Measuring points on a 5 mm part in under one second

You have a 5 mm diameter part, specific points to measure on it, and less than a second for each answer. That is a sizing exercise before it is a brand question: every current platform, smart camera or PC-based, locates points on a 5 mm part inside one second when the pixel budget, the optics and the lighting are right, and most fail quietly when those three are wrong. This checklist works the numbers in that order: field of view to pixels, lens to f-number, light to geometry, then the millisecond budget. You finish with a camera resolution, a lens magnification and working distance, a lighting geometry, and a cycle that fits, ready to pick from the telecentric lenses, coaxial lights and smart cameras ranges.

The decision first

Commit to this set for micrometre-level points on a 5 mm part in under 1 s:

• A monochrome 5 MP camera of the 2448 x 2048 pixel, 2/3" class: one frame holds the whole part with margin.
• A 1x telecentric lens, C-mount, working distance set by the cell: near 110 mm on a compact bench, 200 mm or more when fixturing needs room.
• Backlight for edge and silhouette points; coaxial light for points on a flat shiny face; a dome for curved shiny surfaces.
• Measurement on the device or over a GigE/USB3 link, so image transfer never spends the second.

That set lands the whole cycle near 100-200 ms, five times inside budget. The sections below show where each number comes from.

Step 1: Size the field of view and the pixel budget

Two numbers decide everything downstream: the field of view and the object pixel size.

Field of view: the whole part plus margin, in one frame. Margin rule: 20 percent of the part, at least 1 mm on a 5 mm part, so fixture play cannot push a point out of the image. One frame for all points also removes stage moves from the cycle, which is half of how you stay under a second.

Pixel budget: divide the field of view by the pixel count.

The 2/3" 5 MP class, 2448 x 2048 pixels at 3.45 µm, covers 8.45 x 7.07 mm at 1x and 3.45 µm per pixel. Subpixel interpolation places the black-to-white transition inside a single pixel (subpixel edge detection), so an edge position repeats to about 0.35 µm. Two more rules: the smallest measured feature spans at least 3 pixels, and the tolerance is at least four times the repeatability.

Decision thresholds, committed:

• ±50 µm or looser: any current 1.6-5 MP sensor with its standard lens; lighting dominates the outcome.
• ±2-10 µm: 5 MP with a 1x telecentric, 3.45 µm per pixel, 0.35 µm repeatability. The default for this job.
• ±1 µm: 1.3-1.5x magnification, 2.3-2.5 µm per pixel, and a nest that holds the part within ±0.1 mm.

Run your own field of view and pixel target through the calculator:

The magnification arithmetic from the sensor side is worked in calculate object space pixel size from magnification; the wider measurement workflow, from calibration to tolerance stacking, is in machine vision dimensional measurement for small parts.

Step 2: Choose the magnification and the f-number

Telecentric first, for a computable reason. With a standard lens the magnification changes with part height, so a point away from the optical axis shifts sideways:

A part sitting 0.3 mm high on a 108 mm working distance moves a point 2.5 mm off axis by 0.3 x 2.5 / 108 = 6.9 µm, twenty times the 0.35 µm repeatability. Off-axis points move more than central ones, so every distance between points is bent as well. A telecentric lens makes the chief rays parallel and the term goes to zero.

When you need a telecentric lens for gauging covers the borderline cases; at ±2 µm on a fixtured part you are not borderline.

Start at 1x. On the 2/3" 5 MP class that is 8.45 x 7.07 mm of view, so a 5.00 mm part sits inside with a millimetre to spare on the short side. At 1.36x the object pixel drops to 3.45/1.36 = 2.54 µm, but the short side shrinks to 7.07/1.36 = 5.20 mm, leaving 0.10 mm total on a 5.00 mm part. That is a fixture precision, not a margin. Pass 1x only when the tolerance needs the finer pixel and the nest is that good.

Three telecentrics that bracket the choice:

All 3 share: Mount C-mount. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The LS-TCL-1972 is a 1x object-side telecentric for the 2/3" class, C-mount, 440 mm working distance. The LS-TCL-2118 delivers the same 1x at a 108 mm working distance for compact cells. The LS-TCL-0811 is the fine-pixel option, 1.36x with a 73 mm working distance and a coaxial light port.

On the iris, work between f/5.6 and f/8. The diffraction disc is 2.44 x wavelength x f-number, so at 0.55 µm and f/8 it is 2.44 x 0.55 x 8 = 10.7 µm, about three pixels, still steep enough for tenth-pixel interpolation. At f/11 it reaches 14.7 µm and the edge softens for nothing. Depth of field at 1x and f/8 in this lens class runs past a millimetre, so a few tenths of nest variation sit inside it; the bench procedure for proving that is telecentric depth of field: validate measurement accuracy.

Worked example: three points on a 5 mm pin, one-second cycle

Inputs, each with its source: a 5.00 mm diameter steel pin with three scribed points, each to be located to ±2 µm, answer within 1.0 s, which is the job as quoted at the bench. Fixture: a machined V-nest; assume ±0.3 mm of part-to-part height variation and a part that is static during exposure. Camera class from Step 1: 2/3", 2448 x 2048 pixels, 3.45 µm.

Rule in words: fit part plus 1 mm into the sensor's short side, then confirm the repeatability sits at least four times inside the tolerance.

