In 2D/3D Fusion Inspection, the Bottleneck Is Lighting Interference

LIGHTING

If 2D and 3D take turns imaging a molded part passing at 200 mm/s with a 3 ms gap, the two images end up 600 μm apart, or 77 pixels on the 2D camera.

I worked through the interference that appears when the 2D lighting and the 3D laser share a wavelength band, solved it two ways, wavelength split and time-division, and calculated what it takes to stay within a 10 μm registration tolerance.

About 12 min read

The target is a matte ABS molded part passing at 200 mm/s. Its surface is dominated by diffuse reflection, which makes it a good fit for handing surface scratches and particles to a 2D camera and sink marks and steps to a 3D laser profiler. Once the two channels stand side by side in one station and look at the same spot, the first design decision becomes the wavelength, ahead of the cameras. The moment the 2D lighting and the 3D laser share a wavelength band, each one intrudes on the other’s image.

The interference creeps in quietly. When the white component of the 2D lighting lands in the receiving band of the 3D camera as background light, the peak of the laser line blurs and 40 μm steps start sinking into noise. The usual next move is time-division¹, taking turns between the two channels, but the distance the part moves in between remains as a coordinate offset. A measure meant to prevent missed detections comes back in the shape of registration failures, the leakage cost stays the same, and only the root-cause hunt gets harder.

I calculate the lighting of a fusion station as one optical system. Put the configuration that gives the two channels non-overlapping wavelengths with a band-pass filter on each camera next to the configuration that shares a wavelength but splits time, compare them on the same numbers, and it becomes clear which one fits the line.

Split the wavelengths and the two channels stop seeing each other

The interference comes down to overlapping wavelength bands. Assign blue 450 nm lighting to the 2D channel and a red 660 nm laser to the 3D channel, put a band-pass filter² in front of each camera that passes only its own band, and the two channels can capture at the same instant while barely seeing each other. A signal already buried in background light does not come back through post-processing, so the filter goes on at the optics and the algorithm comes after.

When choosing a filter, I look at the optical density³ of the stop band before the pass band. OD2 cuts the light from the other channel to one hundredth and OD4 to one ten-thousandth. In a layout where a bright source such as the 2D coaxial light leaks toward the 3D camera, one hundredth can leave background light behind, so I set OD4 or higher as the rule.

The left side of the figure is the configuration that splits wavelengths and captures at the same instant, and the right side takes turns with a 3 ms gap. On the left the two channels see the part at the same position, so motion has no room to slip in; on the right the distance moved during those 3 ms lands between the two images as is. The 3.3 mm/s in the lower box is the speed limit when the motion is not corrected at all.

Numbers that remain even with simultaneous capture

Start with the numbers for simultaneous capture. The 2D channel looks at a 40 mm view with a camera 5,120 px wide, so one pixel is 7.81 μm, and the narrowest scratch the 2D side must catch, 30 μm, spans 3.84 pixels. Assuming a sensor with 3.45 μm pixels, a 50 mm lens works at a magnification of 0.44, about 163 mm from the lens center to the part. For the registration⁴ that ties the results of both channels to the same defect, I set the tolerance by my rule at one third of the smallest defect width: 10 μm, or 1.28 pixels on the 2D camera.

Even when the wavelengths are split and both channels capture at the same instant, the part moves during the exposure. At 200 mm/s a 50 μs exposure gives 10 μm of motion blur⁵, or 1.28 pixels, and a 100 μs exposure gives 20 μm, or 2.56 pixels. Keeping blur within the registration tolerance requires exposures of 50 μs or less, so a strobe that delivers enough light in that short window becomes the first condition for simultaneous capture.

In the 3D channel, the point spacing along the travel direction has to be close to the spacing along the line for defects to be captured without gaps. Covering 40 mm with 2,048 points gives a spacing of 19.5 μm, and stacking profiles at that spacing at 200 mm/s takes 10,240 profiles per second, a profile rate⁶ of 10.24 kHz. Assuming the peak position is split down to 1/20 px with the centroid method⁷ at a 30° triangulation angle, the height resolution is 1.95 μm, about one twentieth of a 40 μm step.

Taking turns turns 3 ms into 600 μm

The price of time-division comes out of the same numbers. Assume the gap Δt between the two captures, combining light switching and sensor readout, is 3 ms, and a part passing at 200 mm/s moves 600 μm in between, or 77 pixels on the 2D camera. That is 60 times the registration tolerance, and holding 10 μm with the motion left as is requires a speed of 3.33 mm/s or less, so uncorrected time-division amounts to stop-and-shoot imaging.

