LED binning: colour and brightness variation between orders

Re-order the light that passed validation, swap it in, and the histogram can sit a few counts lower, or the white tint can look slightly different to your eye. Nothing is broken. LED chips are sorted at the factory into brightness and colour groups called bins, and a repeat order can be built from a different bin set than the original. Here you get the numbers behind that variation: how large one bin step is in percent and nanometres, what it does to an 8-bit image and to a threshold set at 150 counts, and the two moments to act — order time, where you state any tight tolerance, and swap-in time, where a reference target and a drive trim keep results identical.

What LED binning sorts

No two LED chips emit exactly alike. Epitaxial growth and wafer processing vary run to run, so brightness and colour spread across production. Every chip is tested on a sorting machine that measures its output and colour, and chips are grouped into bins: narrow bands of brightness and colour. Two chips from different bins look identical under a magnifier and differ slightly in the light they give.

Three properties carry bin codes. Luminous flux sorts brightness, measured as radiant flux in milliwatts for infrared emitters. White LEDs are also sorted on chromaticity, their position around the nominal colour temperature, which you read as tint. Monochromatic LEDs — red, green, blue, infrared, ultraviolet — are sorted on peak wavelength in nanometres. Some ranges add a forward-voltage bin, which matters to the driver more than to the image.

Every light in the shop carries that sorting: the bar lights over a conveyor, the ring lights around a lens, the emitters tabulated on the wavelength ranges from UV to IR page. Form factor does not matter; the chips inside are binned either way.

How large one bin step is

Treat these as industry-typical grids, because each LED maker defines its own bins and revises them between generations. Flux bins step about 5-10% in output. Wavelength bins on coloured LEDs run roughly 1-5 nm wide. White chromaticity bins are a few tint steps around the nominal colour temperature. Two adjacent flux bins therefore differ by up to about a tenth of the light output, two adjacent wavelength bins by a few nanometres, and nothing in the part's appearance says which bin it came from.

White tint is where eyes beat cameras: a chromaticity step can be visible side by side yet vanish in a histogram. If you specify white light, pin the colour temperature and its tolerance, as the white LED color temperature guide sets out. A repeat order of a white diffused bar light such as the LS-BAR-1276 can arrive a touch warmer or cooler than the unit it replaces while the camera sees almost nothing.

Coloured lights sort on peak wavelength instead. The LS-RNG-0100 ring light is specified at 620 nm. A replacement order can land in the adjacent bin a few nanometres away, and neither unit is faulty.

What a bin change does to an image

At fixed exposure, gain and aperture, a flux-bin change scales every pixel by the same factor: the whole image brightens or dims together, geometry untouched, only grey values move.

Here ΔΦ/Φ is the fractional change in light output. On an 8-bit camera, a 10% step shifts a 160-count background by 16 counts, to 144 (160 × 0.10 = 16; 160 - 16 = 144). A 5% step lands halfway, at 152. Whether that hurts is decided by a single number: the distance between your grey levels and your threshold, in counts.

Colour behaves differently. On a monochrome camera, part-versus-background contrast comes from the reflectance difference at the LED's wavelength, the relation worked through in the LED colour and contrast guide. A few nanometres of shift move both reflectances slightly, so contrast changes by a small fraction. The effect grows where contrast rides a spectral edge — a UV or near-IR emitter on the steep shoulder of silicon's response, or a narrowband filter. Mid-visible red and green sit on flatter ground, and the same nanometres barely register.

Worked example: does the replacement break the threshold?

Take an illustrative station and follow the counts:

• 8-bit monochrome camera, manual exposure and gain, white diffused light overhead
• good-part background mean: 180 counts; dark defect mean: 120 counts
• threshold: 150 counts, 30 counts of margin on each side
• replacement light: adjacent flux bin, assumed 10% less output at the same drive and exposure

With the new light, the background scales by 0.90: 180 × 0.90 = 162 counts, and the margin above the threshold falls from 30 to 12 counts (162 - 150 = 12). The defect scales to 120 × 0.90 = 108 counts, and its margin grows to 42 counts (150 - 108 = 42). Inspection still passes, but the background headroom shrank by 18 counts, and the next thing that dims the image eats into it.

