Automated visual inspection

01Overview and value chain

Markers: [EC: EU GMP Annex 1 (sterile products) / 21 CFR Part 11 (electronic records) | OECD: 2.8 Industrial biotechnology | Regulator: FDA (US), EMA (EU), NMPA (CN)]

Automated visual inspection (AVI) is the 100% end-of-line quality gate for sterile injectables: every filled and sealed container is imaged by high-resolution industrial cameras under multi-angle LED arrays while being spun or handled, and machine-vision algorithms decide whether it carries visible foreign matter, a fill-level deviation, a cosmetic glass defect or a seal failure. It replaces the manual light-box inspection in which operators stared at vials against a bright background, a task that is physically exhausting and inevitably lossy. Modern systems resolve particles down to roughly 40 micrometres — about half the diameter of a thick human hair — and run at throughputs such as 600 prefilled syringes per minute, with full audit-trail traceability to satisfy GMP data-integrity expectations. The category’s central engineering problem is the false-reject rate: over-sensitive algorithms discard good product, and the hardest discrimination is between an air bubble and a genuine solid particle, a distinction that also degrades against dark or low-contrast solutions. This is why the field has moved decisively toward deep learning, and why AVI is now specified as a dedicated machine rather than a camera station bolted to a filling line — the equipment specialists tabled here build inspection machines as their core product, distinct from the aseptic filling-line OEMs whose lines merely include an inspection stage.

The key directions of automated visual inspection are:

1. Machine-Vision Particle Detection: high-resolution industrial cameras with multi-angle LED arrays and particle-trajectory tracking that distinguish moving foreign matter from bubbles, resolving visible particles at roughly 40 micrometres in vials, ampoules and prefilled syringes.
2. Deep-Learning Defect Classification: convolutional models that classify defect types automatically and are trained on synthetically generated rare-defect scenarios, addressing the chronic shortage of real defect samples that previously left algorithms unable to generalise to a new product.
3. Container Closure Integrity Testing (CCIT): deterministic leak-detection methods — high-voltage leak detection, vacuum and pressure decay, and camera-based stopper-displacement monitoring — that establish sterile-barrier integrity, a property camera-based cosmetic inspection cannot legally certify on its own.
4. High-Mix Low-Volume Inspection: compact, fast-changeover machines built for clinical trials, ATMPs and CDMOs, where batches are small, formats vary constantly and changeover time rather than peak throughput sets effective capacity.

Sectoral value chain

[Filled & Sealed Units] ──> [Container Handling] ──> [Multi-Camera Imaging] ──> [Defect Classification]
                                                                │
                                                        (reject decision)
                                                                │
                                                                ▼
[Released Batch] <─── [Reject Verification & Records] <─────────┘

Value chain levels

Cross-cutting technologies of the sector:

• Embedded Edge-AI Vision: industrial vision systems with dedicated on-device AI processors that execute inference deterministically at full line speed rather than deferring to a server, supplying the component layer beneath the machine builders.
• Synthetic Defect-Data Generation: AI models that synthesise rare defect scenarios and combine them with real defect libraries, training larger detection models without waiting for rare defects to occur naturally.
• Distributed Inspection Architecture: distributed rather than client-server system design for vision platforms, improving flexibility and scalability while keeping every inspection traceable and every operation controlled.

02US

The US contribution to automated visual inspection is concentrated in the vision component layer rather than in complete pharmaceutical inspection machines: American industrial-vision houses supply the cameras, embedded AI processors and software toolkits that the European and Chinese machine builders integrate.

embedded edge-AI vision components, deep-learning toolkits, machine-builder integration

• Cognex: its In-Sight 3900 embedded AI vision system, built on Qualcomm Dragonwing platforms, delivers up to four times the processing speed of previous Cognex systems, imaging to 25 MP, and real-time edge AI with deterministic, high-throughput inspection at full line speed — the component class on which pharmaceutical inspection recipes are built.
• Component-layer positioning: US vision suppliers sell into inspection machines rather than competing with them, which is why the tabled US entry is a vision-system maker and not a vial-inspection OEM; the same operator-and-component pattern seen in other bioproduction equipment categories.
• Regulatory driver: FDA expectations for 100% inspection of sterile injectables and 21 CFR Part 11 electronic records shape what US-supplied vision software must log, making audit-trail capability a purchasing criterion rather than an optional feature.

