The Defect Zone You Miss When You Switch to 3D Inspection

OPTICS

A 25 μm wide defect that fills three pixels on a 2D line-scan shrinks to 1.25 pixels once it moves to a 3D profiler with 20 μm/px xy resolution.

I calculated the empty zone that opens between planar floors of 25 μm and 60 μm, and looked for the rule that splits defect groups between 2D and 3D in the 2D contrast ΔGV rather than in defect size.

About 11 min read

The targets are a PCB covered in semi-matte solder resist and a glossy ABS molded part. Solder resist scatters light fairly evenly while glossy ABS is dominated by specular reflection, so the two produce entirely different images under the same light. It is easy to expect fewer missed detections after swapping the 2D line-scan inspection of these two surfaces for a 3D laser profiler, but the calculation shows that in the planar direction the smallest visible defect actually gets larger.

If nobody confirms with numbers at the setup stage that a 3D device has coarser planar resolution than 2D, the whole group of defects 25 to 60 μm wide slips through. Micro-bridges on the PCB and small black spots and particles on molded parts fall in this range. They are too small to span even three 3D pixels, so no amount of algorithm work captures them, and once leakage starts, downstream re-inspection and return costs quickly outrun the price of the equipment.

When I evaluate a move to 3D, I measure the contrast that actually comes out in 2D before I look at defect size. Only after calculating the planar and height floors of both 2D and 3D and splitting the defect groups by 2D contrast ΔGV do the defects that should go to 3D become clear. Keeping that order shows, before the equipment arrives, which defects the numbers in a 3D quote newly catch and which ones they miss.

3D draws the money, but planar resolution does not follow

The flow of money toward 3D inspection shows in numbers too. A press release that SNS Insider issued through GlobeNewswire on August 17, 2026 projects the machine vision camera market growing from about USD 6.79 billion in 2025 to USD 20.90 billion in 2035 at an 11.89% compound annual rate, with 3D cameras growing fastest among product types at 15.83% a year. In the same release, inspection accounts for 38.60% of the 2025 market.

On the shop floor that growth rate often turns into the expectation that adding 3D replaces 2D. A laser triangulation¹ profiler reads height down to the 1 μm range through sub-pixel² peak detection, but in the planar direction it stays at the camera pixel spacing. If a 2D line-scan³ on the same target works at 8.33 μm/px, a 3D head at 20 μm/px xy sees the plane 2.4 times more coarsely. 3D is a device that adds the height 2D cannot see, and it does not inherit the small defects 2D used to see.

The figure is a map with the planar size and step height of a defect on two axes. Below 25 μm planar is a blind spot where neither device fills three pixels, between 25 μm and 60 μm is a zone only 2D can capture, and color or contamination defects with steps under 3 μm stay with 2D because 3D cannot tell them apart by height. Defects such as silk-screen bleed, with a height difference around 1 μm, sink into the background of a 3D height map even when they approach 100 μm across. 3D does its work in the upper-right zone at 60 μm planar or more with steps above 3 μm, and even there the overlap with 2D has to be sorted again by contrast.

The planar floor set by the three-pixel rule

My rule is that the smallest defect has to span at least three pixels to be caught reliably. For the 2D line-scan, a 4,096 px sensor with a 7 μm pixel pitch behind a 0.84x telecentric lens⁴ gives an object-side pixel resolution⁵ of 8.33 μm/px, a 34.1 mm view and a planar floor of 25 μm. The 3D profiler at 20 μm/px xy raises the planar floor to 60 μm.

Converting the defects between the two floors into actual pixel counts makes the gap plain. A 26 μm micro-bridge spans 3.12 pixels in 2D and 1.30 in 3D, a 40 μm dent spans 4.80 and 2.00, and a 55 μm black spot spans 6.60 and 2.75. Every defect that cleared three pixels in 2D drops below three in 3D. The three-pixel rule comes from pixel boundaries. With a width of three pixels, at least two pixels are completely filled by the defect wherever it lands, so its contrast survives intact, but with a width of a little over one pixel, a defect straddling a boundary splits half and half between two pixels and its contrast drops by half. That is why the same defect shows on one part and disappears on the next.

Height works differently. When the camera views the laser line at a triangulation angle⁶ θ, a height change of δz moves the line sideways by δz × sin θ. Solved the other way, δz = (xy resolution × sub-pixel fraction) ÷ sin θ, and assuming peaks are split down to 1/20 px at 20 μm/px, that gives 1.41 μm at 45°. I add a factor of two as margin and use 3 μm as the step floor.

Raising the triangulation angle improves height resolution, but it has a price. While resolution improves from 2.0 μm at 30° to 1.15 μm at 60°, the occlusion⁷ zone behind a 1.5 mm tall component, where the camera cannot see, grows as h × tan θ from 0.87 mm to 2.6 mm. Widening the angle also pulls the camera and laser closer to the target, so whether the WD of the 3D head clears the fixtures and handlers decides adoption as much as the optical specs do.

