Multi-image inspection

A single image almost never contains all the information you need. Different lighting conditions reveal different things, and the strongest systems combine several captures.

Several captures fused into one reliable result — Normal, overexposed and angled captures fused into a single, more reliable decision.

A typical multi-image set

Normal exposure for the baseline.
An overexposed image to separate near-identical materials.
Different lighting angles to bring out topography.
A UV image for contamination and coatings.
An IR image for material penetration.

Each reveals different information. Fusing them gives the algorithm far more to work with than any single frame ever could.

A worked acquisition budget

Multi-image only works if it fits the cycle time, so budget it explicitly. Say the line gives you 800 ms per part for imaging:

frame                 exposure   light switch   readout    total
normal, ring          2 ms       1 ms           8 ms       11 ms
overexposed, ring     8 ms       1 ms           8 ms       17 ms
low-angle east        2 ms       1 ms           8 ms       11 ms
low-angle west        2 ms       1 ms           8 ms       11 ms
UV                    25 ms      2 ms           8 ms       35 ms
                                                sum:       85 ms

Five frames in 85 ms - comfortably inside the budget, because strobed LED lighting switches in microseconds and exposure dominates only for the dim UV channel. The expensive part is rarely acquisition; it is the processing of five images instead of one, so measure that too before promising a cycle time.

Simple fusion strategies that work

Per-channel rules: run a dedicated check on each frame (scratches on the low-angle frames, contamination on UV) and OR the results. Simple, debuggable, and you can explain every reject.
Difference images: east-minus-west low-angle frames cancel flat texture and double the response of real topography.
Min/max composites: a per-pixel minimum across lighting angles keeps only what is dark in every condition - very robust for true holes and voids.

Start with these before anything fancier. A fusion you cannot explain to an operator becomes a false-call dispute you cannot win.

The trade-off

More images means more acquisition time and more processing. The art is choosing the smallest set of images that fully separates the defect from everything else.

In practice I start with one frame, prove it insufficient with real samples, and add frames one at a time - each one justified by a defect class the current set misses. Sets designed this way stay small. Sets designed by enthusiasm grow until they blow the cycle time.