Automated visual inspection

Injectable product must be inspected one hundred per cent, unit by unit, for visible particles — and “visible” has never been an optical abstraction: it is what a trained human can see. A camera pointed at a sealed vial photographs glass, liquid, meniscus and stopper, and a ten-micron fragment somewhere among them is invisible in a single frame. Detection therefore is not photography but motion analysis, and the machine’s first mechanical act is to mobilise what it wants to find.

Making a particle visible

The container is spun about its axis so the fluid rotates, then abruptly stopped; the contents keep turning from inertia while the glass is still. Frames taken in this window separate three populations. A real particle continues to move across successive frames and, being denser than the liquid, settles along a trajectory a bubble cannot follow: an air bubble is buoyant, rises, deforms and refracts light differently. A scratch or glass seed embedded in the wall does not move at all. Multi-frame trajectory tracking is thus the core algorithm — the physics does the discrimination, and the software only reads it. Optics set what the sensor can catch at all: back- and dark-field illumination arranged at multiple angles makes a particle a bright or dark silhouette against an engineered background, multi-view cameras cover zones the container wall hides, and lighting recipes are tuned per product because a dark or opalescent solution suppresses contrast for everything.

The bar that humans set

Manual inspection — vial held to a lamp, swirled, watched for a few seconds — built the regulatory concept of the visible particle, and its performance was measured statistically: detection probability rises with particle size, and the pharmacopoeial threshold sits where a panel of trained inspectors detects a defect with defined confidence. That curve is the bar. Fatigue, boredom and variation between and within inspectors were the manual method’s structural flaws; its virtues were an unmatched pattern library and honest treatment of the ambiguous case. Machines entered by having to demonstrate, in comparison tests of the Knapp lineage, that their accept/reject decisions matched or beat the human panel’s — not that they photographed smaller objects.

False accept against false reject

The decision is a threshold on a continuum, and the two errors are priced asymmetrically. A false accept ships a contaminated unit to a patient — a safety and compliance failure — while a false reject destroys a perfectly good vial of product that may be worth more than the machine; at line speeds of hundreds of units per minute, a percentage point of over-rejection is real money. Qualification therefore characterises both curves: detection probability as a function of defect size and type, false-reject rate on known-good units, and the settings become part of the validated recipe. The known limits follow from the method. A camera inspection cannot certify container-closure integrity — a hairline leak that admits microbes is a different physics, handled by deterministic leak tests — and a defect class absent from the training library can pass unchallenged, which is why synthetic and deliberately seeded defect libraries became central to the field.