That does not make time-division the same as stop-and-shoot. The 600 μm is a known motion that can be subtracted, and the remaining error becomes a question of how precisely that motion is known. Subtracting it at nominal speed leaves 6 μm for a ±1 % speed variation and 12 μm for ±2 %, so staying within 10 μm requires the variation to stay within ±1.67 %. Subtracting the actual motion measured from a transport-side encoder⁸ signal shrinks the remaining error to the encoder resolution plus slip and vibration of the part, and that value is known only by measuring it on the line. Assuming one encoder pulse equals 1 μm, its share is only 0.13 pixels, so if registration still wobbles after correction, slip and vibration are the first suspects.

The time the laser is off counts too. If the laser is switched off for 3 ms during the 2D exposure, about 31 profiles at 10.24 kHz go missing and a band 600 μm wide along the travel direction drops out of the height map. If 2D takes one frame every 30 mm, 2 % of the travel passes without height information, and a sink mark sitting in that band is invisible to 3D.

On shiny surfaces some light gets past the filter

The surface decides how far wavelength assignment works. A surface dominated by diffuse reflection, such as matte ABS, is mostly sorted out by wavelength split alone, but surfaces with strong specular reflection, such as glossy molded faces or plated PCB pads, are a different story. Multiple reflection⁹ occurs as the laser line bounces once more inside the part or off a neighboring face before reaching the camera, and since that light is the same 660 nm as the 3D channel, the band-pass filter cannot remove it.

Multiple reflection turns into false peaks that put points higher or lower than the real surface into the height map. On such a surface I would change the laser line angle and watch whether the false peaks move differently from the true line, and until the profile cross-section of a real specimen has been checked, I treat the numbers above only as a starting point. Whether this configuration catches 40 μm steps reliably on a shiny surface cannot be guaranteed before a sample test.

Setup and judgment conditions for both channels

The setups and judgment conditions of both channels go into one table. 2D and 3D handle different physical quantities, so their minimum defect sizes are written separately, and the third row holds the registration condition that ties the two into one defect.

Channel	① Minimum defect size	② Optics · lighting · WD	③ Algorithm parameters
2D appearance (scratches, particles)	30 μm wide = 3.84 px (7.81 μm/px)	450 nm coaxial light with a low-angle ring, 5,120 px camera (3.45 μm pixels assumed), 50 mm lens at 0.44 magnification, about 163 mm from the lens center, WD confirmed against the reference plane in the lens spec, 450 nm band-pass filter on the camera at OD4 or higher, exposure 50 μs or less	Minimum blob area 12 px (3.84 px wide × 3 px long), ΔGV threshold 9 GV (6 times a noise σ of 1.5 GV, assumed)
3D shape (sink marks, steps)	Height deviation 40 μm (about 20 times the 1.95 μm height resolution)	660 nm laser line, 30° triangulation angle, 2,048 points / 40 mm = 19.5 μm per point, profile rate 10.24 kHz or higher, WD at the center of the head measuring range, 660 nm band-pass filter on the camera	Peak detection by the centroid method, valid peak width 3 to 9 px (assumed), residuals of −20 μm or less after a reference-plane fit as candidates (half the minimum height)
Fusion registration	Tolerance 10 μm = 1.28 px (1/3 of the 30 μm smallest 2D defect)	Calibrate by imaging a dot-grid target with both channels together, and measure the offset between the two optical axes	Six-parameter affine transform, pass at an RMS residual of 1 px (7.8 μm) or less. With time-division, correct the motion with the encoder

Of the three rows, the tightest cell is the registration in the third row, not the lens or the sensor. A 10 μm tolerance holds only when both channels see the same instant or when the motion in between is known within ±1.67 %, and once that condition breaks it collapses regardless of the optical specs. Conversely, for an inspection with a generous 2D defect rule such as 100 μm, the tolerance grows to 33 μm, and time-division with only nominal-speed correction survives speed variation up to ±5.6 %.

Lines where time-division or spatial separation fits

Wavelength split is not always the answer. An inspection that judges color difference or discoloration cannot give up white light, so wavelength split does not work, and on a line that stops for indexed imaging the motion during Δt is zero, which makes time-division the cleanest option optically. If the laser wavelength of a 3D head already on hand overlaps the 2D lighting and there is no budget to replace it, that also argues for time-division. Control of the conveying mechanism itself needs a separate review.

There is a third path as well. Separate the 2D view and the laser line along the travel direction and put up a light shield, and the two see less of each other even at the same wavelength. The time gap between the two channels seeing the same spot grows, though, and separating them by 50 mm means a ±1 % nominal speed variation alone shifts them by 500 μm, so encoder correction becomes a practical prerequisite. Whatever the configuration, a shiny surface calls for specimen checks first.

