Counting connected regions in an ROI to verify electrode pairs

Parts that carry paired fine features — electrode fingers, contact prongs, sensor combs — fail in discrete ways: two features fused into one, a finger chipped away, a gap drifted off pitch. A grey-level score blurs those states together; a count of connected pixel regions inside each ROI separates them. This guide sets the inspection up end to end: silhouette lighting, pixel scale, a threshold that holds, an area window that catches what the count misses, and the verdict table for every count. It closes with a worked example sized around a 100 µm gold electrode pair inspected 768 times per card.

The pass/fail rule: count first, area second

Threshold the ROI to a binary image, then count the connected foreground regions. The number is the first half of the verdict:

The count alone cannot clear a part. Two regions with a 60 µm gap measure differently from two regions at 100 µm, and a chipped finger can still leave two regions behind. Gate every count of two behind two measurements: the distance between the two regions (the gap) and each region's area (the window). Pass only when the count is exactly two, the gap is inside its band and both areas are inside theirs. Everything else in this guide exists to make those three numbers trustworthy.

Light the part as a silhouette

Put the light behind the part. The PET film passes the backlight; the vapour-deposited gold blocks it; the camera sees dark fingers on a bright, even field. Contrast then comes from opacity, not reflection, so it survives viewing angle, surface texture and the glare that ruins front-lit metal. That stability is what a count needs: grey-level wobble moves a threshold, and a moving threshold makes the count flicker. The broader surface or silhouette lighting decision covers when backlighting applies; for dark features that must be found on dark surfaces, where no silhouette is possible, dark contaminant detection on dark surfaces is the closest case.

Size the backlight to the part plus fixture with about 10% margin per side. The case card measures 127 × 88.9 mm, so anything above roughly 140 × 98 mm of lit area covers it. A diffuse flat panel is the forgiving choice for presence counting; if you also need to gauge edges, a collimated backlight trades alignment sensitivity for sharper edges.

For small cards in red — a good match for mono sensors and for eyes during set-up — the LS-BKL-1468 lights a 150 × 150 mm area.

If the fixture bolts through the light, keep the bolts out of the ROIs by choosing a panel with a centre opening. The LS-BKL-1397 lights a 300 × 200 mm area at 54 W with a centre hole.

The same families scale by lit area and power:

Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Set the pixel scale before the threshold

Work from the narrowest gap that must still pass, not from the nominal gap. Three pixels across a gap is the floor for a threshold to split two regions at all; five to eight pixels keeps the count steady against focus blur and threshold drift. That sets a hard cap on pixel size:

From a required field of view, check the sensor pixels it takes:

Count in mono — a colour sensor spends half its photons on colour you will discard. For a PC-based cell, the MV-CS200-10UM is a 20 MP mono USB3.0 camera with a 1" rolling-shutter sensor on a C-mount. The LS-CMR-0166 runs a 4000 × 3000 sensor at 30 fps over USB3. A rolling shutter is fine here: the part holds still while you grab the frame.

If you want the verdict issued on the device, smart cameras run thresholding, region counting and area windows in-camera. The MV-SC5060XM-08M-WBN is a 6 MP mono GigE smart camera with an 8 mm M12 lens. A fixed lens may not reach the magnification a fine pitch needs, so compute the pixel scale it delivers before committing; when it falls short, a C-mount area scan camera plus a PC accepts any macro lens you need.

Worked example: a 100 µm gold electrode pair

Inputs, with where each comes from:

• Pair pitch — the gap between the two fingers of one pair: 100 µm (the application's design).
• Measurement repeatability of the imaging chain: ±5% of reading, so ±5 µm at a 100 µm gap (the published figure for this class of set-up).
• Sensor: 1624 × 1236 px (the camera in the case set-up).
• Process: two fingers per cavity, 384 cavities per card, card 127 × 88.9 mm, 10 cards per hour (the case process).

Three rules drive the arithmetic: hold measurement uncertainty to a quarter of the tolerance (the standard 4:1 rule), span the narrowest passing gap with at least 5 pixels, and bracket each pair with one full gap of margin on each side of the ROI.

Step 1 — gap band. Repeatability: 100 µm × 0.05 = 5 µm. At 4:1, the guardable tolerance is 4 × 5 µm = ±20 µm, so the gap passes only between 80 and 120 µm.

Step 2 — pixel scale. The narrowest passing gap is 80 µm, so the cap is 80 µm ÷ 5 px = 16 µm/px. Choose 5 µm/px — a 3.2× margin under the cap, easy to reach with standard macro optics.

Step 3 — ROI size. One gap of margin each side around a pair spanning about one pitch: 3 × 100 µm = 300 µm square. At 5 µm/px that is 300 ÷ 5 = 60 px per side.

Step 4 — coverage. At 5 µm/px the sensor sees 1624 × 5 = 8,120 µm ≈ 8.1 mm by 1236 × 5 = 6,180 µm ≈ 6.2 mm per view — dozens of cavities on the 127 mm-wide card, each with its 60 px ROI inside.