03CN

China has moved from importing inspection machines to building AI-differentiated ones, with domestic pharmaceutical-equipment leaders now shipping intelligent lamp-inspection machines (全自动智能灯检机), while Chinese analyst coverage is unusually explicit about the technical limits that remain.

AI particle-trajectory tracking, domestic lamp-inspection machines, candid limitation reporting

• Truking Technology: its self-developed AI particle-trajectory tracking algorithm identifies visible foreign matter down to 40 micrometres — described as about half the diameter of a thick hair — through dynamic tracking and multi-dimensional analysis; the team also built an AI vision platform that synthesises rare defect scenarios from an accumulated defect database to train larger models, and a new-generation TrukingVision system that replaces the traditional client-server design with a distributed architecture. Coverage reports the machines extending from vials to more than ten package types including oral liquids and blow-fill-seal, a roughly 5% reduction in false-reject rate versus the previous generation, and close to 6,400 domestic and international patent applications filed.
• Domestic sector profile: Chinese analyst reporting describes fully automatic lamp-inspection machines as standard across pharmaceutical, biological-product and premium-food production for detecting visible foreign matter, fill deviations and seal defects in ampoules, vials and prefilled syringes, built on high-resolution industrial cameras, multi-angle LED arrays and machine-vision algorithms with GMP audit-trail compliance.
• Candid limitation reporting: the same analyst coverage states plainly that some models discriminate poorly between bubbles and solid particles, that dark solutions reduce contrast, and that weak algorithm generalisation forces extensive sample training when a new product is introduced — with the projected path running through deep learning, near-infrared assistance, acoustic sensing for loose caps and digital-twin operations.

04EU

Europe holds the dedicated inspection-machine category, and the specialists differ by the problem they solve: syringe throughput, unified inspection software, integrity testing as a discipline distinct from cosmetic vision, and small-batch flexibility for clinical and advanced-therapy work.

Brevetti CEA syringe inspection, Antares Vision INSPECTA platform, Wilco integrity testing, Koerber small-batch ALVA

• Brevetti CEA: the Italian manufacturer specialises in high-precision visual inspection for injectables; its K15 DR integrated system inspects prefilled syringes at up to 600 per minute with a fully automatic tub-and-nest cycle, de-nesting, inspection and re-nesting by two robot arms, no-contact handling to avoid inducing bubbles, high-voltage leak detection, automatic refilling of nests for rejected syringes and 21 CFR Part 11 compliance, with its INTELLIGENT-VISION platform covering particles, fill level, cosmetic defects, needle shield and leak detection.
• Antares Vision: its INSPECTA platform is a unified, modular software suite for pharmaceutical vision inspection that integrates with primary and secondary packaging machinery, providing presence/absence and foreign-object detection, colour control against references, in-line leak detection for aerosols and blister filling control, alongside a traceability stack spanning EU-FMD and DSCSA requirements.
• Wilco AG: the Swiss house is a fifty-year specialist in leak testing and visual inspection for pharmaceutical, biotech and medtech packaging, combining camera and X-ray inspection with AI-supported image processing and integrating multiple test technologies in one system, active in more than eighty countries. Its regulatory point is the sharpest in the category: under the revised Annex 1, visual inspection is no longer a legally valid integrity test for fusion-sealed blow-fill-seal containers, so camera-based cosmetic checks cannot substitute for deterministic container closure integrity testing.
• Koerber (Seidenader): its ALVA inspection machine targets high-mix low-volume production — fast changeovers, automated operation and a compact footprint for small-batch syringes, vials and cartridges — aimed at clinical trials, ATMPs and CDMOs where product variability and changeover cycles, not peak speed, govern throughput.

05Leading companies and research institutes

06Tech stack and innovations

The AVI stack layers optics and handling beneath classification intelligence, with integrity testing running alongside as a separate deterministic discipline rather than a vision output.

