Conversations about AI in quality inspection usually start with questions about the model: which architecture, how much data, what accuracy. Yet the most common reason a pilot underperforms is not the model — it is the image the model had to work with. A model does not see the product. It sees what the lighting showed it.
A model cannot fix an image it cannot read
It's a simple relationship that is easy to forget: if a defect is not visible in the image, no model will detect it, no matter how well it was trained. The reverse is just as true — if lighting produces a false shadow or reflection that resembles a defect, the model can learn to react to an imaging artefact instead of a real product problem. In both cases the source of the error is not training; it is what happened in the frame before any pixel ever reached the model.
That is why process qualification at AIVQC starts with assessing the camera, optics and lighting — before training data is collected, not alongside it. A dataset built on unstable imaging conditions has to be collected again once those conditions are fixed, regardless of how good the model turns out to be.
This applies to every kind of check we describe under applications — missing components, incorrect assembly, labels and print, or contamination — because every one of them starts with the same question: can this even be seen in the image, before anyone talks about the model.
The same defect, a different result depending on the angle of light
Cracks and scratches are the clearest example of this: a fine surface defect is sometimes visible only from a specific lighting angle, and invisible entirely from another. The same applies to product edges and geometry, which are inherently harder to image evenly than a flat surface — flash and moulding deformation often only become apparent once lighting emphasises the contour rather than the surface itself.
The practical consequence: the imaging conditions used to collect training data have to be reproducible at the production station. An image taken under different conditions than the ones inspection will actually run under teaches the model something other than what it will face on the line.
A glossy surface is a different problem than a matte one
Surface type changes the entire lighting problem. Raw, ground, anodised or plated metal reflects light strongly — a reflection can be hard to tell apart from a real scratch without properly qualified optics. A glossy label, film or laminated pack creates the same problem in a different industry: glare can completely obscure the print it was meant to verify. Plastics can be matte, glossy or textured depending on colour and raw-material batch, which means lighting has to be matched not just to the product, but to its specific variant.
There is no single universal lighting setup that works everywhere. There is, however, one question worth asking separately for every case: is this specific defect, on this specific surface, visible in a repeatable way — cycle after cycle, shift after shift.
What to check before training starts
- Whether the defect is visible under the same conditions it will actually be imaged under on production — not only in test or lab conditions
- Whether lighting limits glare and reflections on glossy or laminated surfaces instead of just working around them
- Whether imaging conditions stay stable over time — ambient light, dirty optics and an ageing light source can quietly change what the model "sees" months later
- Whether product edges and geometry are lit to emphasise the contour, not just the surface
- Whether the same lighting setup works across every colour and texture variant meant to run on one station
That's why it's the pilot's first step, not its last
In an AIVQC pilot, assessing camera, optics and lighting comes before any work on the model — precisely because a mistake made at this stage costs the most if it only surfaces after data has been collected and training has run. If the current imaging conditions are not sufficient, the qualification report states that directly, instead of leaving it as a surprise later.
If you are unsure whether the lighting on your line is good enough to detect a specific defect, that is exactly the question the initial consultation resolves — before any decision is made about collecting data or training a model.