A silhouette backlight gives contrast for free: the part blocks, the background passes, and any camera sees an edge. The harder family of jobs asks the light to pass partly through — a fill level through a coloured bottle, a copper trace under solder mask, an insert inside a sealed paper pouch, a missing boss in a translucent housing. There the wavelength decides whether the internal feature reads at all. You get the default band order to try, a bench test with the arithmetic that makes two colours comparable, the counterexample where blue beats red and infrared, and the light to order for each case.
Which wavelength should you try first on a semi-transparent part?
Start with red near 630-660 nm, then near-infrared: for most semi-transparent materials the longer wavelength transmits more deeply, and blue near 450-470 nm scatters or absorbs first in paper, plastics and glass.
Light that enters the wall is either absorbed by dyes and binders or scattered by internal structure. For scatterers small against the wavelength, scattering falls as the fourth power of wavelength, so blue is stripped out of the image path first while red and near-IR slip past fibres, fillers and pigment particles (transparency and translucency). Glass passes red and near-IR readily, and paper or nonwoven pouches are the classic case where red reads what blue cannot.
The longer near-IR peak needs one extra thought: it sits near a water absorption band, so moisture-rich contents darken it faster than the 850 nm band. Reach for it when the shorter band over-penetrates and you want the contents themselves to come back darker.
The bands worth having on the bench:
| Band | Typical LED peak | Usually wins at | Watch for |
|---|---|---|---|
| Blue | 450-470 nm | Coloured glass that blocks red and IR | Strongest scattering in fibrous walls |
| Deep red | 730 nm | Mid case: more depth than red, less than IR | Fewer stocked models |
| Red | 630-660 nm | Paper pouches, presence checks, first trial | Red-absorbing dyes |
| Near-IR | 850 nm | PCB internals, pigmented walls, dry fills | Falling silicon sensitivity |
| Near-IR | 940 nm | Contrast when the shorter IR band over-penetrates | Water-rich contents absorb it |
The backlights category filters by colour and by stocked wavelength (405, 730, 850 and 940 nm), so every band in the table is orderable through the light finder. Reflected-light colour contrast on opaque surfaces is a different problem; that choice is covered in which LED colour gives the best contrast in machine vision.
How do you compare two wavelengths fairly on the bench?
Fix the drive current and the camera exposure, record the background gray level and the through-part gray level in DN, and compare normalized transmission T = (G_part - G_dark) / (G_air - G_dark); the higher T wins.
Raw DN mislead because LED output, lens transmission and sensor response all differ band to band: a through-part reading measures the whole chain, not the part. T divides out the chain by referencing every band to its own air reading, so it reports only what the part itself lets through.
Run it in five steps:
- Set both panels to the same drive current and fix the camera exposure; 100 µs is a workable start.
- Cap the lens and shoot a dark frame; the patch mean is G_dark.
- Remove the part and shoot an air frame; the patch mean is G_air.
- Seat the part without touching focus or exposure and shoot again; the patch mean is G_part.
- Compute T per band; the higher T penetrates deeper.
The second line sets band B's exposure so its air level matches the reference arm A, which is what you want when one threshold must serve both arms.
For a shared threshold, set the IR exposure to 100 µs × 184 / 174 ≈ 106 µs. The through-part level then lands at 16 + 102 × 1.06 ≈ 124 DN, with 131 DN of headroom to saturation. Decision: near-infrared for internal detail.
The arithmetic assumes the sensor is linear below saturation and the light holds steady between frames; shoot the air and part frames seconds apart, and repeat the pair if either mean moved.
Run the red arm from a flat red panel sized to the part:
- Size
- 100-200 mm
- Type
- Standard
- Voltage
- 24 V
The LS-BKL-0416 lights a 136.4 x 136.4 mm field at 850 nm and runs constant-on from 24 V with no driver. Because it cannot pulse, match the two arms with camera exposure time rather than pulse width.
- Size
- 100-200 mm
- Type
- Standard
- Voltage
- 24 V
- Wavelength
- 850 nm
When does blue beat red and infrared?
When the container itself absorbs the long wavelengths: a coloured glass bottle that blocks red and near-infrared can still pass enough blue light to read a liquid fill level, and the usual penetration order reverses.
