What is computer vision?

Computer vision is software that interprets images: deciding whether a part is within tolerance, reading a label, counting items, or spotting that something on a line is not where it should be. The model turns pixels into a decision a system can act on.

The useful mental model is a very fast, very literal inspector who never gets bored and never improvises. It will apply exactly the judgement it was trained on, to every frame, forever, which is its strength on repetitive inspection and its weakness the moment conditions change.

What can computer vision actually see?

Reliably: presence and absence, position and orientation, dimensions against a reference, print and label reading (OCR), surface defects with a consistent appearance, and counting. Less reliably: anything requiring judgement a person would call "it looks off". If two experienced inspectors on your line would disagree about a part, a model trained on their labels will inherit that disagreement rather than resolve it. Label consistency, not model choice, is usually the ceiling on accuracy.

Why lighting decides the project, not the model

The single biggest predictor of whether a vision system works is the rig around it: lighting, camera angle, and how consistently the part is presented. A defect visible under directional light can disappear under diffuse light. A shiny surface will blow out highlights that hide exactly the scratch you are looking for. Vendors sell models because models are what they have to sell, but on most industrial projects the model is close to a commodity and the optics and fixturing are where the engineering time goes. If a supplier quotes on the model and says little about lighting, that is a signal.

The accuracy number nobody converts into parts per shift

"99% accurate" sounds finished. On a line running 10,000 units a shift it means 100 wrong calls per shift. Whether that is excellent or unusable depends entirely on what a wrong call costs. Worth deciding before anyone builds anything: is a defective part reaching a customer worse than a good part being scrapped? The two errors trade off against each other, and tuning for one costs you the other. A system tuned to catch every defect will reject good parts; one tuned to pass all good parts will let defects through. There is no setting that avoids the choice.

When a sensor beats a camera

Vision is the expensive answer. If a dimension can be measured with a gauge, a distance with a laser sensor, a weight with a load cell, or presence with a proximity switch, use that instead: it will be less expensive, faster, and it will not need retraining when the lighting changes. Reserve computer vision for judgements that are genuinely visual: surface appearance, complex shapes, reading text, or checking many features at once. Our audit exists partly to say when a €2,000 sensor does the job a €50,000 vision system was quoted for.

What does a computer vision project cost?

At Crux Digits the path is the same as any other build: a €2,500 audit to confirm the inspection is feasible and worth automating, a €20,000 proof of concept that runs on your real parts and your real lighting for 4–6 weeks, and production from €50,000. The proof of concept matters more here than in most AI work, because feasibility depends on physical conditions no one can assess from a specification, only from your parts, on your line.