Raised mark presence inspection: oblique or coaxial lighting

A 3 mm × 3 mm raised alphanumeric mark with 0.5 mm of relief carries a lot of responsibility on a mixed-model line: it is what separates a good part from the wrong one. Manual checks miss, so the fix is a vision sensor that judges the mark's shadow pattern and hands the line a pass/fail bit. This guide gives the numbers for that set-up: the light elevation that turns 0.5 mm of relief into contrast, the pixel budget, the field of view the part must land in, and the cases where a coaxial light beats an oblique bar light. You leave with a decision per case: fixed station, mixed-model line, or handheld audit.

What the check has to decide

Decide which of three judgements you need, because it changes the software, not the light:

• Presence — a mark is there at all: a blob clears a threshold inside the window.
• Identity — the pattern is the right one for this model: a pattern-match score clears a floor.
• Quality — the mark is complete: the score drops when strokes are damaged or partly formed.

All three feed off the same thing: sharp gray separation between the relief edges and the flat field. Under a top-front light the relief is nearly invisible; the geometry only shows when light rakes across it at a low angle. Where the camera, light and part sit decides most of the contrast before any software runs — lighting geometry for image contrast walks through that. This is not OCR: if the mark text is fixed per model, threshold plus match is faster and more stable than reading characters.

Set the light elevation from the relief height

The shadow length on the surface follows from the relief height h and the light elevation θ, measured from the part surface:

For the h = 0.5 mm relief in this case:

Rule: keep the shadow under about a third of the mark width so it reads as edge contrast without flooding the window. For a 3 mm mark that allows shadows up to 1 mm, so θ ≥ 27° clears the geometry — but strokes inside a 3 mm multi-character mark sit closer than 1 mm apart. Start at 45°, where the shadow is 0.50 mm. If shadows bridge neighbouring strokes, steepen to 60°. Drop toward 30° only when contrast is soft and the stroke pitch is over 1 mm.

Two opposed bars, one from each side, remove the dependence on part orientation and light both walls of the relief. This arrangement is dark field by reflection: the flat field stays dark and the raised walls catch the light — the law-of-reflection split between bright and dark field explains why. Recessed engraving inverts the geometry; engraved text lighting on metal covers that case.

A standard (not diffused) bar gives the hardest shadow edge. A thin bar also mounts close under a low camera: the LS-BAR-0484 is a 312 × 16 mm high-uniformity bar light.

Mount the bar with its lit length running across the mark and its face at 45° to the surface, and set the standoff so the beam covers the window plus the fixture — the worked example below puts that window at 12 mm.

Worked example: a 3 × 3 mm mark with 0.5 mm relief on a stop fixture

Inputs, with where each comes from:

• Mark 3.0 × 3.0 mm, relief h = 0.5 mm — the case.
• Light elevation θ = 45° — from the shadow rule above.
• Fixture repeatability ±3 mm — example input for a stop-and-locator on a manually loaded station.
• Edge margin 1.5 mm — example input for mark placement on the part.
• Target sampling 0.10 mm/px — chosen above the 3 px floor (0.5 mm ÷ 3 ≈ 0.17 mm/px) so the shadow spans 5 px and the mark spans 30 px.
• Sensor active width 4.9 mm — typical 4:3-mode figure for a 1/2.9"-class sensor; refine with your sensor's pixel pitch.

The rules in words: the window covers the mark plus positioning tolerance on both sides plus a margin; the pixel count is the window divided by the target pixel size; the magnification is sensor width over window; and the working distance follows from the focal length.

Arithmetic:

• Window: 3.0 + (2 × 3.0) + 1.5 = 10.5 mm, rounded to a 12 mm square.
• Pixels: 12 ÷ 0.10 = 120 px per side, and 120 × 120 = 14,400 px total.
• Magnification: 4.9 ÷ 12 = 0.41.
• Working distance at f = 16 mm: 16 × (1 + 0.41) ÷ 0.41 ≈ 55 mm.

Result: the pixel budget is 14,400 px, and a typical 1.6 MP mono sensor resolves around 1440 px across the frame — about 12× more than the window needs at 0.10 mm/px. A 16 mm lens holds the 12 mm window at roughly 55 mm working distance, which leaves room for the bar to sit at 45° beside it.

Decision, with its threshold: any mono vision sensor clears this pixel budget by an order of magnitude, so the check is light-limited, not pixel-limited. Spend the effort on the 45° bar and the fixture; if the measured gray gap between shadow and field is under about a third of full scale, change the light, not the camera. The camera resolution calculator reruns this budget for other windows and pixel targets.