The colourants are the container's absorption spectrum. Red and amber glass take their colour by soaking up the red end and the near-IR, and some pass blue with little loss, so the package composition inverts the default order while every rule from the bench test still applies.
That does not rehabilitate colour as a predictor in general. On an opaque part inspected in reflected light, colour contrast is the signal itself. Through a semi-transparent wall, the absorption and scattering spectra are the signal, and the visible colour is only their shadow, which is why the bench reading settles the band and the colour chart does not.
Example readings on a coloured glass bottle holding clear liquid, same camera and 16 DN dark level: red air 200 DN, through 18 DN, T = 2 / 184 = 1.1%; IR air 190 DN, through 17 DN, T = 1 / 174 = 0.6%; blue air 210 DN, through 120 DN, T = 104 / 194 = 54%. Blue transmits roughly 50 times what red manages, and under red the fill edge would sit 2 DN above dark, below any safe threshold. Decision: blue for that container, ordered through the finder's colour filter like any other panel.
For the geometry side of fill-level work, two guides pick up where wavelength stops: liquid level inspection lighting for transparent containers and backlight contrast for sharp meniscus edges. Differential penetration is one member of the wider family covered in how to inspect internal features with machine vision.
Does more intensity fix a wavelength that will not penetrate?
It does not: transmission is a ratio, so doubling the drive current multiplies background and part signal together and leaves T unchanged; more watts pay off only once the band is right and the problem is photon starvation, ambient light or corner falloff.
The ceiling arithmetic makes it concrete. In the coloured-bottle case above, T_red = 1.1%; pushing the air level from 200 DN to a practical 8-bit ceiling near 250 DN lifts the through-part level only to 16 + 0.011 × (250 - 16) = 18.6 DN, about 3 DN of signal above dark. The band is wrong, not the wattage, and no supply fixes that.
Intensity is the right lever in three other situations: the line speed forces an exposure so short that even a good wavelength is photon-starved; ambient light swamps a constant-on panel; or a large field falls dim at the corners. A high-power panel pulsed in time with the exposure delivers the dose without changing the wavelength:
- Size
- 400 mm +
- Type
- High power / strobe
- Voltage
- 24 V
The LS-BKL-0865 spreads 92.2 W across a 400 x 300 mm lit area, the large-field end of the high-power type.
Should a collimated backlight be part of the test?
Yes for gauging edges or wall thickness through a thick or curved translucent part, where a collimated panel restricts the cone so rays cross the wall on a near-normal path; for presence and absence a standard panel is enough.
A standard panel radiates over a wide cone. Through a thick or curved wall, oblique rays take a slanted, longer path, and the internal edge you are gauging smears as the part moves or the surface curves. A collimated beam crosses the wall on a short, repeatable path and holds internal edges in place, at the cost of alignment care: a narrow cone tolerates less panel tilt and wants the working distance held steady.
Keep the split simple. Presence, absence and orientation read fine on a standard panel, whose wide cone even helps fill the field at short working distance. Gauging an edge or a wall thickness through the part is the collimated case. The trade-offs are worked through in collimated backlight or diffuse backlight for edge gauging, and the distance and sizing rules for the collimated layout are in collimated backlight setup: working distance and light size.
Size the collimated panel a little larger than the part:
- Size
- 100-200 mm
- Type
- Collimated
- Voltage
- 24 V
The LS-BKL-0830 is a 120 x 120 mm collimated backlight.
Which controller settings make the wavelength comparison valid?
Give each colour its own channel, drive both arms at the same current, and strobe the light only during exposure; a strobed channel removes the ambient light that would otherwise sit on top of a part signal worth a few DN.
One channel per colour keeps the drive independent, so switching bands never silently changes intensity. Constant-current output holds LED flux steady while the junction heats, which matters because a long bench session warms panels and drifts constant-voltage arms; both output types exist across the light controllers range, with strobe, channel count and power filterable.
Strobing also removes the shop lights. Ambient adds to every reading, and on a part signal worth a handful of DN, a 5 DN ambient offset is not a measurement. Mains-powered ambient beats against the exposure as well; the failure mode and the timing fix are in camera flicker under artificial lighting: diagnose and fix.
