A polished part can pass every other check and still fail the teach: the template stores the reflections of your room along with the outline, and the score dies the first time the part sits 5 mm higher. This guide is for integrators teaching locates on chrome, nickel and bright-plated parts, whether the camera hangs over a tray or rides on the robot arm. You will get a teach method that survives cleaning, the backlight geometry that turns a mirror into a silhouette, the pixel, lens and strobe numbers to size the job, and a pre-order checklist that catches the fixtures which quietly break the method.

Why does a chrome part teach so badly in the first place?

A polished chrome face is a curved mirror: the camera records the room, the ring light and the ceiling, so two identical parts at different heights or angles produce different images and the template score collapses toward noise.

The physics is specular reflection. On a matte surface, light scatters and every point of the finish returns roughly the same brightness, which is what a grey-level template expects. On polished metal most photons bounce off at the mirror angle, so the pixels describe your cell instead of the part. Move the part 5 mm in the tray or tilt it 3° and the highlights travel tens of pixels while the part itself has not changed.

Interior features suffer most: a chamfer or a stamped logo on chrome produces a moving patch of brightness, not a stable edge. The outline behaves differently. Against an evenly lit background it is pure geometry, and geometry does not move when the finish moves.

That decides the approach before you touch the teach button. Do not ask a bright-field template to describe a mirror; change what the camera sees. The vision locate feasibility checklist applies the same logic to any part: separate the part from the background in the image, not in the algorithm.

Can you teach on a whitened part and locate on bare chrome?

Yes: coat the part in the white developer spray welders use for dye-penetrant crack checks, teach the outline on the matte coat, then clean it — a contour-based template still fits the bare part, because it stores geometry, not the finish.

The spray lays down a thin, removable, matte white powder film, and matte is exactly what a teach wants: the finish scatters light like plastic, so the camera sees shape. The sequence that works:

1. Degrease the part and dry it.
2. Apply a thin, even coat — readable shape, not a snow drift. Photographic dulling spray behaves the same way.
3. Teach across the poses production shows: centred, near the field corners, tilted as the tray allows.
4. Clean the film off completely.
5. Verify on at least 10 untouched parts and log the score at the centre and corners of the field.

The coat buys the teach only. A contour or edge model carries geometry, so the bare part keeps matching; a grey-level model carries the white film and fails the moment you clean. Check which model your tool stores before you trust the method, and force an edge- or shape-based one — the edge matching guide shows why edges hold when lighting and finish move. The same reasoning keeps a wrist-mounted camera locate usable when the taught appearance changes.

Which light turns a mirror into a silhouette?

Put the light behind the part and take every front light away: the chrome face reflects the dark space around the panel and renders as a clean dark outline on a bright field, so the locate runs on the boundary instead of the finish.

Three setup rules make the silhouette reliable:

• Nothing reflective sits inside the outline: the part hangs past the panel edge, stands on a matte black fixture, or is gripped in open air. A shiny tray bottom behind the part becomes part of the image.
• The lit area covers the whole field of view, not just the part.
• Every front source switches off, including LEDs built into the camera.

Pick the panel by the job the edge has to do. A standard diffuse backlight is the default for locating opaque parts. A collimated backlight sharpens the edge further, which pays off when the silhouette must also gauge or when translucent material around the part would glow; give a collimated panel roughly 50-100 mm between panel and part so the beam forms, while diffuse panels work at any practical gap. The wider choice between lighting the surface and lighting the outline is set out in machine vision lighting angles.

For the running example field in this guide, a compact red panel fits with margin:

When the outline also has to gauge, step up to a collimated panel:

How bright and which colour should the backlight be?

Size the lit area to the field of view with margin — the panel should cover the field plus 10-20% on every side — and choose the colour for contrast with whatever sits behind the part.

Lit area first: for the running example, a 100 mm field around a 40 mm knob, a 150 mm panel leaves 25 mm per side, a quarter of the field and comfortably above the minimum. Undersized panels vignette the corners, and a locate loses exactly the corners it was trained on.

Power follows the timing. A standard panel at constant on suits a static station where exposure time is free. High power earns its place in two cases: short strobe pulses that must deliver the whole light dose inside a millisecond, and longer working distances where wasted lumens cost contrast. Pulsing also lets a panel run above its constant-on rating, so the same hardware buys a shorter, brighter exposure.

Colour does three different jobs. Red is the default for mono sensors — bright on silicon and inexpensive. Blue or green separates part from surroundings by colour, useful when chrome sits over galvanised basket wire. Infrared at 850 nm or 940 nm hides the locate light from operators and rejects visible ambient, but only a mono sensor sees it.

The high-power panel earns its place in strobed stations:

The infrared panel needs no controller at all and runs constant on from 24 V:

What camera and field of view does a silhouette locate need?

Size the field at 1.5-2 times the part width — a 40 mm part gets a 60-80 mm field — and keep one pixel under a third of the locate tolerance: ±0.1 mm wants 33 µm pixels or finer, which a 6 MP sensor just meets in a 100 mm field.

Worked example — chrome knob, 40 mm across, picked from a tray, locate tolerance ±0.1 mm:

1. Field. Part 40 mm, pose spread ±15 mm, so the part can sit anywhere in a 70 mm window; add margin and set the field to 100 mm.
2. Pixels. Take a 6 MP mono sensor with a 3:2 array, 3072 × 2048 pixels. Pixel size = 100 mm ÷ 3072 = 0.0326 mm = 32.6 µm, inside the 33.3 µm the tolerance allows (0.1 mm ÷ 3).
3. Edge fit. A clean backlight edge locates to about a fifth of a pixel: 32.6 ÷ 5 = ±6.5 µm per edge, and the centre from two edges lands at ±4.6 µm (6.5 ÷ √2 ≈ 6.5 ÷ 1.41).
4. Lens. A 1/2.4" sensor is about 5.5 mm wide in a 3:2 array, so magnification m = 5.5 ÷ 100 = 0.055, and f = WD × m ÷ (1 + m). With f = 8 mm, WD = 8 × 1.055 ÷ 0.055 ≈ 153 mm. Set the working distance to 150 mm and the field lands at about 98 mm — on target.
5. Budget check. ±4.6 µm of algorithm noise under ±0.02-0.05 mm of calibration and fixturing still leaves roughly 2x margin to ±0.1 mm: good for a pick. For gauging, tighten the pixel budget to a tenth of tolerance and re-run the numbers.