A ΔGV of 9 GV divides 2D from 3D

Sort defects by size alone and you get lost in the 25 to 60 μm zone. So I first measure the brightness difference between defect and background, the ΔGV⁸, in the 2D image after the lighting has been tuned. In an 8-bit image with sensor noise at a standard deviation of 1.5 GV, the contrast floor for stable detection is 3σ, or 4.5 GV, and 9 GV, with a factor of two as margin, becomes the line that decides whether a defect stays in 2D.

Moving a defect above 9 GV to 3D only throws away planar resolution. Conversely, a defect that stays below 9 GV even after lighting changes, with a step of 3 μm or more and a planar size of 60 μm or more, is caught only once it moves to 3D. The remaining case is the hard one. A low-contrast defect between 25 and 60 μm planar is not caught even in 3D because it lacks pixels, so the only remedy left is to redesign the angle and wavelength of the 2D lighting to create contrast. Written as code, the branch looks like the block below, and its four return values map to the blind spot, keeping the defect on 2D, routing it to 3D and redesigning the lighting.

# 2D/3D routing decision — branch on contrast, not defect size
XY_2D_MIN, XY_3D_MIN, Z_3D_MIN = 25.0, 60.0, 3.0  # planar/step floors [um]
DGV_MIN = 4.5      # 3-sigma contrast floor at 8-bit noise sigma=1.5 GV

def route(xy_um, z_um, dgv):
    """xy: defect planar size [um], z: step height [um], dgv: gray contrast after lighting optimization [GV]"""
    if xy_um < XY_2D_MIN:
        return "Blind spot -- pixel resolution needs to go up (magnification redesign)"
    if dgv >= DGV_MIN * 2:              # contrast already sufficient in 2D -> 3D not needed
        return "Keep on 2D"
    if z_um >= Z_3D_MIN and xy_um >= XY_3D_MIN:
        return "Route to 3D"                 # has a step, but contrast does not come up in 2D
    return "Redesign 2D lighting and re-judge"      # a zone even 3D resolution cannot cover

The last return value weighs the most in practice. The moment anyone assumes that software post-processing can make up for a shortfall in the optical setup, this zone stays undetected for good, and only resetting the lighting and magnification creates both the contrast and the pixel count this zone needs. When a bridge around 25 μm wide lands on this path, I put the option of raising magnification for more pixels and the option of changing the lighting for more contrast in one table and compare their cost.

Running three defect groups through the routing function

I pick three of the defects plotted in the figure and run them through the function. The contrast values are examples assumed to be measured after the lighting is tuned; the optics column holds the setup of the device each defect is assigned to, and the algorithm column holds the branch the function returned and the judgment conditions that follow it.

Defect group	① Minimum defect size	② Optics · lighting · WD	③ Algorithm parameters · branch
PCB micro-bridge	26 μm planar (2D 3.12 px, 3D 1.30 px), step about 1 μm	2D line-scan 4,096 px, 0.84x telecentric, low-angle dark-field lighting, WD per the lens spec (to be confirmed)	ΔGV 12 (assumed) → keep on 2D. Candidates at 9 GV contrast or more, minimum blob 3 px
PCB insufficient solder volume	300 μm planar, 28 μm step (about 20 times the 1.41 μm height resolution)	3D laser triangulation at θ 45°, 20 μm/px xy, WD per the head spec, 1.5 mm occlusion behind a 1.5 mm tall component	ΔGV 5 (assumed) → route to 3D. Group residuals of 3 μm or more after a reference-plane fit as candidates and judge by volume
Molded-part surface dent	40 μm planar (2D 4.80 px, 3D 2.00 px), 4 μm step	Keep the current 2D setup and redesign only the lighting angle and wavelength; check that the new light fits within the WD	ΔGV 5 (assumed) → redesign 2D lighting and re-judge. Below the 60 μm 3D planar floor, so it cannot be routed to 3D

The value that splits the three rows is ΔGV, ahead of defect size. Even the same 26 μm bridge has nowhere to go if its contrast stays under 9 GV, since it spans only 1.30 pixels in 3D, while the 300 μm solder volume shortfall is comfortably caught by 3D thanks to its 28 μm step even with low contrast. For a low-contrast defect below 60 μm planar, such as the dent, resetting the lighting before asking for a 3D quote costs less.

Conversely, lines where 3D alone is enough

Some lines call for the opposite conclusion. For a product whose surface color and material are uniform, where color or contamination defects never occur and only shape deviation is controlled, removing 2D and keeping only 3D costs little. The same goes for a line where plating stains or gloss variation make the 2D threshold drift from lot to lot and false rejects pile up. Height data sways less with changes in surface reflectance, so the threshold actually becomes steadier. And when the control spec itself is defined in dimensions or volume, with no basis for converting gray values into the spec, 3D is the better fit. All three cases share defects defined by height or shape and no separate control of small color defects below 60 μm planar.

