Taking Machine Vision from Samples to Line Acceptance

Sample validation proves direction, not production readiness. Defect boundaries, imaging stability, cycle-time budget, review workflow, retention and model change control still need explicit acceptance criteria.

Build a sample library from real operating conditions

Version good parts, defects and boundary samples by defect type, minimum size and position. Include material batches, ambient light, oil, dust, vibration and part-pose variation. Every new sample should retain its source, annotation rule and validation result so acceptance criteria do not drift between teams.

Accept cycle time and false calls as system behavior

Total inspection time includes triggering, exposure, image transfer, inference, PLC writeback, rejection and record retention. Acceptance should measure false negatives, false positives and indeterminate cases together, with a manual review path for boundary results.

• Imaging repeatability: fix lighting, optics, exposure and mounting, then document acceptable drift.

• Line cycle time: validate continuous operation at peak production speed, not only individual images.

• Traceability: retain the image, decision, equipment state, model version and review outcome.

• Degraded operation: define stop, pass-through or manual fallback behavior for camera, network and inference failures.

Keep model change under operational control

New materials, suppliers, maintenance and process changes can shift image distributions. Delivery should define retest triggers, model and parameter versions, staged release, rollback and monitoring so improvement is controlled rather than ad hoc tuning.

A production acceptance baseline connects algorithm metrics with line actions and operational ownership. Reproducible issues, traceable decisions and reversible versions are what turn a visual demonstration into a durable production capability.