Size your own field and tolerance with the resolution calculator:

Mono beats colour for this job: no Bayer filter sitting over the edge, and full response to red and infrared. For a station like the example — locate on the device, result to the robot over GigE — a compact smart camera with a fixed 8 mm M12 lens covers it:

Its 1/2.4" mono sensor is what the lens math assumed, and with the part stationary during capture the rolling shutter costs nothing; the strobe pulse simply spans the frame readout. Browse smart cameras for global-shutter or higher-resolution variants of the same idea.

How do you set exposure and strobe so the reflections stop moving?

Strobe the backlight and set the camera exposure to bracket the pulse: a 1 ms pulse freezes motion to 0.1 mm at 100 mm/s, cuts most ambient light, and stops the chrome from drinking the room between frames.

Work the pulse from the motion, not the other way round. Blur equals speed times pulse: at 100 mm/s, 1 ms smears 0.1 mm, about three pixels in the 32.6 µm example field. Cut the pulse to 0.2 ms and the smear drops to 0.02 mm — 0.02 ÷ 0.0326 ≈ 0.6 pixels. If the robot parks and the part is still during capture, the pulse only needs to span the frame readout, and blur is zero anyway.

The same short gate cleans the ambient picture. A 1 ms exposure collects one-tenth of the stray light a 10 ms exposure does, which is why strobed stations shrug off workshop fixtures; when mains-driven lights still beat against the gate, the flicker diagnosis guide sorts pulse timing from exposure faults. Wire the trigger chain in the right order with the camera strobe timing guide. Keep exposure fully manual in production: if auto exposure ever runs during commissioning, set and lock its longest permitted time so motion blur cannot sneak back in. Residual ambient — sunlight through a doorway, a work lamp — gets the treatment in the ambient light control guide.

Turn the front lights off for the silhouette shot: every front source, including the camera's own ring, is a scene the chrome will mirror.

A four-channel box from the light controllers range covers the silhouette panel now and a dome or coaxial light later, driven over serial from the camera or PLC:

When should you light the face instead of the shadow?

Silhouette locate fails when the question is about the chrome surface itself — scratches, smears, dent depth, engraving — so switch the goal and the light: a diffuse dome for cosmetic defects, coaxial for marks in recesses, dark field for cracks at edges.

A dome surrounds the part with diffuse, multi-directional light, so a curved mirror sees an even brightness everywhere instead of one lamp; that is what makes smears and soft dents readable on polished surfaces. As a rule of thumb, plan a dome opening about twice the part width so the viewing cone stays clear. The same trade-off between dome and dark field on bright metal is worked through in the dome or dark field guide.

Engraved characters and flat-bottomed pockets respond to coaxial light dropped straight down the optical axis; the recessed marking guide shows the geometry on metal. Crossed polarizers tame glare on paint and plastic almost completely, but bare metal scrambles polarization, so on chrome expect the glare to drop rather than vanish, at a cost of about one to two stops of exposure. When the question is dimensional — dent depth, warp, a missing chamfer — height data sidesteps reflectance altogether, which is the route the reflective bin picking guide takes.

Run two shots rather than compromise one: the silhouette frame for locate, a second frame under dome or coaxial for the surface check. On a smart camera that is two exposures with the controller switching channels between them.

What should you check before you order the parts?

Before ordering, confirm four things: the part can hang past the backlight with nothing reflective behind it in the field, the fixture stays matte and dark, the field covers pose spread with 20% margin, and the score survives cleaning on at least 10 untouched parts.

Geometry

• The part clears the backlight, and nothing reflective sits anywhere inside the outline in the image.
• Field of view = part + pose spread, with 20% margin on every side (100 mm in the worked example).
• Pixel size at most a third of locate tolerance: 33.3 µm for ±0.1 mm.

Lighting

• Lit area covers the field: a 150 mm panel for the 100 mm example field.
• Panel type matches the job — diffuse to locate, collimated to gauge, high power to strobe, infrared when people work beside the cell.
• Every front light, including the camera's own LEDs, switches off for the silhouette frame.

Camera and timing

• Mono sensor if red or infrared is in the plan; a colour sensor filters infrared out.
• Pulse short enough that speed × pulse stays under a pixel, or part stationary during capture.
• Exposure brackets the pulse and runs fully manual in production.

Verification

• Teach on the whitened piece, then score 10+ cleaned parts at the centre and corners of the field.
• Re-score after any tray or fixture change — the background is part of the silhouette.
• Log scores per shift for the first week; drift with part height means reflections are still reaching the template.

When part and fixture are both metal — a chrome part on a shiny jig — the fixture becomes the second reflector in the scene, and the metal-on-metal lighting guide is the fixture-first playbook. For more pairings of surface and light, the lighting topic page collects the guides, and the light finder filters backlights by size, colour and control.

If the numbers above describe your station, size the panel, pick the camera, and step through the configurator:

If the part cannot lift off a background, or the inspection lives on the chrome face itself, start from the dome, coaxial or 3D routes in the section above instead. And if the layout resists all of them, describe it through the contact page before spending on hardware.