The hardest vision problems are often not the model: optics, lighting, line speed, product variation and what happens after a detection all matter.
Start with the cost of a wrong decision
A vision inspection project does not begin with a camera resolution or an AI accuracy figure. It begins with the defect definition and the consequences of false accepts, false rejects and unnecessary line stops. A missed safety-critical defect and a cosmetic over-rejection have different tolerances, escalation paths and business owners. Write these down before issuing a vendor request.
Treat image formation as part of the system
Lens, illumination, exposure, sensor placement, motion blur and mechanical repeatability define what the algorithm can see. Build the image-formation envelope for the real production line: day and night shifts, material batches, surface finishes, cleaning regimes, vibration and operator handling. An algorithm cannot reliably recover features never captured by the optical system.
Demand representative acceptance data
Collect examples of good parts and every consequential defect class. Split trials by production condition or time where possible, not just random frames from one run. Agree the confusion matrix, per-class detection rates, false-reject cost, throughput and out-of-distribution escalation before tuning. Beware impressive single-number performance claims without a traceable denominator or a documented test set.
Define who is allowed to change the inspection
Every proposed threshold, recipe or model update should have a test plan and a responsible approver. Record which version was active for a production batch, how exceptions were handled and how the result reached PLC, MES or quarantine workflow. A detection should not become an unreviewed quality disposition merely because an AI system produced it.
Prototype the failure mode, not just the happy path
Ask vendors to show lighting drift, rejected images, network interruptions, camera misalignment, false alarms, model rollback and manual override. Confirm whether the plant can restore a known-good configuration without vendor intervention. The procurement decision should weigh uptime and maintainability alongside detection capability.
Make the next vendor conversation more productive.
Use the free project fit check to generate an initial list of questions for your team.
Open planning tool ↗Primary sources and limitations
These links support the public product or research context. They do not substantiate comparative performance conclusions. Source content may change; verify directly with the supplier.