Inline dimensional measurement of brass parts with machine vision

A machined brass fitting about 35 mm tall needs its dimensions checked to ±0.2 mm inline, inside a 2 s cycle, and the measurement is being added to a machine that already does appearance inspection with image processing. That combination points to one answer: a second 2D camera in the same station, behind a low-magnification telecentric lens, looking at a backlight. ±0.2 mm on a 35 mm part is a 0.6% relative tolerance — comfortable territory for telecentric vision — and 2 s leaves ample time to stop the part, strobe and grab. This guide sizes the camera and lens with the arithmetic shown, picks the lighting a specular metal actually needs, and states plainly where a laser profiler is the better tool.

The decision up front

Use a 2D area-scan camera with a low-magnification telecentric lens and a backlight. One stopped frame measures every outer dimension of the fitting — height, body width, hex across flats — with margin counted in microns, and it drops into the image-processing software the station already runs. A 1" 20 MP sensor behind a 0.166x telecentric puts 14.5 µm of part on each pixel, so one tolerance spans almost 14 pixels and sub-pixel edges land an order of magnitude inside the band.

Choose a laser line profiler when the dimension is a height you cannot silhouette: a step on the end face, warpage across a surface, or a feature the light path cannot reach. Cameras measure edges; profilers measure where the surface is. That difference settles the choice, and 3D vs 2D machine vision: when to choose 3D walks the boundary in detail.

The numbers that set the design

Three numbers fix everything: the tolerance (0.2 mm), the largest part dimension (35 mm) and the pixel pitch of the sensor. The rule that connects them: one tolerance must span at least 10 object-space pixels, and you design in the 10-20 zone. The basis is simple. A sub-pixel edge detector holds a clean backlit edge to about 1/5 of a pixel, a dimension needs two edges, and a gauge you can trust should consume no more than a tenth of the tolerance it polices.

Work it left to right. Choose the field of view to cover the part plus 10-20% for fixturing and sway, divide the sensor width by it for magnification, then divide the pixel pitch by that magnification. The full derivation is in telecentric lens field of view calculation and how camera field of view changes with working distance. Stop buying resolution once the tolerance spans 10-20 pixels; after that, money belongs in the light, the mount and the calibration. Solve the same arithmetic for camera pixels instead:

Two lens-side numbers matter as much as pixels. The lens must resolve the sensor's Nyquist limit — 1000 ÷ (2 × 2.4 µm) = 208 lp/mm — or edges soften before the detector ever sees them; see lens resolution vs pixel size: matching MTF to your sensor and sensor resolution in lp/mm. And the lens must be telecentric, so magnification does not drift with part height — entocentric vs telecentric lenses: when to use each covers that trade.

Worked example: the 35 mm fitting

Inputs: part 35 mm tall and 25 mm across the widest feature; tolerance ±0.2 mm; cycle 2 s. Camera: a 1" 20 MP class sensor — 5472 × 3648 pixels of 2.4 µm, which is 13.1 × 8.8 mm of silicon. Lens: 0.166x telecentric.

• Field of view = 13.1 mm ÷ 0.166 = 78.9 mm wide, and 8.8 mm ÷ 0.166 = 53.0 mm high. The fitting fills two-thirds of the frame height and leaves room for a fixture.
• Object-space pixel = 2.4 µm ÷ 0.166 = 14.5 µm.
• Pixels per tolerance = 200 µm ÷ 14.5 µm = 13.8 — mid-zone.
• Two-edge noise at 1/5-pixel edges: 2 × (14.5 ÷ 5) = 5.8 µm ≈ 0.006 mm, about 3% of the tolerance band.
• Whole-pixel edges, with no sub-pixel fitting at all, cost 2 × 14.5 = 29 µm = 0.029 mm — under a seventh of the band. From here the budget belongs to optics and calibration, not pixels.

Cycle time: a 20 MP mono frame is about 20 MB, and USB3 moves it in roughly 60 ms. Add a strobe of 0.15-0.5 ms and a few tens of milliseconds of edge fitting, and the measurement closes near 0.1 s — 1.9 s of the cycle remains for stop, settle and release. The 2 s budget is not tight; the same logic squeezed into faster lines is worked in measuring points on a 5 mm part in under one second.

For the full fitting in one frame, the low-magnification pick from the telecentric lenses range is:

The LS-TCL-1422 is a 0.166x bi-telecentric C-mount lens for 1" sensors at a 258 mm working distance, with a 16-22 mm image circle. Working distance is a design point on a telecentric, not a tuning knob — order the class you need rather than shimming one into place. Match it with a mono body from the area scan cameras finder:

The MV-CS200-10UM is a 20 MP USB3.0 camera on a 1" IMX183 sensor, in mono — what silhouette edges want. The two IMX183 bodies differ in interface and color:

All 2 share: Sensor type CMOS, rolling shutter; Sensor model Sony® IMX183; Pixel size 2.4 μm × 2.4 μm; Sensor size 1"; Resolution 5472 × 3648; SNR 41.5 dB; Gain 0 dB to 24 dB; Binning Supports 1 × 1, 2 × 2, 4 × 4. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Take the mono USB3 body for a measurement-only head beside the controller; take the GigE color body when one camera must also feed appearance grading, since GigE spans longer cable runs. Both are rolling-shutter sensors, which is fine with the part stopped and the backlight on continuously, and wrong for measuring a moving line — for that, filter the camera finder for global shutter.

