Unified 2D 3D inspection: one edge platform, one workflow

Most inspection lines grow sideways. A smart camera checks the print, a laser profiler checks the height, and a bench PC runs both tools while the PLC glues two verdicts together in a script nobody wants to own. A unified 2D/3D inspection platform collapses that stack: both checks run in one environment on hardware at the line, and one pass/fail per part leaves the station. This page gives you the two numbers that decide whether your checks belong together — pixels across the smallest surface defect, and Z resolution against the height tolerance — a worked example that sizes a station for a 40 mm cap, and the cases where two separate boxes are still the right answer.

What a unified 2D/3D inspection platform is

A unified platform executes 2D and 3D checks in one software environment that runs on the inspection hardware itself — the smart camera, the 3D sensor, or a vision controller beside them. One recipe per part carries both checks. One result leaves the station, carrying pass, fail and which measurement failed.

The two modalities answer different questions. A 2D system judges appearance: contrast, color, edges, print and surface defects. A 3D system judges geometry: height, coplanarity, warpage and volume, most often by a laser line swept across the part or structured light projected over the full field. If a check does not clearly need depth data, settle that first — choosing between a 2D and a 3D camera for conveyor pick and place works through the decision.

The conventional alternative is an assembly, not a design: one tool per device, a PC hosting both, and glue logic in the PLC combining two verdicts. Every added layer brings its own license, start-up screen and failure mode.

What changes when both checks share one environment

• One trigger cycle. Both acquisitions run under one handshake, and the station issues a single verdict instead of two that someone must combine.
• The merge lives in software. The result carries which check failed, which is what maintenance needs at handover, not a hunt through ladder logic.
• Edge execution. Images are judged where they are captured, so nothing crosses the plant network before the decision and the station keeps deciding when that network is busy. The latency and bandwidth arguments are the same ones behind edge inference camera integration for factory automation.
• One recipe per part. Changeover touches one screen, not two tools and a script.

The hard part is not the concept; it is getting a 2D camera and a profiler into one tool. Generic acquisition software only half solves that, and the realistic paths are laid out in camera acquisition software compatibility: integration options. Whether the tools inside the pipeline are rules-based or AI matters less than you would expect — both run in the same environment, and the trade-offs are covered in AI vs rule-based smart camera inspection: how to choose.

The 2D number: pixels across the smallest defect

Start from the smallest defect the check must catch, not from the camera you happen to own. The sampling rule of thumb in machine vision is 3 pixels across that defect: at 1 px a feature can fall between sensor pixels, at 2 px blur and contrast eat the margin, and 3 px is the smallest that segments reliably.

Work it in that order: field of view equals part width plus positioning spread on both sides; field of view divided by 3 gives the largest acceptable pixel size; field of view divided by that pixel size gives the minimum horizontal pixel count.

The calculator opens the camera finder filtered to sensors with enough pixels for your field of view, so the arithmetic ends on a filtered shortlist rather than a wish.

The 3D number: Z resolution against tolerance

Height checks are metrology, so borrow the 10:1 gauge rule from measurement practice: the total tolerance band divided by ten is the largest Z repeatability the check can live with.

A ±0.05 mm coplanarity spec is a 0.10 mm band, so the shortlist starts at sensors printing 0.01 mm or better at your working distance. Commit to that figure at the stand-off you will actually mount, not to a best-case number quoted at a different one.

Hardware splits by coverage. A laser profiler measures one profile at a time and pairs it with a mono intensity image — right when the part travels under the sensor or a single edge carries the check. A structured-light or projected-light camera captures the whole field in one shot — right for a stationary part under a dome. Set the mount geometry with the 3D camera working distance and positioning guide; profiler numbers only hold at the calibrated stand-off. Once the height map exists, the check reduces to measuring between surfaces — measuring surface height difference from a 3D point cloud shows the method.

Worked example: a 40 mm cap with 0.10 mm print defects and a ±0.05 mm coplanarity check

Inputs for the example station, a typical electronics case — substitute your own:

• Part: 40 mm wide cap, presented flat.
• Positioning spread: 5 mm in any direction.
• 2D check: catch print defects down to 0.10 mm.
• 3D check: coplanarity tolerance ±0.05 mm.

