Connector metrology
Every pin located, and the debris, from one inference
Measurement and defect detection are usually two stations. On a pin field they come out of the same pass.
One trained recipe, one telecentric capture, 72 pins. The network returns a sub-pixel centroid for every pin and a pixel level contamination map at the same time, and the part orientation is recovered in 3D from the same set of points.
The raw capture, and the same frame with the model output drawn on it. A region of interest on every pin, a marker on every detected tip feature. Both rows of the report come from this one inference.
What true position means on a pin field
How far a pin sits from where the drawing puts it, reported as a diameter rather than a plus and minus. The tolerance zone is a circle around the nominal centre, so the reported value is twice the radial deviation.
RTP = 2 × √( Δx² + Δy² )
A pin 0.10 mm out in x and 0.10 mm out in y has a radial error of 0.141 mm and a true position of 0.283 mm. Quoting the radial figure halves every number on the report, which is the most common way these results get misread.
From frame to centroid
Four stages, repeated on every pin
The same pipeline runs on all 72 pins in every capture, and the same network produces the contamination output alongside it. The measurement is not bolted onto the defect check, and the defect check is not bolted onto the measurement.
Raw capture
One frame of the whole connector through a telecentric lens. No per pin imaging, no indexing stage, no second station.
Locate each region of interest
The model finds every pin cell in the frame and places its own search region. Nothing is windowed by hand, so a part with more pins costs a larger frame rather than a longer recipe.
Segment the tips at pixel level
Inside each region the tip features are segmented rather than matched. A mask tolerates the plating variation, the seating variation and the lighting variation that break a template.
Reduce to a sub-pixel centroid
Each pin reports one centroid, the mean of its two detected tip features. That single point is what every downstream measurement uses: true position, neighbour pitch, and the pose fit.
Debris in the same pass
Contamination arrives already attributed to a pin
Foreign object debris turns up as metal swarf, plating flakes, glove fibre, dust and whatever the previous operation shed. A classifier trained on the list you have already seen fails on the first thing you have not, so detection is posed as a departure from a clean pin cell rather than as a set of named classes. Each region comes back with a pixel accurate bounding box and the pin it is sitting on.
Accuracy
The model is pixel accurate. The optics decide the microns.
The centroid lands within one pixel in the worst case, and a centroid taken from a segmentation mask is inherently a sub-pixel quantity, so that bound is conservative. What it is worth in microns is then a property of the sensor and the telecentric lens rather than of the network. That makes the accuracy upgrade path a sensor swap: the lens, the mount, the illumination and the trained recipe all stay as they are.
| Sensor | Resolution | Pixel pitch | Scale on part | Field of view | Worst case |
|---|---|---|---|---|---|
| 8.3 MP | 3864 x 2192 | 2.0 µm | 9.35 µm / px | 36.1 x 20.5 mm | ±10 µm |
| 65 MP | 9344 x 7000 | 3.2 µm | 3.33 µm / px | 31.1 x 23.3 mm | ±3.3 µm |
Roughly 2.8 times the linear resolution over the same part means each pixel covers about three times less of it. This is also the reason to name a tolerance before anyone specifies hardware: the sensor that holds it falls out of arithmetic, and paying for the larger one when the drawing does not need it is the most common way these stations get overbuilt.
3D pose recovery
Getting the part plane out of a 2D image
A telecentric lens projects orthographically, with parallel rays and no perspective foreshortening. That is what makes tilt recoverable rather than merely annoying. A tilted pin grid images full size along the tilt axis and compressed by the cosine of the tilt angle perpendicular to it, so the compression itself carries the orientation.
An affine fit, decomposed
Fitting a two by two affine to the 72 measured pin positions and decomposing it by SVD returns yaw, tilt magnitude, tilt axis direction and the split into pitch and roll in closed form. No iteration, no second camera, no structured light.
Immune to the bent pin
The fit runs as a consensus over random 30 pin subsets, so one displaced pin cannot drag the recovered plane off the agreement of the other 71. A naive fit to all of them absorbs the outlier into the datum and reports the part as good.
Tilt stops being error
Applying the inverse affine normalises the grid back to a flat reference frame. Every downstream number is reported against that frame, so how the part happened to sit no longer shows up as a position deviation.
Reference frames
Three ways to report it, one set of pins
The underlying measurement does not change. Only the frame it is expressed in changes, so the report can match the drawing instead of the drawing having to match the station.
Absolute
Each pin reported as an x and y in the camera frame. Simple, and the right answer when the part is held rigidly, but it moves whenever the part moves in the field of view.
Relative to neighbours
Pin to pin pitch and adjacency distances. This is the frame that catches a missing pin, a doubled pin or a local pitch error, and it does not care where the part sits.
Relative to a datum
The corners of the connector body define a part local frame and pin positions are reported as offsets from it. This is how drawings normally specify true position, so it is usually the one that matters.
Robustness
Nine parts, one recipe, no re-pointing
The nine captures span the variation a line actually produces: differences in seating, in illumination, and in how the part is rotated in plane. All of them ran through the identical recipe.
Every capture returned all 72 pins with two tip features each. Nothing was re-pointed, re-focused or re-trained between parts, which is the difference between a demonstration and something that survives a shift.
The hard parts
Four things that decide whether the number is real
Measurement fails differently from detection. A detection either finds the defect or does not. A measurement can be confidently wrong, and a confidently wrong pin position is a good part scrapped or a shorted one shipped.
