More Than One Imaging Method on a Line: How Structured-Light 3D, X-ray, Infrared and Electron Microscopy Share One Decision Logic

Running more than one imaging method on a line is no longer the exception: visible light for appearance, structured-light 3D for height and coplanarity, X-ray for solder joints hidden under packages, infrared for thermal distribution, electron microscopy for sub-micron morphology. Historically each additional modality meant another standalone system, another rule set and another acceptance standard. This article is about the alternative: decoupling the decision logic from the imaging method.

ACI in this article stands for Automated Cognitive Inspection — the next-generation term for AI AOI: keep the A and the I, and replace the O (optical) with C (cognitive), because the decision basis is cognition rather than the imaging method. The full argument is set out in From AI AOI to ACI.

1. Does a new imaging method require a new system?

In most plants the answer has long been yes. One system for visible-light AOI, another for X-ray, another for 3D — each with its own programming model, its own defect library and its own alarm thresholds. Operators learn three interfaces, process engineering maintains three standards, and a changeover means retuning all three. This is not because unification is technically impossible. It is because those systems' criteria grew out of their imaging methods: they compare pixels, and pixels from different modalities are not comparable to begin with.

Once the criterion compares pixels, the imaging method becomes the system boundary: change the illumination or the modality and the reference image is void, so the whole rule set has to be rebuilt. That boundary is not a law of physics. It comes from how the model was built.

2. What the word “optical” blocks

The O in AOI stands for optical. Semantically it writes the other modalities out of the category — calling a station that runs structured-light 3D alongside X-ray an “automated optical inspection” station is simply inaccurate. The practical consequence is worse: procurement and process engineering read the word literally when drafting standards, so the X-ray portion gets its own separate procedure. The name draws the boundary before the solution is even read.

That is the technical reason for renaming AOI as ACI: what is actually shared is cognition, not imaging. Further reading: From AI AOI to ACI.

3. Which layer is actually shared

Split the inspection chain apart and the sharing happens in the middle layer, not at either end:

• The imaging layer is not shared: structured light, X-ray, infrared and electron microscopy each have their own physics and calibration, and forcing them together would be wrong.
• The feature layer is partly shared: different modalities yield different feature dimensions, but the question “what does normal look like for this component” is posed identically.
• The decision layer is shared: once the criterion reads “is this component within its normal range of appearance” rather than “how many pixels does this image differ from the reference”, the decision logic becomes independent of the imaging method.

This is also why one world-model foundation can serve both defect inspection on a line and quantitative analysis of electron-microscope images: the imaging differs by orders of magnitude, but the question being answered is of the same kind.

What is shared is not the camera. It is the judgement of what counts as normal.

4. Where each modality remains irreplaceable

A shared decision logic does not make the modalities interchangeable. Each carries a responsibility the others cannot take over:

• Structured-light 3D answers geometric questions — height, coplanarity, volume. Two-dimensional imaging cannot obtain these in principle, and inferring them from greyscale is unreliable.
• X-ray answers what is under the package and inside the joint. Once a BGA or QFN is placed, external imaging can no longer see its solder balls.
• Infrared answers questions of temperature and energy distribution, typically for locating anomalies under power.
• Electron microscopy answers sub-micron morphology and composition, pushing inspection beyond the reach of conventional industrial cameras.

The selection question is therefore not “which imaging method” but “which classes of question must this line answer”. Answer that and the modalities fall into place — while the decision logic need not fragment along with them.

5. A workable sequence

Unifying modalities need not happen in one step. A common order is: first move the main station (usually visible light) from pixel comparison to component cognition and observe what changes in changeover effort and false calls; then bring the second modality into the same decision framework, which requires adding features for it rather than rebuilding rules; only at the end merge the acceptance standards and inspection SOPs into one document.

The advantage of this order is that every step can be validated and rolled back on its own. Merging three systems into one in a single move concentrates all the risk at the moment of cutover while the benefit only arrives at the end — engineering-wise, the least favourable arrangement.

6. Why false-call definitions drift between modalities

When three systems coexist, the painful part is usually not escapes but the fact that false calls are not counted the same way. The visible-light station records “a bright spot caused by reflection” as a false call; the X-ray station records “void ratio within tolerance”; the 3D station records “height difference within tolerance”. All three are called false calls, yet they refer to different things. At the monthly review the three numbers can neither be added nor compared.

The root cause is again how the criterion is written. When it reads “pixel difference from the reference exceeds a threshold”, a false call can only be defined as “the difference exceeded but nothing was actually wrong” — and what counts as “nothing wrong” is buried in each station's own threshold. When the criterion reads “is this component within its normal range of appearance”, false calls gain a single statement: the model judged it outside the normal range and review judged it inside. That statement is independent of the imaging method, so all three stations' false calls belong in one table.

The value is not a tidier report. It is that improvement gains a direction: aggregate false calls by component type and you can see which component's normal appearance is under-described; aggregate by station and all you learn is which machine is noisy.

7. How annotation is unified

A shared decision layer does not remove the need to annotate. Feature dimensions differ by modality and must still be covered per modality. What can be unified is how annotation is organised:

• Same object: annotation attaches to the component instance, not to image coordinates. The same component is a region in a 2D image, a height distribution in a point cloud and a set of solder balls in an X-ray — all pointing at one object.
• Same vocabulary: defect names and severity use one shared list. Each system's private “NG-3” or “defect type 07” cannot be translated across; with one vocabulary, review opinions from different modalities can corroborate each other.
• Same boundary: what counts as acceptable normal variation is defined once by process engineering and shared by all three modalities, instead of being rewritten inside each threshold.

With those three in place, the cost of adding a modality drops from “build another system” to “add one set of feature annotations”. That is what unification actually saves in engineering terms.

8. The three layers, side by side

The preceding sections, collected into one table for self-assessment:

• Imaging —— No;Physics and calibration differ;New equipment and calibration flow
• Feature —— Partly;Dimensions vary by modality; the question does not;Add one set of feature annotations
• Decision —— Yes;The criterion hangs on normal component appearance;No rule rebuild
• Review & statistics —— Yes;One false-call definition adds up across stations;No separate definition for the new modality

9. When not to unify modalities

There are clear cases where this does not apply, and saying so is more useful than recommending.

• The line runs one imaging method and will not add another soon: the benefit is the rebuild cost saved when a modality is added, so with no addition there is no benefit, while rewriting criteria still costs time.
• The two modalities sit in different plants, shifts or acceptance ownership: unifying the decision layer means unifying definitions, and definitions are usually harder to unify organisationally than technically.
• The existing systems run stably and changeover is rare: the benefit of a restated criterion is realised mainly during changeover and false-call work.
• There is no process-side definition of normal variation: without an owner for that, unification merely merges three vague standards into one vague standard.

If any of the four holds, continuing with the current approach is the right call. How the criterion is stated can be migrated in stages; there is no reason it must happen all at once.

Full solution for this scenario: the full inspection solution for AI AOI Software

This article was generated by AI. Customer cases are simulated scenarios based on real product capabilities and figures are illustrative; see product pages for official benchmarks.