Defect detection demos are convincing. A camera sees a scratched part and flags it. Deployed on a real line, the same setup produces false positives all morning and gets switched off by the afternoon shift.
Lighting is the model
More vision problems are solved by controlled lighting than by better architectures. Consistent, diffuse, purpose-built illumination turns a hard classification problem into an easy one. Ambient light through a factory window changes hourly and seasonally, and the model has no way to know why the images shifted.
Mounting and vibration
A camera that moves two millimetres invalidates every spatial assumption. Mount to structure, not to the machine that vibrates, and check alignment on a schedule rather than after a failure.
Plan for drift
- Suppliers change materials. Surface finish changes. The model was not told.
- Cameras accumulate dust and sensors age.
- Retain a labelled sample from production continuously, so retraining is a routine task rather than a project.
Decide the failure mode deliberately
A system that flags too much gets ignored; one that misses defects is worse than nothing because it creates false confidence. Choose which error you prefer, set the threshold to match, and tell the operators which one they are getting.