Automated Visual Inspection (AVI)
Head of Innovation
Digital Transformation, AI, IoT, Machine Learning, and Cloud Technologies. Nearly 15 years driving innovation
Automated Visual Inspection (AVI): System Design, Camera-Lighting Selection, and Production-Line Deployment
Most automated visual inspection (AVI) projects fail at the procurement stage. Engineering teams pick a camera from a vendor catalogue, attach a generic ringlight, hand the system to a data-science team to "train a model on it," and then spend six months trying to recover from imaging choices that were wrong before a single image was captured. AVI is fundamentally a system-design discipline. The model only does what the imaging chain allows. This article works through the system-design decisions that determine whether an AVI deployment hits 99% accuracy in eight weeks or stalls at 92% forever.
The audience is plant engineers, controls integrators, and the manufacturing-IT teams who own the integration layer between vision and the rest of the production stack. We will walk through the system architecture, the camera-and-lighting selection method, and the production-line deployment sequence we use on customer engagements.
Define the Inspection Specification Before You Pick Hardware
The AVI specification document is one or two pages and answers six questions. Skip any of them and the hardware decision is guesswork.
- Smallest detectable defect — in millimetres on the part surface, with the rule of thumb that you need 3-5 sensor pixels across the smallest feature. A 0.3 mm scratch on a 200 mm field of view requires roughly 2,000 pixels across, i.e. a 5 MP camera or larger.
- Field of view (FOV) — the physical area each frame must cover. Drives both pixel pitch and lens working distance.
- Cycle time per inspection — derived from line rate. A 60 ppm line gives you 1 second per part total, of which inspection is typically a 200-400 ms budget.
- Lighting tolerance to ambient — does the station have a light-tight enclosure, or does sunlight through a roof skylight create a shift change at noon? This determines whether you need active illumination plus shielding or passive imaging.
- Acceptable false positive and false negative rates — set by quality, not engineering. A safety-critical pharma application might accept 3% false positive in exchange for zero false negative; a cosmetic-grading consumer application might accept 1% false negative to keep yield high.
- Failure-mode containment — what happens if the vision system fails? Stop the line, tag the unit for human re-inspection, or run-and-flag for downstream review?
A spec answering these six questions plus a representative defect catalogue is the input to the hardware decision. Without it, every reviewer has a different mental model of the requirement and the design conversation goes in circles.
The Camera Selection Method
Camera choice splits along two axes: area scan vs line scan, and 2D vs 3D.
| Camera type | Sensor | When to use | Common manufacturers |
|---|---|---|---|
| Area scan, mono | CMOS Sony IMX, ON Semi Python | Discrete parts, snapshot inspection, fixed FOV | Basler ace, FLIR Blackfly, IDS uEye |
| Area scan, colour | CMOS with Bayer filter | Label, print, colour grading, food | Basler ace U, Cognex In-Sight |
| Line scan | CMOS line sensor 2k-16k pixels | Continuous web (paper, foil, textile, sheet metal), cylindrical parts on rotating fixtures | Teledyne DALSA, Basler racer, Hamamatsu |
| 3D profiler (laser triangulation) | Mono CMOS + laser line | Surface profile, weld bead, height/depth measurement | LMI Gocator, Keyence LJ-X, SICK Ranger |
| 3D structured light | Projector + stereo CMOS | Painted-surface inspection (deflectometry), free-form geometry | Zivid, Photoneo, Cognex 3D-A |
| Stereo vision | Two CMOS area sensors | Depth in robotics-bin-pick, guidance | Intel RealSense D400, ZED 2i |
CMOS has effectively replaced CCD across the industrial-imaging market over the past decade. Modern CMOS sensors deliver lower read noise, higher frame rates, and global shutter at price points where CCD used to be the only option. The remaining CCD niches are very-low-light scientific imaging and a handful of legacy line-scan applications. For a new AVI design in 2026, the default is global-shutter CMOS unless you have a specific reason to choose otherwise.
Frame rate matters as much as resolution. A 12 MP camera at 16 fps is useless on a 60 ppm line that needs 5 captures per part because you cannot synchronise the trigger. Real-world targets are 30-100 fps for most discrete-parts AVI and 50-100 kHz line rates for line-scan continuous-web inspection.
Need expert help with automated visual inspection (avi)?
Our cloud architects can help you with automated visual inspection (avi) — from strategy to implementation. Book a free 30-minute advisory call with no obligation.
The Lighting Selection Method
Lighting is the single biggest determinant of AVI performance. Eight techniques cover 95% of industrial cases, and the choice flows from the optical interaction between the surface and the defect, not from "what we have on the shelf."
- Diffuse dome (cloudy-day) — eliminates specular highlights on reflective curved surfaces. Default for shiny metal, painted plastic.
- Coaxial — beamsplitter places light along the camera optical axis. Reveals scratches and surface texture on flat reflective surfaces.
- Backlight — silhouettes the part. Default for dimensional measurement and transparent containers (vial fill levels, bottle shape).
- Bar / linear — line-scan partner; provides uniform brightness across the line direction.
