Most inspection designs start with resolution: what is the smallest defect, what pixel size does that need. It feels rigorous. It is also, in my experience, the wrong anchor - because resolution is negotiable in many more ways than cycle time is.
The line runs at the speed it runs. That number was decided by people far above the inspection project, and it will not move. Design from it.
The budget comes first
Before any optics discussion, write the time budget:
line rate: 1,800 parts/hour -> 2.0 s per part, hard ceiling
part transport in: 0.4 s
positioning/settle: 0.3 s
part transport out: 0.3 s
available for imaging + processing: 1.0 sOne second. Every design decision now has a price tag in milliseconds, and trade-offs become arithmetic instead of debate.
How resolution explodes the budget
Resolution does not cost linearly - it costs quadratically in data, and often worse in time. Suppose one camera position covers the part at 20 um/px, and someone wants 10 um/px "to be safe":
at 20 um/px: 1 field of view -> 1 capture, ~0.1 s, 1x data
at 10 um/px: 4 fields of view -> 4 captures + 3 moves
4 x 0.1 s + 3 x 0.25 s (move and settle) = ~1.15 sThe "safer" resolution just consumed the entire one-second budget on acquisition alone - before a single algorithm runs. This is why halving the pixel size is never a small change: it quadruples the data and frequently adds the most expensive ingredient of all, motion.
Every field-of-view boundary also adds a stitching seam - and seams are where defects get missed and false calls breed. One field of view at adequate resolution routinely beats four fields at beautiful resolution.
Spending the budget well
- Buy back time with light. Exposure time is often the cheapest thing to shrink: doubling illumination intensity halves exposure. Photons are cheaper than mechanics.
- Overlap, don't queue. Process part N while acquiring part N+1. Pipeline design can nearly double effective throughput with zero new hardware - but only if the software is built for it from the start, not retrofitted.
- Move the camera less. Every mechanical move costs settle time. A larger sensor that avoids one move usually pays for itself immediately.
- Match the algorithm to the leftover time, honestly. Measure processing on the real image sizes early. "We will optimize it later" is how systems end up running at 80% of line speed - which means they do not run at all.
When the budget does not close
Sometimes the requirement genuinely does not fit the time. That is a valuable early result - say it clearly and offer real options:
- Inspect a sample (every Nth part) instead of 100%.
- Split inspection across two stations, each doing half.
- Relax the smallest-defect requirement - return to whoever owns the risk and ask what a miss really costs.
- Add a parallel lane. Expensive, but honest - unlike a system that silently cannot keep up.
What never works: quietly hoping the system will somehow be fast enough. The line rate always wins.
Resolution determines what the system can see. Cycle time determines whether the system exists. Anchor on the one that cannot move.