Arithmetic, with the two rules above:

• Field of view: 40 mm + (2 × 5 mm) = 50 mm.
• Maximum pixel size: 0.10 mm ÷ 3 px = 0.0333 mm/px.
• Minimum horizontal pixels: 50 mm ÷ 0.0333 mm/px ≈ 1,500 px.
• 1.6 MP sensor at 4:3: width = √(1,600,000 × 4 ÷ 3) ≈ 1,461 px → 50 ÷ 1,461 = 0.0342 mm/px → 0.10 ÷ 0.0342 ≈ 2.9 px on the defect → below 3, fails.
• 5 MP sensor at 4:3: width = √(5,000,000 × 4 ÷ 3) ≈ 2,582 px → 50 ÷ 2,582 = 0.0194 mm/px → 0.10 ÷ 0.0194 ≈ 5.2 px on the defect → passes with 1.7× the minimum.
• Z side: tolerance band = 2 × 0.05 mm = 0.10 mm → 0.10 ÷ 10 = 0.01 mm maximum repeatability.

The 4:3 assumption converts megapixels into an array width; a 16:9 sensor changes the width, not the method.

Decision: specify a 5 MP class sensor for the surface check. The 1.6 MP unit only works if the fixture holds the spread to 1 mm per side, which shrinks the field of view to 42 mm: 42 ÷ 1,461 = 0.0287 mm/px → 0.10 ÷ 0.0287 ≈ 3.5 px, a pass with little room. For the height check, shortlist profilers that print 0.01 mm repeatability or better at the set stand-off. Both devices share one controller and one recipe, and the station issues one verdict per part.

The 5 MP branch is the sensor below — 5 MP color resolution, 1/1.45" sensor format, global shutter, Fast Ethernet, M12 mount:

The 1.6 MP branch is the sensor below — 1.6 MP color resolution with an integrated 12 mm lens, manual focus, 1/2.9" sensor format, global shutter:

Which box runs which check

For the height row, a laser profile sensor such as the one below returns a calibrated height profile plus a mono intensity image from every scan:

When the surface check needs glare-free light anyway — shiny or curved caps — the fastest route to one station is a combined system: the model below pairs a 12 MP area camera with a 300 mm white dome light and checks appearance and surface shape in one unit:

The two-devices-one-verdict row needs a box that takes both devices and drives the lighting: the controller below offers PoE camera ports, 8 built-in light channels, 32 GB of memory and an Intel Core i5 processor:

How one cycle runs

Trigger in, 2D exposure out of a global shutter, profiler scan, merge, one discrete or Ethernet verdict to the PLC — all inside the takt. GigE Vision and GenICam keep the camera side open: acquisition and parameters are described by the standards regardless of which environment merges the results. The controller's built-in light channels fire the lights in the same cycle, which keeps the 2D image consistent from part to part; what drifts, and how to pin it, is in ambient light variation in pick and place vision: control it.

Two failure modes live here. If acquisition outruns processing, frames queue and the station drifts behind the conveyor; the fixes are in acquisition faster than processing: fix the frame rate gap. If you keep one process per device instead of one merged pipeline, contain the failure the way multiple vision software instances: one camera per process describes — a crash then costs one check, not the station.

When two boxes are still the right answer

• Checks at different stations. Print inspection upstream, coplanarity after final assembly: no single stand-off serves both, so keep two devices and unify only the reporting.
• Full-wrap geometry. A rotary indexer or several profilers around the part keeps the hardware distributed; merge the results in one software environment instead.
• Color-critical 2D. A profiler's intensity image comes from laser light and is mono, so the color camera stays; unify at the controller.
• Takt below the sum of both acquisitions. Unifying software does not speed physics; add a second station rather than stretch the cycle.
• A 2D model too heavy for the box. When the surface check grows a network that will not fit at the edge, split execution deliberately — plan edge vision inference deployment across camera and host covers how.

Checks before you order

• Smallest defect ÷ 3. Gives the largest acceptable pixel size; divide the field of view by it for the minimum horizontal pixels. Why: below 3 px, blur and threshold noise swallow the defect.
• Tolerance band ÷ 10. Gives the largest Z repeatability the height check tolerates, read at your stand-off. Why: the 10:1 gauge rule keeps measurement uncertainty out of the verdict.
• Port count. Camera plus profiler plus PLC link plus plant network, against the controller's LAN ports. Why: running out of ports mid-project forces a switch.
• Light channels. One per light, plus strobe timing if the line moves. Why: built-in channels only help if they match the lights you picked.
• Shutter type. Global shutter whenever the part moves during exposure. Why: a rolling shutter shears moving parts.
• Lens mount headroom. Most smart-camera models take M12; larger formats move to C-mount, where the lens choice widens. Why: the mount decides the lens range before the pixels do.

Build the station

Run your own two numbers first — pixels from the smallest defect, Z repeatability from the tolerance band — then pick the boxes. The smart cameras and 3D sensors ranges cover the sensing side, the camera finder filters both, and the vision controllers range carries the merge. If your part looks nothing like the worked cap, describe it through the contact page and we will run the numbers with you. Station-level reading beyond this piece lives on the systems topic page.