Twelve GigE Vision cameras capture up to 14 images of every filled ampoule at Aspen's pharmaceutical lines, replacing manual particle inspection with a deep-learning pipeline built on MVTec HALCON. The system reached production status on two lines after increasing foreign-particle detection and reducing false negatives, according to Vincent Trombetta, Automatic Visual Inspection Expert at Aspen.

Why machine vision and not human eyes?

The inspection target is small, translucent plastic ampoules filled with viscous drug solutions. Bubbles in the liquid mimic foreign particles, and contaminants may sit on the side wall, sink to the bottom, or fade into the medium. A trained human operator faced a slow, error-prone task that Mickael Denis, Manager Operationnel Vision Industrielle at Aspen, called central to quality assurance.

"Quality assurance of pharmaceutical products is extremely important. Therefore, it was essential that the new solution matched the detection rates of the previous process, which involved inspection by human operators, or ideally even surpassed them," Denis said.

How does the imaging chain work?

The ampoules are blown, filled, and sealed in one clean-room machine, then transferred to an inspection station. A 12-camera array follows the GigE Vision standard and is positioned at multiple stations around each vial. The system produces up to 14 images per ampoule, which the engineers use as a redundancy mechanism: a defect counts only when it appears in a minimum number of frames, suppressing spurious hits.

Key imaging and processing parameters:

• 12 GigE Vision cameras per line
• Up to 14 images per ampoule captured at different stations
• Industrial PC host running MVTec HALCON
• Deep-learning classifier for foreign particles
• Classical matching and blob analysis for fill level, colour, and closure

Why deep learning instead of rule-based vision?

Classic machine vision depends on deterministic feature rules. Aspen's medium made those rules fragile. Bubbles, viscosity artefacts, and variable lighting defeated hand-coded thresholds. Trombetta's team concluded that only a neural approach could learn the visual signature of real contamination in this matrix.

"It was clear that such a task could only be automated using deep learning technologies," Trombetta said. "We also had to adapt to the particularly demanding validation processes that apply in the pharmaceutical industry."

Deep learning needs training data, and pharmaceutical production lines provide it: large volumes of labelled images under controlled lighting. That data supply was the practical justification for choosing a learning-based pipeline over a rule-based one.

What tasks does the system run in parallel?

The HALCON software stack runs several checks on every captured frame:

• Foreign-particle detection in the liquid (deep learning)
• Fill-level verification (matching and blob analysis)
• Colour conformance check (classical methods)
• Closure specification check (classical methods)

A pass/fail decision is written for each ampoule at the end of inspection. Test vials are placed manually on the conveyor, then travel through the imaging stations without operator intervention.

What compliance questions remain?

Aspen validated the system against the pharmaceutical industry's process-validation regime, which the source describes as "particularly demanding" but does not name by reference standard. Two lines now run in production. The company plans further machine-vision automation on adjacent processes, though no specific standard revision (such as USP <1790> or GAMP 5) appears in the source material. The unresolved buyer question: how the validated false-negative rate compares quantitatively with manual inspection baselines, and whether regulators will accept multi-frame consensus as a primary release criterion.