Data quality vs data quantity: why clean images beat big datasets.

Why does data quality beat data quantity for inspection AI?

For a good-parts-only inspection model, the consistency of a small training set matters more than its raw count. The model builds a picture of normal from the reference images, so what it learns is only as clean as what you show it: 20 tightly controlled good images describe normal more precisely than thousands of images shot under drifting light and framing.

This runs against the instinct that more data is always better. That instinct comes from supervised learning on huge, varied datasets, where volume averages out noise. An anomaly model that learns the normal distribution from a small set works the opposite way: every reference image directly shapes the boundary of normal, so a bad image is not diluted by volume, it moves the boundary.

The practical consequence is that curation beats collection. Time spent making sure each reference image is sharp, consistently lit and genuinely good does more for accuracy than time spent gathering more images. Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators, and it trains from about 20 good reference images, which is exactly why each of those images has to earn its place.

What makes a training image clean?

A clean training image is one where the only thing that varies from image to image is the part itself, not the conditions around it. The lighting, the framing, the focus and the fixturing should be as constant as they will be at run time, so the model learns the part rather than the noise of an inconsistent capture.

Three conditions carry most of the weight. Lighting should be the same intensity, colour and angle in every frame, which is why the unit uses configurable lighting (diffuse, directional or coaxial, at 24V) tuned once and held constant. Framing should keep the part in the same position and scale, so the same feature lands in the same pixels. Focus should be sharp, and a 12 MP global-shutter camera captures a moving part without the motion blur a rolling shutter would add.

The reason this is non-negotiable for a small set is that the model cannot tell intended variation from accidental variation on its own. If the lighting changes between references, the model may learn that "normal" includes that lighting swing, which loosens the boundary and lets real defects slip inside it. Consistency at capture is what keeps the learned normal tight.

What counts as representative variation versus accidental noise?

Representative variation is the real spread of a good part that the model must accept; accidental noise is everything about the capture that should have stayed constant but did not. A good training set contains the first and excludes the second, and telling them apart is the core judgement in curating the set.

Representative variation is the range a good part genuinely has: different colour lots, tolerated surface finishes, the small cosmetic differences quality already accepts. If the model never sees that range, it will reject good parts at the edges of it, so the set has to cover the real good-part population, not one perfect exemplar. This is where a slightly larger set earns its keep, when the extra images add real variation rather than repeats.

Accidental noise is the opposite: a shadow from a door left open, a part sitting crooked in the fixture, a smudge on the lens, one image shot at a different exposure. None of it belongs to the part, but the model cannot know that, so it folds the noise into normal. The goal is a set that is broad on real variation and narrow on everything else.

How do a few bad images poison a good-parts-only model?

A few bad images poison a good-parts-only model by stretching its definition of normal until real defects fit inside it. Because the model learns the boundary of normal from the reference set, a defective part mistakenly included as good teaches it that the defect is acceptable, and a badly lit or blurred image teaches it that the blur is part of normal too.

The damage is not proportional to the number of bad images, which is what surprises people. On a set of 20, a single contaminated image is 5% of everything the model knows about good, so it carries real weight. Adding more clean images later does not automatically undo the harm, because the bad image is still in the set moving the boundary. The fix is to remove it, not to outvote it.

That is why a good-parts-only workflow puts the review step on the input, not the output. It is cheaper to reject one questionable reference image before training than to chase the false rejects or escapes it causes afterward. For how drift can quietly degrade even a clean model over time, and the retraining loop that corrects it, see the companion post on model drift and retraining.

What does a training-set curation checklist look like?

A training-set curation checklist is a short set of pass/fail questions you apply to every reference image before it goes into the set. The point is to make the review deliberate rather than assumed, because the model will treat whatever you give it as ground truth for normal.

Work through each candidate image and keep only the ones that clear every line:

1. 01Is this part genuinely good, confirmed by quality, not just assumed?
2. 02Is the lighting the same intensity, colour and angle as every other reference and as run time?
3. 03Is the part in the same position and scale, fully in frame, not cropped or rotated?
4. 04Is the image sharp, with no motion blur, using the global-shutter capture?
5. 05Does the lens and part surface look clean, with no stray smudge, shadow or reflection that is not always there?
6. 06Across the whole set, is the real good-part variation covered (lots, tints, tolerated finishes) without adding accidental noise?

Run the same checklist again whenever you recapture to retrain, because a set that was clean last quarter can pick up new noise from a moved fixture or a changed light. Clean input is not a one-time task; it is the standard you hold every time the model learns.