1. Optics, Illumination and Handling:
   • High-resolution industrial cameras paired with multi-angle LED arrays image each container from several viewpoints while it is spun to mobilise particles; lighting geometry is tuned per recipe to the container material and the colour and transparency of the solution.
   • Contact-free handling matters as much as optics: robotic tub-and-nest de-nesting and re-nesting avoids inducing the very bubbles that would then be classified as defects, and supports plastic as well as glass syringes carrying liquids, viscous products and suspensions.
2. Classification Intelligence (Rule-Based to Deep Learning):
   • The discrimination that defines system quality is bubble versus solid particle; dark solutions cut contrast and weak generalisation forces retraining for each new product, which is why convolutional models that classify defect type automatically have displaced purely rule-based logic.
   • Synthetic defect generation resolves the sample-scarcity problem directly: rare defect scenarios are generated from an accumulated defect database and combined with real data to train larger models, replacing the previous practice of waiting for suitable rejects to appear.
3. Container Closure Integrity Testing:
   • High-voltage leak detection is integrated into inspection machines, while vacuum and pressure decay and camera-based stopper-displacement monitoring quantify sterile-barrier integrity; deterministic methods are what the revised Annex 1 expects, and for fusion-sealed blow-fill-seal containers visual inspection is explicitly not a valid substitute.
   • Combining leak testing with camera and X-ray inspection in a single system, with AI-supported image processing, lets one station return both a cosmetic verdict and an integrity verdict.
4. Architecture, Records and Changeover:
   • Distributed system architectures replace traditional client-server designs to improve flexibility and scalability while keeping every inspection traceable, every dataset queryable and every operation controlled — the practical shape of GMP data integrity.
   • For clinical, ATMP and CDMO duty the binding constraint is changeover, so compact high-mix low-volume machines optimise fast format changes and automated setup instead of peak units per minute.

07Value chains and production pipelines

Industrial pipeline of an automated visual inspection campaign (EU GMP Annex 1, 21 CFR Part 11, FDA / EMA / NMPA frameworks)

┌───────────────────────────┐      ┌───────────────────────────┐
│ 1. Recipe & Format Setup  │ ───> │ 2. Container Presentation │
└───────────────────────────┘      └───────────────────────────┘
                                                  │
                                                  ▼
┌───────────────────────────┐      ┌───────────────────────────┐
│ 4. Defect Classification  │ <─── │ 3. Multi-Camera Imaging   │
└───────────────────────────┘      └───────────────────────────┘
               │
               ▼
┌───────────────────────────┐      ┌───────────────────────────┐
│ 5. Integrity Testing      │ ───> │ 6. Reject Handling &      │
│                           │      │    Batch Records          │
└───────────────────────────┘      └───────────────────────────┘

Stage 1: Recipe and Format Setup

The container format — ampoule, vial, prefilled syringe or cartridge — and the solution’s colour and transparency fix the lighting geometry, camera configuration and classification model. Vial programmes typically span several diameters across a 2 to 50 millilitre range, and in high-mix low-volume work this setup step, not imaging speed, is the throughput constraint.

Stage 2: Container Presentation

Filled and sealed units are fed from the line or de-nested robotically, then spun or rotated so that any suspended particle moves relative to the container. Handling is deliberately contact-free, because mechanically induced bubbles are indistinguishable from defects to a camera and inflate the false-reject rate before classification even begins.

Stage 3: Multi-Camera Imaging

Each unit passes a series of camera stations under multi-angle illumination, generating a multi-view image set per container at full line speed — up to 600 prefilled syringes per minute on a high-throughput syringe machine. Embedded edge-AI processors execute inference deterministically at that speed rather than deferring analysis downstream.

Stage 4: Defect Classification

Algorithms assign each unit an accept or reject verdict with a defect class: visible foreign matter such as fibres, glass or metal fragments, fill-level deviation, cosmetic glass faults including cracks and delamination, and closure faults such as a missing stopper or cap. Particle sensitivity reaches roughly 40 micrometres, and the quality of the bubble-versus-particle discrimination sets how much good product the line discards.

Stage 5: Integrity Testing

In parallel with the cosmetic verdict, deterministic integrity methods establish that the sterile barrier holds — high-voltage leak detection integrated into the inspection machine, or vacuum and pressure decay and stopper-displacement monitoring. For fusion-sealed containers this test is mandatory rather than optional, since camera-based inspection cannot legally certify integrity.

Stage 6: Reject Handling and Batch Records

Rejected units are segregated, optionally re-verified through fail-safe reject verification, and nests are automatically refilled where the format requires it. Every unit result, operator action and parameter change is written to an audit trail supporting 21 CFR Part 11 electronic records, and the accumulated defect data feeds back into model training for the next campaign.