This judgment is likely to be needed more often. A press release that SNS Insider issued through GlobeNewswire on August 17, 2026 states that the 3D camera segment will grow at 15.83% a year over 2026 to 2035, faster than the 11.89% of the whole market, and that area scan cameras held 56.40% of the market in 2025. I read those two numbers as a sign that configurations putting 3D and 2D together in one station will grow, more than 3D pushing 2D out, and that is why the first page of my setup review is a wavelength layout. This article explains technology and is not investment advice. See our Disclaimer.

Field Note

If the registration error between the 2D and 3D images jumps at a conveying speed of 200 mm/s, I check whether the wavelength bands of the two light sources overlap before suspecting vibration. If they overlap, I keep the band-pass filters in place, switch off the light of one channel at a time and write down how far the background gray value of the other channel’s image drops, and if the line runs time-division, I capture the trigger signals on an oscilloscope to measure the real Δt. Those two numbers are what decide whether to split the wavelengths or refine the motion correction.

Field Checkpoints

With the background gray values and the real Δt written down, check both channels with the questions below before finalizing the spec.

• Do the WD of the 2D lens (from the reference plane in the lens spec) and the center of the 3D head measuring range both match at the same fixture height? (The 163 mm from the lens center is a thin-lens approximation.)
• Is the stop band of the band-pass filters OD4 or higher? (Was the filter chosen by more than its pass band?)
• Is the 2D exposure 50 μs or less? (10 μm of blur, 1.28 pixels, at 200 mm/s)
• Does the 3D head reach a profile rate of 10.24 kHz or more? (Based on a 19.5 μm point spacing along the travel direction)
• Has the registration residual of both channels been measured with a dot-grid target and recorded at an RMS of one pixel or less? (A baseline for when vibration or heat shifts it later)
• With time-division, have the real Δt and the speed variation been measured with an oscilloscope and the encoder? (Within ±1.67 % for nominal-speed correction)
• For glossy molded faces or plated pads, has the profile cross-section of the laser line been checked on real specimens?

Frequently Asked Questions

What about color inspection that needs white light?

An inspection that judges color difference or discoloration cannot give up white light, so wavelength split does not work. In that case go with time-division, and consider either subtracting the motion between the two captures as measured by the encoder or capturing during a stop. The conveying side needs a separate review.

How was the 10 μm registration tolerance chosen?

It is my rule of one third of the smallest defect width the 2D side must catch, 30 μm. At 7.81 μm/px that is 1.28 pixels, so holding the calibration residual to one pixel or less lets the defects seen by both channels be tied to the same defect. As the defect rule grows, the tolerance grows in the same ratio.

Glossary

• [1] Time-division acquisition time-division acquisition
  Capturing two channels separately by switching their lights on in turn with a time gap. It avoids each other’s light at the cost of motion between the two captures. ↩
• [2] Band-pass filter band-pass filter
  An optical filter that passes only a set wavelength band and blocks the rest. It goes in front of the camera lens. ↩
• [3] Optical density optical density, OD
  How strongly a filter blocks light, written as a logarithm. Transmittance is 10 to the power of −OD, so OD2 is 1 % and OD4 is 0.01 %. ↩
• [4] Registration registration
  Aligning the coordinates of two images so the same location maps to the same point. ↩
• [5] Motion blur motion blur
  The stretching of the image along the travel direction because the part moves during the exposure. It is speed × exposure time. ↩
• [6] Profile rate profile rate
  The number of cross-section profiles a 3D laser profiler captures per second. It sets the point spacing along the travel direction. ↩
• [7] Centroid method centroid method
  A method that estimates the peak position more finely than a pixel by taking the centroid of the brightness profile across the laser line. ↩
• [8] Encoder encoder
  A sensor that turns rotation or travel into pulses. It serves as the reference signal for imaging triggers and motion correction. ↩
• [9] Multiple reflection multiple reflection
  Light reflecting once more inside the part or off a neighboring face before entering the camera. It creates false peaks. ↩

References

• SNS Insider, Machine Vision Camera Market Size 2026-2035 press release, GlobeNewswire, 2026-08-17 (3D camera segment 15.83% a year, whole market 11.89%, area scan 56.40% of 2025) Read the release
• Formulas: motion = speed × Δt; motion blur = speed × exposure time; profile rate = speed ÷ point spacing; height resolution δz = (point spacing × sub-pixel fraction) ÷ sin θ; nominal-speed correction residual = motion × speed variation; transmittance = 10^(−OD); thin lens u = f × (1 + 1/m)
• Assumed values: a sensor with 3.45 μm pixels, Δt of 3 ms, 1/20 px sub-pixel peak detection, 1.5 GV noise standard deviation, one 2D frame every 30 mm and a valid peak width of 3 to 9 px are assumptions for the worked example

Related reading — The Defect Zone You Miss When You Switch to 3D Inspection · How 3D Structured Light Measures a Solder Lead Lift That Brightness Cannot Show · When 3D+AI Inspection Answers the Shape Defects 2D Can’t Catch