Step 5 — area units. One pixel is 5 × 5 = 25 µm². As an illustrative measurement, a threshold returning 1,856 px for one finger gives 1,856 × 25 µm² = 46,400 µm² ≈ 0.046 mm². Store the window in µm² so a lens change cannot silently invalidate the limits.

Step 6 — load. 384 × 2 = 768 ROIs per card; at 10 cards per hour that is 7,680 checks per hour, so 3,600 s ÷ 7,680 ≈ 0.47 s per check. The vision step is never the bottleneck; part handling is.

Decision. Pass a cavity only when the count is exactly two, the gap lands between 80 and 120 µm, and both finger areas sit inside the ±3σ window set from known-good cards. Count of one — reject as fused. Three or more — reject as chipped. Count of zero, or a count that flickers — stop and fix the imaging, not the parts.

Thresholding that survives production

Lock everything that sets grey level. Drive the backlight from a regulated supply, fix the exposure and gain, and turn every auto mode off. Aim the histogram so the bright field sits above roughly 200 grey levels and the gold below roughly 60 at 8 bit, then threshold at the midpoint near 130; that split leaves the widest room for drift before the binary image changes.

Set the threshold from a known-good master image, not from each frame. An adaptive threshold re-balances itself against whatever sits in the ROI — on a chipped finger it re-balances exactly wrong, and can erase the defect you are hunting.

Filter with a minimum area only. A minimum area of a few pixels drops dust specks out of the count; a maximum-area filter is dangerous here, because it silently deletes a fused pair instead of reporting it. Let the count and the area window do the diagnosing.

Count before you clean up, then clean up for measurement. A dust particle on the backlight inside a finger leaves a hole, and a hole does not split a connected region — the count is safe. The hole does corrupt the area, so fill holes before the area measurement, not before the count. When you also measure positions, watch the coordinate conventions in blob centroid coordinates and, for sub-pixel centres, the inverted grayscale centroid method.

Anchor the ROI against fixture play

A count is only as good as the ROI's placement. Anchor each ROI on a stable printed feature — the contact pad, not the fragile fingers — using edge-based pattern matching, which holds position while grey levels wander. Then measure the fixture: image one known-good card twenty times, re-clamping between grabs, and take three standard deviations of the pair-centre positions as your minimum ROI margin. The worked example's one-gap margin already tolerates ±100 µm of placement error per side; keep your measured 3σ spread well inside it. When several cameras or image sources feed one inspection, binding each ROI to its image source prevents the classic fault of judging one part with another part's ROIs.

Set the area window from known-good parts

Measure the limits; do not guess them. Run at least 20 known-good cards through the final lighting, optics and exposure, measure every finger's area, and set the window to the mean ±3σ of those runs. Calibrate the pixel scale first with a backlit calibration plate and store the window in µm², so the limits survive a lens or working-distance change.

Then attack the window on purpose. A card with one deliberately bridged pair must return a count of one; a card with a deliberately chipped finger must return three or more regions or an area outside the window. If a defect class slips through both gates, tighten whichever gate let it past — and re-run the 20 good cards before you ship the new limits.

For dimensional numbers beyond pass/fail — actual gap distributions, capability indices — the same silhouette set-up feeds camera resolution choices for dimensional inspection.

When the count rule is the wrong tool

The count works because the nominal state is known and simple: two separate features. It misleads when:

• The design legitimately connects the features — tie bars, serpentine traces, keyed bridges. The nominal count is not two, so the criterion must come from area or shape instead.
• The features stand above the film plane in relief. A silhouette merges them with the substrate; switch to a surface geometry from machine vision illumination types.
• Many identical features share one view and drift together. Counting each ROI wastes effort; lattice matching across the array finds a missing or extra cell directly.
• You need graded measurements and trends rather than a boolean — then treat the task as measuring, not counting.

Checks before you commit the set-up

• Lit area covers part and fixture with about 10% margin per side. The case card needs at least 140 × 98 mm of lit area. The LS-BKL-0750 lights a 180 × 100 mm area with room left for the fixture.
• Mono sensor; colour only if a second wavelength is doing other work in the same cell.
• Rolling shutter is fine while the part holds still; insist on global shutter only if frames are grabbed on the move.
• Trigger and discrete I/O for the robot or indexer handshake, so frames land only when the part is settled.
• Threshold, ROI geometry, pixel scale and area window stored per product, with the pixel scale recorded next to the limits.
• Golden samples on the shelf: one bridged card and one chipped card, run at every set-up change.

Where to go next

Build the silhouette first: the backlights category, the light finder and the light configurator narrow panels by size, colour and type. For the camera side, the area scan cameras category and its camera finder filter by resolution, shutter and interface, with the camera configurator as the structured start; on-device counting options live in smart cameras & vision sensors. A structured trial before ordering is laid out in machine vision lens and lighting selection: a test plan, and questions about a specific pitch go through the contact page.