Vision inspection with edge inference and human review
A fabricator stamping galvanised brackets and panels for air-conditioner makers needed consistent inspection across shifts. Over 20 weeks we installed cameras and edge inference on one production line, with automatic rejection, human review and an Arabic operator interface.
Screens
The challenge
One person inspected parts beside a moving line, and detection varied by shift. Customer returns arrived weeks later, with little evidence tying a fault to its batch or die, so recurring defects were hard to investigate.
Our approach
We paired controlled lighting and cameras with inference on an edge box so decisions could happen at line speed. Working with the QC lead, we labelled about 4,000 images from the client's reject bins over three weeks. Clear rejects divert automatically. Borderline cases wait for a person. Grades post to the ERP against batch and die.
Outcome
- Every part graded in about a third of a second at line speed
- Defects reaching customers reduced by 57% across the first two quarters
- 12% of parts held for human review to limit costly false passes
- Defect rates available by batch and die during the same shift
“It does not blink and it does not get to the end of a shift. And when it is not sure it stops and asks, which is more than the checklist it replaced ever did.”Turki Al-Ghamdi — QC lead, Safeeh Industries
Original interfaces and workflows are protected by NDA. Images are UI recreations using synthetic data and figures. Client identity is shared with permission.