A smaller sensor also makes the grade

The classic 5 MP 2/3" format — 2448 × 2048 pixels at 3.45 µm, or 8.5 × 7.1 mm of silicon — stays inside the 10-pixel rule. The match in the telecentric range is LS-TCL-1390, the 2/3" C-mount object-side telecentric at 0.179x with a 260 mm working distance. That magnification frames 8.5 ÷ 0.179 = 47.5 mm by 7.1 ÷ 0.179 = 39.7 mm, so the 35 mm part fits with about 2 mm to spare above and below. Object-space pixel: 3.45 µm ÷ 0.179 = 19.3 µm, and 200 µm ÷ 19.3 µm = 10.4 pixels per tolerance — on the floor of the zone, but on it. The trade is fixture margin, so part presentation must be disciplined.

Lighting: silhouette the brass

Polished, oiled brass is a mirror. Front light either glares into the lens or vanishes, and edge contrast ends up a property of the day's finish. For gauging, turn the part into a silhouette: backlight on the far side, camera on the near side, and the edge becomes the shadow boundary — finish, tarnish and coolant film drop out of the measurement. The collimated backlight or diffuse backlight for edge gauging comparison favors collimated light at this accuracy, sized at least to the field width. Round shiny sections reward the same discipline; see telecentric illumination for measuring round shiny parts accurately. Keep domes and coaxial front light for the appearance station — how brass behaves under them is shown in lighting solder joints on brass: dome plus coaxial fill.

For dimensions on the end face — a turned step, a counterbore seen as a diameter — send the light through the lens instead. The coaxial route in one body:

The LS-TCL-2194 is an object-side telecentric C-mount lens for 1" sensors, 0.5x at a 110 mm working distance, with built-in coaxial light. That magnification frames 13.1 ÷ 0.5 = 26 mm by 8.8 ÷ 0.5 = 18 mm — an end face, not the whole fitting, which is the point of a second station.

If the part must keep moving, hold blur under one pixel: at 100 mm/s, a 150 µs strobe smears 100 mm/s × 0.00015 s = 15 µm ≈ 1 pixel at 14.5 µm per pixel, while a 0.5 ms pulse smears 50 µm ≈ 3.4 pixels. Stop the part, as the 2 s cycle allows, and the question disappears. Drive the strobe from a controller triggered by the station PLC — start at the lighting step of the configurator.

Vision or laser: where each wins

The competing sensor is a laser displacement gauge or 2D laser profiler. It wins when height is the dimension — steps, warpage, clearance under an overhang — or when the far side of the line has no room for a backlight. It loses on integration: a separate controller, a separate program and a measurement engine the existing vision station cannot host, all spent on a tolerance the camera already covers thirty times over.

The decision, case by case: outer dimensions on a stopped part — camera and backlight. End-face heights read in silhouette — camera and coaxial telecentric. Heights across a whole face, or hidden geometry — profiler. No light path behind the part — profiler.

Checks before you order

• Clear aperture. A telecentric front element must exceed the field diagonal: √(78.9² + 53.0²) = 95 mm for the worked example. Plan the frame opening, the lens mass and the bracket around it.
• Standoff on both sides. The lens working distance in front; behind the part, room for a backlight wider than the field.
• Calibration. Put a certified gauge or dot grid at the part plane and calibrate pixels-to-mm in the finished fixture, never from nominal magnification. Distortion is the budget item: 0.1% residual at 30 mm off-axis is 30 µm, more than the edge noise; lens distortion correction for accurate pixel-to-mm measurements shows the fix.
• Focus at the real working distance, with the production light fitted — the procedure in focus the lens at working distance applies to any body.
• Depth range. A fitting that settles ±2 mm high must stay inside the telecentric depth of field; the validation method is in telecentric depth of field: validate measurement accuracy.
• Swarf and coolant. Brass chips on the backlight glass cast false edges; give the window an air purge and a wipe interval.
• Which edge. Burrs and chamfers move an edge by more than the tolerance; agree with the process engineer whether the gauge sees the theoretical sharp corner or the burr root. For features stacked in height, read concentricity measurement of features at different heights.

When the usual answer is wrong

• The part arrives in any orientation. A silhouette still yields across-flats and overall height, but hex orientation must come from fitting, not assumption, and one-sided features need a fixture or a second camera at 90°.
• The tolerance tightens to ±0.02 mm. The architecture holds; the budget does not. Magnification rises toward 0.5x, the field falls to about 26 mm, you buy a second view, control the mount temperature and run a formal gauge study. The methods that take cutting tools to micrometre precision are in subpixel edge measurement of cutting tools.
• Threads and knurls. An edge detector on a thread measures crest noise, not a diameter. Define the derived dimension — shadow envelope or pitch-diameter proxy — before promising a number.
• Sizing from scratch rather than adding to a station. Work the same chain from part to lens to camera as in machine vision dimensional measurement for small parts and telecentric lens selection: match magnification and image circle; the numbers above are the fitting-sized instance of it.

The 60-second checklist

• Tolerance spans at least 10 object-space pixels: 200 µm ÷ (pitch ÷ m).
• Field of view covers the part plus 10-20% on the constraining axis.
• Telecentric lens, so part height cannot change the scale.
• Backlight at least as wide as the field, collimated, strobed, PLC-triggered.
• Mono body for gauging; global shutter if the line runs; color only if one camera grades appearance too.
• Calibration gauge ordered before the camera bracket is cut.
• Laser profiler only for heights and hidden geometry — interlocked and labelled.

When the drawing is on the table, build the station in the configurator — camera, lens, light and trigger in one pass — or send the part drawing and tolerance list through the contact page and we will size it with you. More station-level reasoning is collected on the systems topic page.