How to detect small transparent parts with machine vision

A die face that has to be recognized in one shot can fail for two separate reasons, and fixing only one of them leaves the cell unreliable. First, the feature you search for must be big enough in the image: for parametric and template matchers, plan on about one tenth of the field of view. Second, a transparent yellow body gives the camera almost no stable contrast under room light, so even a correctly sized pip scores badly. This guide works the numbers for a 16 mm die with a 4 mm pip at 70 mm working distance, picks the lighting geometry that makes translucent plastic behave, and lists the checks to run before you trust the cell. The same method fits any small translucent target: caps, connectors, lenses, small medical parts.

How small can a part be and still be detected?

Plan on the feature you search for spanning at least one tenth of the field of view; a wrist camera that images about 100 by 75 mm at 70 mm working distance therefore has a practical floor near 10 by 7.5 mm for anything you must not miss. Below that size, detection is luck, not engineering.

The rule comes from how matchers score. They normalize against the whole search image, so a feature covering a few percent of it competes with background clutter, shading gradients and lens vignetting; a small change in focus or illumination then pushes the score across the accept threshold. Scaling the image up in software does not change the ratio — the matcher still sees the same fraction, with softer edges.

Absolute pixel count matters too, but it is rarely the limit at these fields of view: 100 mm across 1280 pixels is 0.078 mm per pixel, so even a 4 mm feature holds 51 real pixels. The same conversion for any sensor is in the object space pixel size calculation.

Measure before you model: put a ruler in view and read the width it spans, because every number below hangs off that one measurement. Note the direction of travel, too — if you widened the view to cover a whole table by stitching wrist camera snapshot positions, every small feature got smaller at the same time, and the large field of view lighting balance problem arrived with it.

Is the pip on a transparent die too small at 70 mm?

Yes, it is: a 4 mm pip covers about 51 of 1280 pixels, roughly 4 percent of a 100 mm field of view — two and a half times below the one-tenth rule — while the 16 mm die itself clears the rule at 16 percent. Run the arithmetic with a common wrist-camera format: 1280 x 960 pixels on a 4.8 x 3.6 mm sensor, 3.75 µm per pixel, optics giving about 100 x 75 mm at 70 mm working distance.

• Object pixel size: 100 / 1280 = 0.078 mm per pixel.
• Die, 16 mm across: 16 / 0.078 ≈ 205 pixels, and 205 / 1280 ≈ 16 percent of the width. The die match works — which is what you see.
• Pip, 4 mm across: 4 / 0.078 ≈ 51 pixels, and 51 / 1280 ≈ 4 percent. The pip match fails.

Three ways out, cheapest first:

Crop the search, not the sensor. Match the die in the full frame, then search for the pip inside a 20 mm crop centered on it: 4 / 20 = 20 percent of the search region. The one-tenth rule binds the region the matcher searches, not the whole sensor, so two matcher stages fix this with zero hardware.

Narrow the field. To put a 4 mm pip at 10 percent, the field may span 4 / 0.10 = 40 mm. With the same sensor the magnification is m = 4.8 / 40 = 0.12, and:

The nearest stock lens, 8 mm at the same 70 mm, gives m = 8 / (70 − 8) ≈ 0.13 and a field of 4.8 / 0.13 ≈ 37 mm; the pip lands at 4 / 37 ≈ 11 percent and matches in one pass.

Buy a bigger die. This helps only in proportion: a 22 mm die carries a pip near 5.5 mm, still 5.5 percent of a 100 mm field. Not a fix.

And enlarging the image in software is not on the list: interpolation spreads the same pixels over more samples and dilutes the edge gradient the matcher scores.

Decision: implement the crop search now — it is free — and order the 8 mm lens when you want single-pass matching or smaller features. Where pips must also be measured rather than found, dimensional inspection systems for small parts add the metrology margins on top.

Why does a transparent die fail when an opaque one passes?

Because the contrast you were counting on does not exist: under room light, a translucent yellow body shows pip-to-body differences that swing from about 30 grey levels toward zero as the tilt changes a few degrees, so the match score never settles. An opaque die scatters light from a stable surface; a transparent one does three things instead. It transmits, so the camera reads the table through it. It refracts at the edges, which show as thin bright and dark lines rather than broad gradients. And it reflects specularly off both faces, so glints from the ceiling sweep across the face with every robot pose.

Threshold tuning cannot fix physics. The fix is to own the light field: light from every direction at once, so no single direction can glint — a dome; light at grazing angles, so flat faces go dark and refracting edges glow — dark field; light down the lens axis for flat shiny faces — coaxial. Machine vision illumination types walks the full set of geometries; the next section picks one.

What lighting gives a transparent part stable contrast?