Coaxial light when the face is flat and shiny

On a specular flat face, an oblique bar paints a glare stripe that moves with part tilt, and the mark competes with it. A coaxial light folds the LED output into the camera's optical axis through a beam splitter: the flat face returns the light straight back into the lens and reads bright, while the walls of the raised characters throw it away and read dark. The mark becomes a dark outline on a bright field, insensitive to small tilts, and narrow-angle coaxial versions reach into recesses. In-line illumination covers when specular parts need this; a part inside a metal can is the same physics — coaxial diffuse wins there.

Decision: flat and shiny face → coaxial. Matte, curved or wrinkled → the 45° bar. Both of these small coaxials run from 24 V: the LS-CXL-0092 covers 70 × 70 mm and the LS-CXL-0027 is a 60 × 60 mm unit rated 4.4 W red / 7.2 W white.

Red suits a mono sensor here: silicon responds strongly in the red, and pairing a red light with a matching bandpass filter strips out line lighting and ambient drift — see bandpass filter with red LED light. Which color separates the mark from the field is a material question; part and lighting color for grayscale contrast gives the selection rule.

Camera and signalling: mono, global shutter, one pass/fail bit

Choose mono: relief contrast is grayscale geometry, and color channels add nothing to the threshold while costing sensitivity. Choose a global shutter unless the part is fully stopped and static. The sensor decides on-board and hands the PLC a single pass/fail bit — keep that contract, because images belong on your screen, not on the wire: send pass-fail results, not images. For the trigger-and-result exchange over Ethernet, smart camera to PLC over TCP shows the pattern.

The HK-TD032-12-WBN is a 1.6 MP mono vision sensor with a global shutter, a 1/2.9" sensor, an M12 lens mount and a Fast Ethernet interface.

Drive the LEDs from a DC or strobe light controller; mains-coupled flicker turns up as brightness stripes that mimic defects, and black bars from flicker vs exposure is the diagnostic. For faster indexing, the LS-SMC-0063 is a 1.6 MP GigE smart camera with a 16 mm lens rated at 60 fps, and that 16 mm lens is the worked example's optic.

Positioning decides the cost, not the camera

On a mixed-model line, do not re-fixture for every model — widen the window instead. The window formula is the one from the worked example: window = mark + 2 × allowance + margin. At ±10 mm position tolerance it grows to 3.0 + 20 + 1.5 = 24.5 mm, call it 25 mm, and the pixel count only rises to 250 per side. The real costs are a wider light footprint and a longer locate search, both small at this scale. Set the window per model at changeover and teach that model's mark pattern there.

Run locate-then-verify: an edge-based locate across the window first, and judge the mark only when the locate score clears its floor. A part that never arrived should alarm as "not presented", not count as a bad part — otherwise the position-miss misjudgements you are trying to eliminate simply move into the software. Edge-based pattern matching for stable results is the method.

The handheld case, honestly: a handheld vision sensor that alarms with light and sound is the right tool for audits, first-off checks and recovery after a shift gap. For 100% on-line checking it is not enough on its own, because the part must land in the window every cycle — that is what the stop or indexing fixture buys.

Checks before you commit

• Golden sample of 20 parts — 10 good marks, 5 wrong-model, 5 missing. Set the threshold midway between field gray and shadow gray, and the match-score floor midway between the lowest good score and the highest bad score. The populations must show a clear gap; under about 0.2 on a 0–1 correlation scale, fix the light before the software.
• Tilt test — rock the part ±5° in the fixture. The shadows must still separate; if they do not, move to coaxial.
• Wrong-part test — run one wrong model through deliberately. It must fail identity, not just presence.
• Ambient test — lights on, lights off, dock doors open. If the threshold drifts, shroud the station or add red light with a bandpass filter.
• Cycle test — 100 cycles through the fixture and confirm the mark stays inside the window every time.

The decision, per case

• Fixed station with a stop fixture: thin bar light at 45°, mono global-shutter vision sensor, threshold plus pattern match, one pass/fail bit to the PLC.
• Flat, shiny face: swap the bar for a coaxial light; the camera math from the worked example is unchanged.
• Mixed-model line without per-model fixtures: widen the window to mark + 2 × tolerance + margin, locate then verify, per-model window and pattern at changeover.
• Audit only: the handheld sensor with light-and-sound alarm for spot checks; 100% checking needs the presentation fixture.

Parts list for the fixed station: a standard bar light, a strobe-capable light controller, a mono vision sensor, and the stop fixture. Pick the light from the bar lights or coaxial lights pages — the light finder filters by size and color — and the sensor from the smart cameras and vision sensors page. The lighting configurator steps through camera, lens and light together, and more lighting guides cover the neighboring cases.