PCB & FPCInspection

Traditional automated optical inspection is rule-based and very good at what it was programmed for: component presence, placement offset, polarity marks, and solder volume against a threshold. It struggles when the defect is a judgment about appearance rather than a measurement, and it needs reprogramming for every new board. An AI approach learns from example boards instead, which shortens changeover on short product cycles and handles the cosmetic and mixed defects rules describe badly. Most real lines end up using both, rules where the tolerance is numeric and a model where it is visual.

Because they do not sit flat. On a rigid board the working distance is constant, so you set focus once and resolution is the binding constraint. A flex circuit curls and lifts, so the distance from the lens changes across the part and depth of field becomes the limit instead. You either constrain the part mechanically, close the aperture and add light to buy depth of field, or accept multiple captures. It is a lighting and optics problem before it is a software problem.

Often not, and it is better to find that out with arithmetic than after buying. A board is physically large while its features are very small, so a single field of view wide enough to see the board leaves too few pixels on the smallest feature. Either use several cameras or several positions, or inspect at the region level where the small features actually are. The mixed reflectivity of a board is a second reason to split it up: an exposure that suits matte laminate tends to blow out bright solder in the same frame.