On a populated board, most solder joints are cosmetically variable and perfectly acceptable. The rare genuine defect — a bridge, insufficient solder, a lifted lead — is the exception. So the everyday behaviour of an inspection is dominated not by catches but by how it treats the thousands of good, imperfect joints it sees. Tune it to catch everything and it flags a flood of them; an operator has to clear each false call by hand, and once the rate is high they stop trusting the machine and start rubber-stamping it. At that point the inspection has failed, regardless of how many real defects it could technically find.
The fix is to teach the system what a good joint really is, in all its variety. A good fillet varies enormously in shape and sheen, and a model that has only seen a narrow slice reads that legitimate variation as a fault. Training on the true distribution of acceptable joints is what lets the system question only genuine deviations. Multi-angle imaging helps, because some of the worst defects — a lifted lead, a hollow joint — are three-dimensional and look acceptable from directly above; capturing the joint's actual shape, not just its top-down appearance, is what tells a real defect from a cosmetic one.
Then the threshold is set on cost, not on an accuracy number. The right operating point balances the cost of a false call against the cost of an escape, weighted by the real consequence of each defect type rather than by how often it appears. A rare but critical defect is worth catching even at the price of a few more questions; a cosmetic quirk is not worth a single one. And every operator verdict is folded back into training, so the false-call rate keeps falling and the trust keeps building — which is the only thing that keeps the system switched on.