Engineer the image before tuning the model
The model cannot recover information the camera never captured. Lens selection, field of view, working distance, lighting geometry, exposure, motion blur and trigger timing determine whether the defect is visible consistently.
A production pilot should therefore document the imaging envelope: acceptable illumination, part position, speed, focus, occlusion and environmental variation. The goal is repeatability, not one perfect frame.
- Camera and lens selection
- Lighting geometry and intensity
- Trigger source and timing
- Working distance and field of view
- Motion and exposure limits
Define the economics of errors
False reject and false accept rates have different operational consequences. A false reject may create rework and operator fatigue; a false accept may ship a defect or create a safety risk.
Acceptance criteria should therefore be written in business terms before training starts. Some classes may require near-zero escapes and human review, while lower-risk classes can tolerate different thresholds.
- Defect class by consequence
- False-accept cost
- False-reject cost
- Manual review capacity
- Escalation threshold
Make evidence part of the product
A PASS/FAIL bit is rarely enough for quality, traceability or continuous improvement. Retain the image or approved evidence crop, timestamp, model version, confidence, part identity, camera identity and final operator disposition where required.
That evidence allows quality teams to audit decisions, investigate disputes, retrain models and correlate defects with process conditions.
- Original or governed evidence image
- Detection/measurement overlay
- Serial, batch or work-order context
- Model and camera version
- Human review outcome
Monitor the vision system, not only the model
Model accuracy can degrade because the scene changed rather than because the model changed. Camera offline state, focus drift, dirty lenses, lighting degradation, shifted fixtures and image-distribution changes all matter.
A production platform should surface camera health, inference latency, no-read rate, confidence distribution, class distribution and review overrides. These signals create an early warning before the inspection becomes unreliable.
- Camera availability
- Image quality and focus
- Inference latency
- Confidence drift
- No-read and override rate
- Model/version deployment state
Put the decision where the operator already works
The fastest way to create resistance is to add another isolated dashboard. Vision results should enter the existing quality, HMI, MES or Andon workflow with clear review and disposition steps.
For an Ignition customer, that can mean surfacing VisionX evidence in Perspective while the inference runtime remains at the edge. For an MES-led site, the same evidence can be attached to the genealogy or quality-hold workflow.
- Single operator workflow
- Explicit review ownership
- Reason codes for overrides
- Traceability to production context
- Feedback loop for retraining
Primary sources used for this perspective
Product names and trademarks belong to their respective owners. Delfers engineering perspectives are intended as architecture guidance; outcomes depend on site conditions, system design and implementation.