Conductive substrate inspection

Anonymized case study. No customer or product is identified.

The problem

Conductive substrates made of ceramic and bonded metal layers carry defects that are subtle at normal exposure: the defect and the surrounding material reflect light almost identically.

A reflective metal substrate with a low-contrast feature — A bonded metal substrate. The defect and the background reflect light almost identically until the lighting is tuned to separate them.

Constraints

Defects micron-scale and low-contrast.
High throughput required for production.
Surface finish varied between batches.

Approaches tried

Single coaxial image - revealed reflectivity differences but missed texture-type defects.
Overexposure - deliberately saturating the image separated the defect brightness from the background.
Multiple lighting angles - brought out topography that flat lighting hid.

Each approach was scored the same way: contrast of the worst defect sample against the noisiest good sample. The single coaxial setup managed roughly 2x noise on the hardest defects - technically visible, practically a false-call generator. The combined strategy reached 8-10x on the same samples.

Solution

A multi-image strategy: a normal exposure as baseline, an overexposed frame to separate near-identical materials, and angled lighting for topography. The algorithm then fused these into a single reliable decision.

The acquisition sequence ran four frames in under 100 ms using strobed lights, so throughput survived. Processing was kept deliberately simple - per-frame thresholds and a rule-based fusion - so that line engineers could understand and adjust every stage without calling anyone.

What failed first in the field

Not the optics and not the algorithm: the batch-to-batch surface finish. A new etching parameter upstream changed the background texture enough to shift the overexposure clip point, and false calls spiked for two days. The fix was procedural, not technical - the inspection recipe gained a per-batch calibration step using a known-good sample from each lot.

Lessons learned

The contrast was created by the lighting design, not the algorithm.
Batch-to-batch surface variation, not the defect itself, was the hardest part to keep stable.
A simple algorithm the site could maintain beat a smarter one they could not. The system survived because every false-call dispute could be traced to a specific frame and threshold within minutes.