ND filter exposure control: pick OD by stops

A neutral density (ND) filter is an exposure tool. When the iris is already closed, the exposure time is already at its floor and the strobe pulse cannot get any shorter, the last stops you need have to come out of the optics. The whole selection reduces to one conversion — every 0.3 of optical density (OD) halves the light that reaches the sensor — so once you have counted the stops you are short, the OD follows by a single multiplication. This guide gives the conversion and the OD ladder, the four exposure ceilings that force a filter onto the lens, a worked example that opens a lens three stops without moving the exposure, and the absorptive-versus-reflective decision that matters as soon as near-infrared is in the picture.

Optical density, transmission and stops

Optical density is a logarithmic attenuation scale. A filter of density OD passes the fraction T of the light, and the conversion runs both ways:

The 0.301 is log10(2): one stop is a factor of two in light, and 10^(-0.3) = 0.50, so OD 0.3 is exactly one stop. Each extra 0.3 halves the light again. Exposure time runs the same conversion in reverse: fit a filter, keep the aperture, and the exposure that matched the unfiltered scene multiplies by 10^OD. Filter ranges are usually named by OD × 100, so a part labeled ND060 is OD 0.6 and ND200 is OD 2.0.

The ladder below covers every practical machine-vision case. Read it two ways: count the stops you need and read off the OD to order, or take the OD you have, divide by 0.3, and that is what it buys you.

This guide treats the filter as an exposure decision in stops. The neutral density attenuation sizing guide starts instead from a required transmission and works forward to the density; use whichever direction matches your problem.

The four ceilings that force a filter

An ND filter earns its mount when every ordinary control is exhausted. In practice that means one of four ceilings:

1. Exposure floor. At the frame rate the inspection needs, the shortest exposure the camera accepts still leaves the histogram clipped. Frame rate and exposure are tied together through readout time, so the stops cannot come from a shorter shutter.
2. Aperture fully closed. The iris sits at F22 and the scene is still hot. Stopping down further is unavailable — and unwise anyway, because diffraction blur grows with the f-number:

At λ = 0.55 µm and N = 16 the blur disc is 2.44 × 0.55 µm × 16 = 21.5 µm across — five to six pixels on the 3.45 to 4 µm pixels common in machine vision. That diameter is an estimate at mid-visible wavelength; the growth with f-number is exact.

3. Fixed strobe pulse. The LED driver has a shortest pulse that still lights the field uniformly, and shortening it further leaves the pulse half-formed and the illumination uneven. The energy per pulse has to leave optically.
4. You want the iris open. Depth of field is the goal: throw background clutter out of focus, or move to the f-number where the lens performs best. Opening the iris adds light, and something must remove it.

One trap sits in the control loop: an automatic exposure or iris algorithm will quietly cancel a fixed filter by opening exposure or gain back up. Set the exposure manually first — how automatic iris, exposure and gain behave — and treat the filter as part of the optical setup from then on.

Worked example: open F8 to F2.8, keep the 2 ms strobe

The camera is the MV-CS200-10UM, a 20 MP mono USB3.0 model on a 1" sensor with a C-mount. The lens is the LS-FFL-0567, a 16 mm C-mount lens rated F2.8-22 on a 1.1" format at 20 MP, with <0.3% distortion.

The setting: a steel fitting on a roller conveyor, strobed. The driver's shortest uniform pulse is 2 ms, and at F8 the image sits at target gray. Background rollers and bracket edges keep triggering false calls, so you want F2.8 to throw them out of focus.

The arithmetic:

• F8 to F2.8 runs F8 → F5.6 → F4 → F2.8: three stops.
• Three stops multiply the light by 2^3 = 8.
• OD to remove them: OD = 0.301 × 3 = 0.903, so order OD 0.9.
• Check: T = 10^(-0.9) = 0.126. Wide open, the filtered image sees 8 × 0.126 = 1.01 times the F8 level. Exposure stays at 2 ms and the histogram holds.

The nearest coarser stock shows what moves the answer. With OD 1.2 instead, 8 × 0.063 = 0.50 lands the scene one stop dark, so the pulse stretches to 4 ms or a stop of gain goes in — both worse than picking the density the arithmetic names.