- Ring — cheap, omnidirectional, decent default for moderately diffuse surfaces.
- Dark-field — light at low angle to the surface. Surface scratches become bright on a dark background, ideal for engraved-text reading and machined-surface inspection.
- Structured light / fringe projection — encoded patterns enable 3D reconstruction and deflectometry for painted-surface defects (orange peel, sags).
- Multi-spectral (UV, NIR, SWIR) — UV reveals organic contamination and security inks; NIR penetrates plastic packaging; SWIR distinguishes materials by absorption signature.
The single most-skipped step is imaging proof of concept before model training. Capture 50-100 representative parts with the proposed camera-lens-light setup, label them by hand, and confirm that defects are visually distinguishable in the raw image. If a human expert cannot pick them out of the captures, no model will. This is a one-day exercise that saves three-week training detours.
Acquisition, Triggering, and the Frame-Grabber Layer
Discrete-part AVI requires hardware triggering synchronised with the part moving past the camera. The standard chain is: encoder on the conveyor, encoder pulse to the PLC, PLC issues a digital trigger to the camera or frame grabber, camera captures, image flows over GigE Vision / USB3 Vision / CoaXPress to the inference host. Software triggering on a free-running clock works for low-rate prototypes and fails on production lines because parts and frames go out of phase within minutes.
For 10 GigE or CoaXPress applications, the frame grabber (Matrox, Euresys, Pleora, BitFlow) handles direct-memory-access transfer, removing CPU bottlenecks. For most modern installations on Jetson Orin or AGX Xavier, the camera connects directly via GigE Vision with PoE+ power, no frame grabber required.
The Production-Line Deployment Sequence
An AVI system is deployed in a sequence we have run dozens of times. Skipping or compressing stages is the most common cause of post-launch escapes.
- Bench validation (week 0-2) — assembly of camera, lens, lighting, and edge box on a lab bench. Capture and grade representative parts. Confirm imaging spec.
- Mechanical mounting design (week 2-4) — fixture, vibration isolation, light enclosure. Done in CAD with the controls integrator. The mechanical-mount and light-shield design is where most AVI projects underspend.
- Initial model training (week 3-6) — capture 1,000-3,000 production images across SKUs and defect classes. Train baseline classification, segmentation, or anomaly model. Validate against a held-out test set.
- PLC and MES integration (week 4-7) — wire OPC-UA or Modbus TCP to the existing PLC; map MES tags for traceability; configure reject-actuator timing.
- Shadow run (week 6-9) — system runs in parallel with manual inspection, no actuation. Compare AI calls to inspector calls daily. Tune thresholds.
- Cutover and ramp (week 9-12) — system takes over actuation. Manual inspector role shifts to override review. Ramp from 10% of parts to 100% over 1-2 weeks.
- Steady-state monitoring — dashboards on accuracy, false-positive rate, throughput, and model drift; monthly retraining cadence.
The Edge-Compute Choice for Production AVI
Edge-compute selection is the third pillar after camera and lighting. The dominant options:
- NVIDIA Jetson Orin Nano / NX / AGX — TensorRT-optimised inference, 20-275 TOPS depending on SKU. Default for new AVI designs that need 30+ fps on multi-megapixel cameras.
- Intel OpenVINO on Core i7 / Xeon-D industrial PC — better C++ tooling integration and richer Windows ecosystem. Default for customers with existing Windows-based vision infrastructure.
- AWS Panorama — managed appliance that hides the device-management problem. Good for multi-site fleets; weaker for highly customised pipelines.
- Industrial GPU PC (Advantech, Beckhoff, Siemens IPC) — RTX A2000/A4000 cards, full PC OS, full flexibility. Default when inference budget is multiple high-MP cameras at high frame rate.
The AVI inference target is end-to-end under 50 ms per image, including pre-processing, inference, and post-processing. Latency budgets above that fail to keep up with most modern lines.
How Opsio Helps
Opsio's engineering team designs and deploys AVI systems end-to-end — imaging spec, camera-lighting engineering, mechanical mounting, edge inference, and PLC/MES integration. We default to NVIDIA Jetson Orin or industrial GPU PC platforms and to GigE Vision cameras from Basler, FLIR, and Cognex. Customers consolidating multiple inspection projects under a single managed-service model use our visual inspection services as the cluster pillar; adjacent capabilities include computer vision consulting for the model-architecture work, manufacturing defect detection for line-specific implementations, and IoT services provider capabilities for the connectivity layer between edge boxes and the cloud retraining stack.
About the Author
Head of Innovation
Jacob leads innovation at Opsio, specialising in digital transformation, AI, IoT, and cloud-driven solutions that turn complex technology into measurable business value. With nearly 15 years of experience, he works closely with customers to design scalable AI and IoT solutions, streamline delivery processes, and create technology strategies that drive sustainable growth and long-term business impact.
Editorial standards: This article was written by a certified practitioner and peer-reviewed by our engineering team. We update content quarterly to ensure technical accuracy. Opsio maintains editorial independence — we recommend solutions based on technical merit, not commercial relationships.