Part alignment in machine vision: what it actually buys you

Alignment is the least glamorous stage of an inspection and one of the most consequential, because it decides how much you have to spend on mechanics. A vision system that requires the part in exactly the same place every time is really a specification for a fixture. One that finds the part first can inspect whatever the conveyor delivers.

That trade is the whole subject. Alignment costs cycle time and buys you tolerance to a part that moved. Here is how to price both sides of it.

What the alignment stage does

You give the system a template image of a good part and mark the region that identifies it. On every capture after that, the aligner locates that pattern and works out how the part has translated and rotated relative to the template. Everything downstream then runs in the part's coordinate system rather than the camera's.

The consequence is that your inspection regions travel with the part. A region drawn around a latch stays on the latch when the part arrives 8 mm to the left and rotated 6 degrees. Without alignment that same region would be looking at empty background, and the inspection would report a missing latch on a perfectly good part.

This is why alignment failures and inspection failures are worth counting separately. A part the aligner could not find has not been inspected at all, and a station that quietly passes those is not inspecting a fraction of your production.

Edges or learned features

Classical pattern matching works from edges. It builds a geometric model of the template's edge structure and searches for the arrangement that fits best. It is fast, predictable, and decades proven, and it fails in one specific way: when there are not enough clean edges to work with.

A learned aligner takes the other approach. It learns the visual features of the part from the template rather than relying only on edge geometry, which is what lets it hold onto a part whose edges are unreliable. Three cases where that matters:

One practical note if you do switch: the two engines respond to their settings differently, so sensitivity and confidence need re-tuning rather than carrying across. Budget a few minutes for that instead of assuming a like-for-like swap.

The speed modes are a cycle-time decision

Alignment happens before any inspection runs, so whatever it costs comes straight off your cycle-time budget. On the OV80i the aligner offers three modes with published timings, and seeing them as numbers rather than adjectives makes the choice obvious:

The span between Fast and Accurate is about 200 ms. On a line at one part per second that is a fifth of your budget, and it is the single largest lever on inspection latency that has nothing to do with the AI models. Worth deciding deliberately rather than leaving on whatever it defaulted to.

Rotation range, and the limit worth knowing

The rotation range setting bounds how far the aligner will look. A part beyond that angle relative to the template is simply not detected, which is a design decision rather than a bug: a narrower search is faster and produces fewer false matches.

Set it from the mechanics rather than optimistically. If the fixture holds the part within a few degrees, a tight range is faster and safer. If parts arrive loose on a belt, you need the range to cover reality or you will be investigating alignment failures that are really a range set too narrow.

The learned engine handles roughly plus or minus 25 degrees. That is a real ceiling and it is better to plan around it than to discover it. Two ways past it when parts can arrive at any orientation: mechanically constrain the presentation so it falls inside the window, or capture more than one template so each covers a slice of the rotation. Neither is exotic, and both are far cheaper to design in than to retrofit.

Scale invariance, and why measurement turns it off

Scale invariance lets the aligner match a part that appears slightly larger or smaller than the template. Useful when part height varies or the camera distance is not perfectly repeatable, because the same part at a different distance images at a different size.

It is also the one setting that interacts with another block, and the interaction is worth understanding rather than fighting. If you are measuring, scale invariance has to be off, and the product enforces this. The reason is straightforward: a measurement converts pixels into millimetres using a fixed scale. An aligner allowed to absorb size changes silently rescales the image, which means the very variation your measurement is trying to detect gets normalised away before the measurement sees it. You would get a beautifully consistent number that no longer reflects the part.

So the rule is simple. Detecting defects on parts whose apparent size wanders: scale invariance on. Reporting a dimension against a tolerance: off, and control the working distance mechanically instead. Our guide to telecentric lenses covers the optical way of removing that same variation.

When to skip alignment entirely

Alignment can be switched off, and on a genuinely well fixtured station you should. If the part is located by a nest or a clamp and arrives in the same place to within a pixel or two, the aligner is spending 70 to 270 ms per capture confirming something the mechanics already guaranteed.

The honest way to decide is to measure rather than assume. Capture a few hundred parts as they normally arrive and look at how much the position actually varies. Two outcomes, both useful: the variation is small and you can drop alignment and reclaim the cycle time, or it is larger than anyone expected, which is worth knowing before you tune inspection regions around an assumption that does not hold.

Setting it up, in order

1. Capture a template on a good part presented the way production actually presents it, not the way it looks when someone places it carefully by hand.
2. Mark the template region on stable features. Larger regions with real detail align better than small ones. Exclude anything that legitimately varies between good parts, otherwise you are asking the aligner to match on the thing that changes.
3. Set the rotation range from the mechanics, not from optimism. Measure how much parts actually rotate.
4. Start on the balanced mode, then move only if cycle time or match rate gives you a reason.
5. Decide scale invariance from the job: on for detection with varying apparent size, off if anything is being measured.
6. Re-run the setup against captures you already have. A batch realign over existing images tells you the match rate before the station sees a live part, which is the same discipline as backtesting the inspection itself.
7. Decide what a failed alignment does. It is not a pass and it is not the same as a defect. Route it somewhere a human sees it.

That last point is the one most often left to chance. An alignment failure means the part was never inspected, so treating it as a pass puts uninspected product into your good stream. Counting alignment failures separately, and giving the line a signal when the rate climbs, turns a silent gap into an early warning that a fixture is drifting or the lighting has moved. For where this fits in the wider station, see our guide to machine vision systems.