A dome light is the first choice for a glossy translucent die, a low-angle ring light makes the outline glow when you need shape or orientation, and a coaxial unit fits flat shiny faces carrying engraved marks. The table is the decision:

For the die, take the dome: the pips sit on the face, and you want that face evenly lit with no directional glint at any robot pose. Dome lights for shiny, curved and wrinkled parts covers the geometry in depth.

The dome to start from is the LS-DOM-0085: a 166 mm outer diameter and a 116 mm aperture.

Size any dome so its aperture is at least twice the part's largest dimension; 116 mm against a 16 mm die leaves generous margin.

The low-angle ring is the backup and often the better answer for orientation alone. It skims the surface at a grazing angle: flat faces stay dark while edges, where light bends into new directions, scatter brightly — on a transparent die the whole outline glows. The surface or silhouette lighting-angle question decides between the two families.

The LS-RNG-0256 is a 96 mm OD LED ring light with a low 0-15° beam.

Coaxial light shares the lens axis through a beam splitter: flat shiny faces light evenly and engraved marks read with even contrast. It is the wrong first tool for a clear die — the camera sees the background through the body — but the right one for marks on flat glossy faces. The LS-CXL-0027 is a 60 x 60 mm coaxial light rated 4.4 W red and 7.2 W white.

Each family has its own section: dome lights for the die, ring lights for the dark-field option, coaxial lights for flat engraved faces.

How do you set the station up, step by step?

Budget half a day: lock the field of view first, then shroud the station, mount the light, set exposure near 5 ms at f/5.6, and re-teach the matcher only after the image holds steady across every robot pose you plan to use.

1. Measure the field of view with a ruler in view and compute the object pixel size. If the feature sits under one tenth of the width, decide the crop search or the lens swap before touching lighting.
2. Fit the lens: for the worked example, 8 mm at 70 mm working distance gives a 37 mm field and puts the pip at 11 percent.
3. Shroud the station: a hood with black sides removes the ceiling and the neighboring cell's strobes from your image.
4. Mount the dome on the lens barrel with the part centered under the camera hole, and confirm the aperture clears the part with margin.
5. Power the dome from a 24 V controller channel and run it continuous while you align.
6. Set exposure and aperture: start at 5 ms and f/5.6, then adjust until the die body sits mid-grey, about 110 to 140 of 255, with no saturated glint.
7. Teach two stages: the die parametrically in the full frame, then the pip inside the 20 mm crop, with each accept threshold near 60 percent.
8. Freeze gain and exposure, and log the score of every frame for verification.

How do you verify the result before release?

Run the die in 20 poses spread over at least four rotations, log every score, and require the weakest pip score to clear your accept threshold by ten points with the room lights on and again with them off. Judge the tail, not the average: a cell that averages 85 but dips to 55 fails at shift change.

Two checks catch the rest:

• Contrast: histogram the die face — the pip should sit at least 40 grey levels from the body in every rotation.
• Ambient: repeat the run with the lights off; if scores move more than a few points, extend the shroud or pulse the light and lock the exposure to the pulse, the method strobed lighting ambient rejection measures.

If the weak frames cluster just after the arm stops, the matcher is seeing motion, not contrast — measure it with OpenCV frame tracking for camera vibration.

When is the one-tenth rule the wrong rule?

The rule is conservative: with a backlight behind a clear part you can hold a silhouette at 3 to 5 pixels across, and inside a cropped search region the rule binds the crop, not the whole sensor frame. Two escapes are worth knowing. Backlighting turns a transparent part into a binary silhouette — near-total contrast, clean thresholds, sub-pixel edges — the right route when only outline and dimensions matter. And when the pips are molded rather than painted, a dark-field ring lifts them as bright dots on a dark face, which a blob tool holds with a handful of pixels, no matcher needed. Where neither escape applies — full-frame matching of low-contrast surface marks — the only honest fix is fewer millimeters in the field of view.

What should you check before ordering the light?

Check three numbers: the dome's aperture at least twice the part's largest dimension, the light's outer diameter clearing your mount and the robot envelope, and a controller channel matched to the light's 24 V input. Then run the list:

• Aperture: the dome opening clears the part and the lens barrel with room to spare.
• Lens coverage: at the narrower field the lens must still cover the sensor corner to corner — calibration target field of view and lens coverage margins shows the margin to demand.
• Controller: one channel per light, rated above the light's current.
• Mechanics: dome height adds to the station; confirm the robot approach and the camera focus distance still fit.
• Drawer spare: a sheet of polarizing film rescues the occasional glossy part — cross polarization explains the set-up.

Build from the light: the lights finder filters by family and size, the configurator below starts at the lighting step, and the lens step of the configurator covers the lens swap. To adapt the numbers here to your part, send the field of view and working distance through the contact page and we will size the set-up with you. More lighting decisions sit under the vision lighting topic.