On strongly specular mirror plating, hairline-finished surfaces with a grain, and translucent molded parts, though, this judgment does not hold as is. In translucent resin the laser digs below the surface and scatters, widening the peak of the line, while on mirror plating the reflection either misses the camera or bounces once more off a neighboring face and creates a false peak. Whether 3D does better than 2D on such materials cannot be guaranteed before a sample test. In my case, I would measure the peak width and height standard deviation of the line profile on real specimens before deciding.

One line of calculation to write before the quote

When a quote for a 3D head arrives, I multiply the xy resolution in the spec sheet by three to get the planar floor and write it next to the smallest defect spec under control. That one line shows the missed-detection zone after adoption in advance, and writing the triangulation angle and WD alongside it brings mechanical interference into the same table. A routing table in numbers is also useful to whoever draws up the budget, because the return costs of a leaked defect zone can grow larger than the price of one 3D head. This article explains technology and is not investment advice. See our Disclaimer.

Field Note

When reviewing a switch from a 2D line-scan to a 3D profiler, I measure the ΔGV of the problem defect groups under the current lighting on 100 or more images and look at the histogram. If the lower tail of the distribution is above 9 GV, that defect group stays on 2D, and only the groups whose tail reaches below 9 GV get their step and planar size measured again and move up as 3D candidates. Which way a defect near the boundary goes on the real line is something I confirm with specimens.

Field Checkpoints

Once the histogram has sorted out the defect groups that stay on 2D, check the questions below before choosing a 3D head.

• Has it been confirmed, on the drawing and by measurement, that the WD of the 3D head does not interfere with fixtures, handlers or existing brackets? (Shrinking the triangulation angle to gain WD lowers height resolution along with it.)
• Is the planar floor, the xy resolution of the proposed 3D head times three, smaller than the smallest defect spec under control?
• Has the ΔGV of the problem defect groups been measured on 100 or more images, and is the lower tail of the distribution above 9 GV?
• At triangulation angle θ, does the occlusion zone h × tan θ behind the tallest component stay clear of the inspection area? (1.5 mm for a 1.5 mm tall component at 45°)
• Have the peak width and height standard deviation of the line profile been measured on translucent resin or mirror-plated specimens?
• If the head allows a laser wavelength change, have both wavelengths been compared on the same specimen?

Frequently Asked Questions

The 3D head resolves height to about 1 μm, so why does it miss small defects?

Height is read far finer than a pixel through sub-pixel peak detection, but the planar direction stays at the pixel spacing. At 20 μm/px xy the planar floor is 60 μm, and a defect narrower than that gets squeezed into two pixels before its height can even be measured properly.

Where does the 9 GV ΔGV rule come from?

It assumes a sensor noise standard deviation of 1.5 GV in an 8-bit image, takes 3σ, or 4.5 GV, as the floor that separates a defect from noise, and adds a factor of two as margin. If you measure the noise of your own camera and σ comes out different, the rule changes with it.

Glossary

• [1] Laser triangulation laser triangulation
  A method that projects a laser line onto the target and calculates height from shifts in the line position read by a camera at an oblique angle. ↩
• [2] Sub-pixel sub-pixel
  A technique that estimates the peak position more finely than one pixel from the shape of the brightness profile. How finely it can split depends on noise and line shape. ↩
• [3] Line-scan camera line-scan camera
  A camera with a single-line sensor that builds an image by stacking lines while the target passes. ↩
• [4] Telecentric lens telecentric lens
  A lens whose object-side chief rays run parallel to the optical axis, so the target barely changes size when it moves slightly forward or back. Each lens has a fixed magnification. ↩
• [5] Object-side pixel resolution object-side pixel resolution
  The length one pixel covers on the target, written in μm/px. ↩
• [6] Triangulation angle triangulation angle
  The angle between the laser line and the camera line of sight. A larger angle improves height resolution and lengthens occlusion zones. ↩
• [7] Occlusion occlusion
  A zone behind a tall object that the camera or laser cannot reach, so its height cannot be read. ↩
• [8] ΔGV gray value difference
  The difference in gray value between defect and background. In an 8-bit image it is a difference between values from 0 to 255. ↩

References

• SNS Insider, Machine Vision Camera Market Size 2026-2035 press release, GlobeNewswire, 2026-08-17 (market size, 3D camera compound annual growth, share of inspection applications) Read the release
• Formulas: planar floor = 3 × object-side pixel resolution; view = pixel count × resolution; magnification = pixel pitch ÷ resolution; height resolution δz = (xy resolution × sub-pixel fraction) ÷ sin θ; occlusion = h × tan θ; contrast floor = 3σ
• Assumed values: 1/20 px sub-pixel peak detection, 1.5 GV sensor noise standard deviation, per-defect ΔGV (12, 5, 5) and a 1.5 mm component height are assumptions for the worked example

Related reading — When 3D+AI Inspection Answers the Shape Defects 2D Can’t Catch · How 3D Structured Light Measures a Solder Lead Lift That Brightness Cannot Show · In 2D/3D Fusion Inspection, the Bottleneck Is Lighting Interference