1. Required field of view: 5.00 + 1.00 = 6.00 mm.
2. Short side at 1x: 2048 pixels x 3.45 µm = 7.07 mm. 7.07 > 6.00, so 1x passes with 1.07 mm spare.
3. Object pixel at 1x: 3.45 µm. Edge repeatability 3.45/10 = 0.35 µm; a two-edge point-to-point distance repeats to √2 x 0.35 = 0.5 µm. Against ±2 µm that is a 4x margin. Pass.
4. Standard-lens check at 108 mm: 0.3 x 2.5/108 = 6.9 µm of shift, which fails. Even at 440 mm it is 0.3 x 2.5/440 = 1.7 µm, most of the tolerance spent on geometry. Telecentric, no debate.
5. Cycle: trigger and settle 20 ms, exposure 2 ms, readout at 30 fps 33 ms, point tools on the device 100 ms, answer over TCP 5 ms. Total 160 ms against 1000 ms, a 6x margin.

Decision: a 2/3" 5 MP monochrome camera. On the lens, keep the LS-TCL-2118 at its 108 mm working distance, and use a collimated backlight silhouette if the points are edges. If the three points are marks on the flat end face, keep the lens and swap the backlight for the coaxial route in Step 3. Order, then prove the 0.5 µm figure over 30 placements before the cell ships.

Step 3: Light it for two grey levels

Point tools lock onto a transition, so build an edge that runs from black to white inside one or two pixels. Geometry first, colour second, brightness last.

• Edges, diameters, silhouettes: backlight behind the part. Collimated gives the sharpest gauging edge, and collimated backlight or diffuse backlight for edge gauging sets out when each wins.
• Marks, engraving or pads on a flat shiny face: coaxial light, through the lens port when there is one.
• Curved shiny surfaces, ball ends and domed heads: a dome, the route dome lights for shiny, curved and wrinkled parts works through.

Two coaxial units that suit this size of part:

The LS-CXL-0036 is a high precision coaxial light, 74 x 50 x 50 mm on 24 V. The LS-CXL-0027 is a 60 x 60 mm unit at 4.4 W red or 7.2 W white: red on bare steel, white when the mark itself carries colour. When two finishes refuse to separate, run the contrast test in which LED colour gives the best contrast in machine vision rather than guessing.

If the part moves inline, freeze it. Allowed motion is half an object pixel, so the exposure must sit at or below 0.5 x 3.45 µm = 1.73 µm of travel: at 250 mm/s that is 1.73/250 mm per ms = 6.9 µs, so strobe at 5 µs from a controller. A static part on a nest takes 1-2 ms of continuous light, which is gentler on the LED.

Step 4: Budget the second

Write the milliseconds down before choosing a processor:

• Trigger and part settle: 20 ms
• Exposure: 2 ms static, 5 µs strobed
• Readout, 5 MP at 30 fps: 33 ms
• Point tools on the device: 50-150 ms
• Answer over TCP or serial: 5-10 ms

Total 110-215 ms. Speed is not the constraint; image transport is. A 5 MP mono frame is 2448 x 2048 = 5.0 MB. Down a Fast Ethernet link at about 10 MB/s that is 500 ms before a single tool runs, and the second is spent. Over GigE at about 100 MB/s it is 50 ms, and USB3 is several times faster again. Two architectures clear the bar: a smart camera that measures where the pixels are and sends a few bytes, or a GigE/USB3 camera into a PC.

One body that takes the C-mount telecentric and runs the tools itself:

The LS-SMC-0017 is an x86-platform smart camera, 8.8 MP, global shutter, 1" sensor, C-mount. The spare pixels let you crop to the part and read out faster. For the coarser end of the tolerance band, an integrated vision sensor with its own lens:

The MV-SC3050XC-16M-ATN-H is a 5 MP color vision sensor with a global shutter, a 1/1.45" sensor and an M12 lens mount. Its fixed-lens families include no telecentric. That geometry suits tolerances near ±50 µm, not ±2 µm, so for micrometres stay on the C-mount route. Colour costs accuracy too: the Bayer mask softens the very transition you locate, so mono for point measurement.

If the answer lands in a PLC and you must prove elapsed time on the machine, measuring execution time between two points in a PLC scan gives the method.

When the usual answer is wrong

• Buying pixels instead of contrast. A 12 MP camera on a grey, low-contrast edge measures worse than 5 MP on a binary edge, and no calculator rescues bad lighting.
• Cranking magnification. At 1.36x the short side of the view is 5.20 mm, so a 5.00 mm part leaves 0.10 mm total. One burr on the nest and a point clips out of frame.
• Colour sensor for measurement. Demosaicing interpolates away the edge you are locating; mono, every time.
• Rolling shutter on a moving part. The image shears as the part passes; use a global shutter.
• Streaming full frames over Fast Ethernet. 500 ms gone before processing starts.
• Taking the 0.35 µm repeatability on faith. That figure is what binary contrast buys; measure it over 30 placements.

The pre-order checklist

Geometry

• Field of view = part + 20 percent margin, 6 mm for a 5 mm part, whole part in one frame.
• Part height range measured, not estimated: it sets the depth of field needed and settles the telecentric question.

Pixels

• Object pixel chosen so pixel/10 is at most a quarter of the tolerance; smallest measured feature at least 3 pixels.
• Repeatability confirmed on 30 placements, not read off a projection.

Optics

• Telecentric; magnification derived from the field of view, not from preference.
• C-mount on the camera body if a telecentric lens goes on it.
• f/5.6 to f/8; stop down only to cover the measured height range.

Light

• One geometry per point type: backlight for edges, coaxial for flat faces, dome for curves.
• Monochrome sensor; strobe width at or below half an object pixel divided by line speed.

Timing

• On-device tools or a GigE/USB3 link; the budget sums under 250 ms.
• Elapsed time proven on the machine, not in the demo.

Order against the telecentric lens finder, the lighting finder and the camera finder, or start at the lens with the system configurator, because the lens sets the mechanical envelope everything else lives in. The full bench procedure is in machine vision lens and lighting selection: a test plan, and planning views on round parts specifically is covered in small round part inspection: planning views and measurements.