Ageing is the companion effect. A light at its L70 point emits 70% of its initial output — that is the definition of L70 — and the LED lifetime and strobing guide shows why strobed operation reaches it later. Stack a 10% bin step on end-of-life output: 0.90 × 0.70 = 0.63, so the background lands at 180 × 0.63 = 113 counts, below the 150 threshold, and good parts start failing. The margin you budget must cover the bin step plus years of service, or you trim the drive at every swap.

Why one order matches and the next does not

Light makers track the bins that go into each product and usually build one purchase order from one production run, often drawing on just two or three adjacent bins across the whole product. The units within your order match each other closely. Place the same order again months later and it can be built from a new run with a different bin set, so order-to-order consistency is a small step, not an exact match.

A step within the maker's normal bin range is not a defect and gives no grounds for rejection. It becomes your problem only where the application leaves little margin: colour measurement, spectral-edge contrast, golden-image comparison. Where validation genuinely needs tight wavelength or brightness, state the tolerance in nanometres and percent before ordering and have it agreed on the order — the contact page is the place, with the light's model and the old unit's measured output.

Multi-light stations add a wrinkle: swap one unit of three and the different bin can show as a brightness seam across the image. Trim each swapped unit individually until the reference target reads level again.

When a bin step matters and when it does not

The margin count decides, so put your own numbers against these rows:

How to keep binning from moving your results

At swap-in, hold the camera fixed: manual exposure and gain, every auto mode off. Auto exposure quietly stretches the exposure time to hide a dimmer bin and hands you motion blur in exchange, the trap described in the turning every auto mode off guide.

Then measure. Image a reference target with the old light and keep the file; repeat with the new light and compare the ROI mean, using the ROI method in the light intensity measurement guide and the target practice from the reference target for illumination guide. Bring the new mean back to the old with drive current or exposure, inside the light's rated drive — not by editing thresholds.

Trim on the light controllers, not the camera, where you can. A constant-current controller such as the LS-CTL-0180, 4 channels and 120 W for 24 V lights, holds the set current against supply wobble, so the trim stays where you put it; the current control guide covers setting it without overdriving.

Strobed stations need one extra check: average output follows peak current times duty, so a changed pulse width masquerades as binning. The duty cycle matching guide works that arithmetic. A digital controller such as the LS-CTL-0487, 2 channels at 24 V and 240 W, stores per-channel current and timing, so a swapped light is matched in software.

If you trim with an analog control voltage instead, expect output to track the control value non-linearly; the DAC LED brightness non-linearity fix page works through why current drive keeps the match proportional. For stations that must not drift at all, close the loop: the camera reads a reference patch every cycle and the controller trims the drive, the scheme in the closed-loop lighting control guide.

Checks before you blame the light

A swapped light that looks different is not automatically a binning story. Work down this list; binning is what remains when everything else matches:

1. Controller channel carried over? Read back the set current or voltage on the new channel; a channel left at a default explains most dimmer-light reports.
2. Strobe timing identical? Average output scales with peak current times duty, so a different pulse width looks exactly like a different bin.
3. Voltage at the connector? A long cable on a constant-voltage light drops volts before the LEDs see them.
4. Camera auto modes off? Auto exposure and auto gain hide one change and create another.
5. Ageing or binning? Ageing drifts over months of running; binning appears in the instant of the swap and tracks the order date.
6. A clean step of 5-10% with everything matched? That is binning. Trim drive until the reference target reads the old mean and carry on.

The decision, in one pass

Expect a one-bin step of about 5-10% in brightness and a few nanometres in wavelength between orders, treat it as normal rather than a fault, keep threshold margins that cover it — 40 counts or more on an 8-bit image where the optics allow — and re-baseline every swap against a reference target, trimming drive current instead of thresholds. Where the application cannot absorb a bin step, write the tolerance into the order before you buy. That is the whole discipline.

If you are speccing lights or replacements now, the light finder filters by size, colour and wavelength, the controller finder by channels and output type, and the lighting step of the configurator walks the whole chain; the lighting topic page collects the rest of the reading. For a tolerance clause or a matched replacement, contact us with the light's model and your reference-target numbers.