Sequencing closes the loop for production: run red, infrared and blue arms back to back in one fixture and keep whichever band the recipe calls for, the pattern described in multi wavelength light: sequence colours to get more from one camera.
A digital strobe controller with a channel per arm covers the test:
- Output
- Constant voltage 24 V
- Channels
- 4 channels
- Control
- Digital
- Feature
- Strobe
The LS-CTL-0070 is a 4-channel digital controller rated 200 W with a strobe feature and RS232 control.
What else in the optical path biases a wavelength test?
Sensor response, lens coatings and stray filters all change with wavelength: silicon sensitivity falls toward the long near-IR, Bayer colour sensors pass little beyond roughly 650-700 nm, and an IR-cut window in the path can erase an infrared image entirely.
Silicon's near-IR falloff is why the IR air frame read 190 DN against red's 200 in the worked example. T absorbs that difference, but the absolute signal is lower, so budget extra exposure or add a bandpass filter matched to the LED peak, which also rejects ambient colour and sharpens the part edge.
Lens coatings transmit unevenly across the visible and near-IR, and the mechanism is laid out in lens coating and spectral transmission: why contrast drops. Past 940 nm you leave silicon's comfortable range, and the material-by-band trade-offs are set out in SWIR wavelength selection by material: silicon, moisture, plastics.
Which light do you order for each case?
Match the light to the decision the image must make: soft packaging presence runs white or red on a standard panel, internal detail runs near-infrared on a monochrome camera, coloured-glass fill levels run blue after a bench test, and gauging through the wall adds collimation.
- Presence and absence in sealed flexible packaging: red or white, standard panel, constant-on. The LS-BKL-1468 lights a 150 x 150 mm field. With 10% margin per side that is a 120 x 120 mm part envelope, and stocked fields run from under 50 mm to 400 mm and up.
- Internal detail through boards, housings and pigmented walls: short-wavelength near-infrared, monochrome camera, with the bench T telling you how much exposure the line speed leaves you.
- Fill level through clear containers: clear glass and clear plastics pass red and near-IR readily, so red is a sound default; coloured glass is the blue case from the bench section, and the near-IR options are compared in NIR lighting for fill level inspection through packaging.
- Gauging through a thick translucent wall: collimated panel at whichever band the bench test chose; wavelength first, collimation second.
Set the wavelength, field and channel count in the backlight configurator, and size the strobe side in the light controller finder. If the best T on your part sits within a few DN of dark on every band, send the readings through the contact page and the search can widen to other bands or SWIR optics.
Questions
Does the part's colour predict which wavelength will penetrate?
Not by itself. Transmission through a semi-transparent wall follows the dye, filler and crystal structure across the spectrum; the visible colour is only part of that picture, and two parts of identical colour can transmit red and infrared very differently. The matched-exposure reading of T settles it in minutes on the actual part.
Why does a colour camera see almost nothing under an infrared backlight?
Colour sensors sit behind a Bayer mosaic and usually an IR-cut filter that block most energy beyond roughly 650-700 nm, so a near-IR panel reaches the sensor at a tiny fraction of its output. Run wavelength work on a monochrome camera with no IR-cut window, and check that the lens itself transmits the band.
Can one station run red, blue and infrared without swapping lights?
Yes, in two ways: lights that combine several emitter bands in one body, addressed band by band through a sequencing controller, or separate panels on separate channels of one multi-channel controller. Either way each arm keeps its own drive current, and the stocked bands (405, 730, 850 and 940 nm) bound what a changeover can reach.
How large should the backlight field be compared with the part?
Cover the part with 10-20% extra lit area on each side so normal conveyor wander never brings a dark edge into the measurement window, and keep the camera's field matched to the lit area. A field larger than the camera view is harmless; an undersized one shows brightness steps that break fixed thresholds.
What exposure time can a moving line tolerate under a backlight?
Work backwards from the blur budget: t (ms) = blur (mm) / speed (mm/ms). At 300 mm/s = 0.3 mm/ms, a 0.1 mm blur budget allows about 0.33 ms, or 333 µs; slower speeds scale it up linearly. If the part signal lands too low in that time, raise dose with a strobe or a brighter panel before changing band.