Telecentric optics are doing real work here
They are not a nicety on this part. Orthographic projection is what removes perspective error from the measurement and, separately, what makes the tilt recoverable in closed form. A standard lens would put the part height into the position numbers and leave no clean way to take it back out.
The scale has to be traceable
Microns per pixel is the constant that converts everything, and it is the only thing standing between a pixel accurate centroid and a wrong micron figure. Derive it from a certified artefact, record it, and re-check it on a schedule. A scale that has drifted produces plausible numbers rather than obviously wrong ones, which is worse.
The datum choice changes every number
A fit to all pins spreads error evenly and flatters a part with one bad pin. A fit to the datum features named on the drawing reports what the drawing asks for. The two disagree, and only one of them is the specification.
Repeatability before accuracy
A station that is right on average but scatters part to part cannot disposition anything. Repeatability is a fair thing to require before the station rejects its first part, not after it rejects the wrong one.
How to specify it
- Name the position tolerance from the drawing, and say whether it is diametric. Half the disagreements about whether a system is capable turn out to be a factor of two.
- Name the frame the drawing uses. Absolute, neighbour referenced and datum referenced reporting all come from the same centroids, but they produce different numbers on the same part.
- Name the smallest particle that matters, then let the field of view, the lens and the microns per pixel follow. Our optics calculators do that arithmetic for standard and telecentric paths.
- Say how much the part can tilt and rotate in the fixture, because that is what decides whether pose recovery is a convenience or the thing holding the measurement together.
- Agree the acceptance study up front. Repeatability on real parts and a correlation against your reference method, with the thresholds written down before anyone runs it.
Manufacturers running Overview AI in production
Manufacturers running Overview AI in production
FAQ
Frequently asked questions
Can one model measure pin position and find debris at the same time?
Yes. The same network produces two outputs in a single inference pass: a sub-pixel centroid for every pin, and a pixel level contamination map. On a 72 pin connector that is 72 positions and every piece of debris on the part from one capture. Each contamination region is attributed to the pin it sits on and reported with a pixel accurate bounding box, so the defect arrives already tied to a location rather than as a loose reject flag.
What is true position on a connector?
True position is how far a pin actually sits from where the drawing says it should, expressed as a diameter rather than a plus and minus. The tolerance zone is a circle around the nominal centre, so the reported value is twice the radial deviation. A pin 0.10 mm out in x and 0.10 mm out in y has a radial error of 0.141 mm and a true position of 0.283 mm. Reporting the radial figure instead of the diametric one halves every number on the report, which is the single most common way these results get misread.
How accurate is the pin centroid?
The centroid lands within one pixel in the worst case and is typically tighter, because a centroid computed from a segmentation mask is inherently a sub-pixel quantity. That makes the physical accuracy a function of the optics rather than of the model. At 9.35 microns per pixel the worst case is plus or minus 10 microns. Putting a higher resolution sensor behind the same telecentric lens and the same trained recipe takes the scale to 3.33 microns per pixel and the worst case to plus or minus 3.3 microns. Nothing else in the station changes.
How is part tilt handled without a second camera?
A telecentric lens projects orthographically, with parallel rays and no perspective foreshortening, which is what makes the tilt recoverable from a single 2D image. When the connector face is tilted, the projected pin grid is full size along the tilt axis and compressed by the cosine of the tilt angle perpendicular to it. Fitting a two by two affine to all 72 measured pin positions and decomposing it by SVD returns yaw, tilt magnitude, tilt axis direction and the split into pitch and roll in closed form. The inverse affine then normalises the grid back to a flat reference frame, so tilt stops appearing as measurement error.
Does one bent pin corrupt the alignment?
No, because the pose fit runs as a consensus over random subsets of the pin field rather than as a single fit to all of it. With 72 pins available, repeatedly fitting 30 pin subsets and keeping the consensus means a single displaced pin cannot pull the recovered plane off the agreement of the other 71. That matters because the outlier is usually the thing you are trying to measure, and a naive best fit quietly absorbs it into the datum and reports the part as good.
Which reference frame should positions be reported in?
Whichever one the drawing uses, and the same set of centroids supports all three. Absolute position reports each pin in the camera frame, which is simple but moves when the part moves in the field of view. Neighbour referenced reporting gives pin to pin pitch and adjacency, which catches a missing pin, a doubled pin or a local pitch error and does not care where the part sits. Datum referenced reporting builds a part local frame from the corners of the connector body and reports offsets from it, which is how drawings normally specify true position.
What size of debris can be detected?
It follows from the optics rather than from the model. The scale in microns per pixel sets the floor, and a feature needs several pixels across it to be separable from sensor noise and from the texture of the plating underneath. Contrast matters as much as size, since a dark fibre lying across a pin cell is far easier than a metal fragment on a metal surface. The honest way to specify it is to name the smallest particle that matters, then work backwards to the field of view and lens that put enough pixels on it.
Related
Debris and FOD Detection
Finding foreign object debris inside connector housings and cages before it becomes a field short.
Read more →Dimensional Measurement
Distances, diameters, angles and gaps against a drawing tolerance, and what makes the number trustworthy.
Read more →Bent Pin Inspection
Deformed and displaced pins on high density connectors, and where position measurement takes over.
Read more →Bring us a drawing
Send the position tolerance, the pin count and the smallest particle you need to catch. You get the optics and an honest answer on whether one station can hold both, before anyone talks to you about hardware.