Industrial inspection
Surface defect detection on a stamping line
Inline detection of surface defects on stamped panels, inside the existing press cycle.
- 100%panels inspected (from sampled)
- <90 msinference per panel
- ~0.5%false-reject rate
Stack
- Basler area-scan cameras
- Dome + dark-field lighting
- PyTorch segmentation
- TensorRT on Jetson Orin
- PROFINET to line PLC
The problem
A Tier-1 automotive supplier inspected stamped panels by hand at the end of the line. Throughput limited the check to a sample, so cracks and draw marks reached assembly. They wanted full coverage without slowing the press.
Approach
- 01 Mounted area-scan cameras with controlled dome and dark-field lighting over the exit conveyor.
- 02 Trained a segmentation model on a few thousand labelled panels, with a synthetic set for rare crack types.
- 03 Ran inference on an edge GPU and gated rejects through the line PLC over PROFINET.
- 04 Added a review screen so operators could confirm borderline calls and feed corrections back.
Outcome
Every panel is now inspected within the press cycle rather than sampled. In trials the escape of visible surface defects dropped sharply, and the false-reject rate settled low enough that operators stopped second-guessing the system. Labelling effort, not model accuracy, was the main cost.