Do not trust a depth-of-field preview to sign off the change; simulated previews overstate what stays sharp — why sharp previews mislead. Shoot the real part at F2.8 with the filter fitted before the recipe is locked.

Counting the shortfall from the histogram

For the saturated case the stops come straight out of the image. Halve the exposure time repeatedly until clipping just disappears; each halving is one stop. Suppose the exposure floor is 50 µs at your frame rate and clipping only clears at 12.5 µs. You are two stops over: OD = 0.301 × 2 = 0.6, T = 25%, and at the 50 µs floor the filter brings the sensor back to the 12.5 µs equivalent (50 × 0.25 = 12.5). Check which units the raw value reports before halving anything — interpreting raw exposure and gain values.

Lamps, arcs and molten surfaces are a different order of shortfall: six to ten stops, not two. That is OD 2.0 (1% transmission, 6.6 stops) to OD 3.0 (0.1%, 10 stops), and at those levels the reflective decision in the next section stops being optional.

Absorptive or reflective: where the rejected light goes

Both types deliver the stated OD across their rated band; they differ in where the rejected photons go and in how far the rating extends past the visible.

Absorptive glass is the default on visible LED lines: neutral where the sensor works, stackable, and nothing reflects back into the lens.

Near-infrared is where the choice stops being academic. Absorptive ND glass is usually characterized in the visible, while silicon keeps sensing out to roughly 1000 nm — mono sensors furthest, since they carry no color filter stack. A visible-rated OD 2.0 filter can pass a large share of an 850 nm LED array's output, and gray contrast quietly collapses. When the illumination is a narrow IR band, select the band instead of attenuating everything: a band-pass matched to the LED peak sets contrast and attenuation in one element. The LS-LXR-0225 is an 850 nm band-pass with an M30.5 x 0.5 thread — the kind of part that replaces guesswork on an IR line.

On color cameras the same leakage arrives as IR contamination, and the fix is an IR-cut filter rather than an ND: IR-cut filter options for color cameras. If the inspection band itself is the open question, SWIR vs NIR vs visible imaging frames the choice.

Stacking filters

Densities add. OD 0.6 plus OD 1.2 in series is OD 1.8, and the transmissions multiply to prove it: 25% × 6.3% = 1.6%, and 10^(-1.8) = 1.6%.

Three habits keep a stack honest:

• Mix types, never match them: two reflective filters facing each other bounce light between the coatings and print a ghost pair beside every highlight. Pair an absorptive with a reflective instead.
• Coated, mounted filters carry their surface reflections in the stated density, so the sum holds. Bare glass rated by internal density adds Fresnel reflections of about 4% per uncoated surface; four surfaces pass 0.96^4 = 0.85 of what the sum predicts, about 0.07 OD denser.
• Keep stacks short anyway: every extra element costs margin and adds a surface to keep clean.

Fit and ordering checks

• Front thread. Match the lens front thread exactly — the band-pass above is M30.5 x 0.5 — or order unmounted blanks for a slip mount. The lens accessories category lists filters, prisms and targets by type and mount.
• Density at your wavelength. Read the transmission curve at the LED's peak, not at 550 nm; the number that matters is the OD there.
• Refocus after fitting. A plane-parallel plate in a converging beam shifts the focal plane slightly; refocus and lock, and re-check after the line's temperature settles — the locking screws guide covers keeping the setting put.
• Coated surfaces when a bright source sits in frame, to keep its reflection from ghosting.

When an ND filter is the wrong tool

• Glare on a shiny part. An ND dims the specular highlight and the signal together. Crossed polarizers suppress the highlight itself: cross polarization to reduce surface glare.
• Band contrast. If the defect only separates in one band, the filter is chosen for contrast first: choosing machine vision optical filters.
• A dark image. An ND only lengthens the exposure you are trying to shorten; the fix lies in lighting, not attenuation.

The decision, case by case

Count stops first, multiply by 0.3, order the mount that matches the front thread, read the curve at your wavelength, then refocus and lock. Start in the accessories finder, or assemble camera, lens and filter in the configurator started at extras. If the stops will not add up, send the scene to us through the contact page — and the lenses topic page collects the rest of the optics decisions.