Lighting designed for the defect
Low-angle, dome, backlight, polarised or strobed — chosen in our optics lab against your real samples so the defect has the highest possible contrast before any AI is involved.
Vision & Inspection
Catch the defect your line keeps shipping
We design and install AI inspection cells that see the scratch, short shot, open seal or missing part at full line speed — and act on it before the part reaches a customer.
The problem
Rule-based vision breaks on the defects that matter
Classic machine vision is excellent at measuring things that look the same every time: a hole position, a barcode, a label edge. It struggles with defects that vary — a hairline crack that runs in any direction, a seal wrinkle that looks different on every pouch, a cosmetic blemish on a textured surface. Teams compensate with tighter thresholds, and then fight false rejects for the rest of the year.
Human inspection catches more variety but fades with fatigue: attention on a repetitive visual task declines over a long shift, and a second shift rarely judges the same way as the first. Neither approach gives you a record of why a part was rejected.
An AI inspection system learns what good looks like and what each defect looks like from your own parts. It runs on hardware beside the line, decides in milliseconds, and keeps an image of every reject so quality teams can audit every decision.
How it works
Optics, compute and judgement in one cell
Most inspection projects succeed or fail on the first two items on this list. We start there.
Cameras that match the physics
Lensnode V global-shutter cameras freeze motion at line speed; thermal and multispectral variants see what visible light cannot.
Decisions at the edge
A fanless Gatepoint gateway runs the model beside the line with typical frame-to-I/O decisions under 40 ms — no network round trip.
Models trained on your parts
We combine supervised defect classes with anomaly detection, so the system flags new defect types it has never seen instead of passing them.
Reject logic your PLC trusts
Encoder tracking, reject-gate timing and handshakes are part of the job, not an afterthought. See PLC integration.
An audit trail for every part
Perflens Console stores each reject image with the model version, confidence and reason, ready for a customer audit.
What the model sees
From pixels to a decision you can audit
Drag the handle to compare the raw camera frame with the model's defect heat map. Each highlighted region is stored with the reject image.
Process
From your samples to a running cell in about four months
-
Feasibility Sprint
You send good and bad parts. We image them under several lighting geometries, train a baseline model and deliver a written go/no-go with modeled ROI.
-
Proof of Value
A loaner cell runs on your line in shadow mode against success criteria we agree in writing before we start.
-
Build and integrate
Mechanical mounting, enclosures, PLC and MES integration, factory acceptance test in our lab, then site acceptance test on your line.
-
Hand over and care
Operator training, documentation, and Perflens Care to monitor drift and retrain as materials, suppliers and seasons change.
Deliverables
What you receive
- Inspection specification and defect catalogue, signed off by your quality team
- Lighting and optics design with sample images for every defect class
- Installed Lensnode cameras, illumination and Gatepoint compute, mounted and guarded
- Trained model with a model card documenting data, accuracy and known limits
- PLC reject integration, stack-light logic and MES/historian data feed
- FAT and SAT reports with measured detection and false-reject rates
- Operator and maintenance training plus as-built documentation
- Exportable model files (ONNX) and full ownership of images and data
Typical performance
Typical system parameters
Every line is different; these are the ranges our standard designs are built around. Your Feasibility Sprint gives you numbers for your parts.
- Line speed
- Up to 600 parts per minute per camera, depending on field of view and resolution
- Decision latency
- Under 40 ms frame-to-I/O for typical models on Gatepoint G4
- Smallest detectable defect
- About 3–4 pixels across — e.g. 0.1 mm at a 50 mm field of view with Lensnode V
- Cameras per cell
- 1–12, synchronised by hardware trigger
- Integration
- Digital I/O, EtherNet/IP, PROFINET, Modbus TCP, OPC UA, MQTT
- Environment
- IP67 standard; IP69K stainless washdown option; −20 to 55 °C
Honest fit
When AI inspection is — and isn't — the right tool
A good fit when
- Defects vary in shape, size or orientation and rule-based vision produces too many false rejects
- Human inspectors disagree with each other, or results drop on later shifts
- You need an image record of every reject for customers or regulators
- Line speed makes manual inspection impractical
Probably not when
- The only requirement is a precise dimensional measurement — a calibrated gauge or classic vision tool is cheaper
- The defect is invisible without handling, cutting or destructive testing
- You produce fewer than a few hundred parts a day and inspection takes seconds per part
See it in context
Related example engagements
Heat-seal inspection at 180 packs per minute
Low-angle red light and a Gatepoint G2 find leaking seals on snack pouches in a modeled food-line scenario.
Read more
Weld-nut verification on existing cameras
How a stamping line could stop shipping brackets with missing nuts without buying new cameras.
Read moreFAQ
Questions we hear often
How many defect samples do you need?
For a Feasibility Sprint, 30–50 good parts and as many defective parts as you have — even 5–10 per defect type is useful. Anomaly-detection models learn mostly from good parts, and we can add synthetic defect data for rare defects once we have real examples to validate against.
Will it slow down my line?
No. The system inspects at line speed and decides in under 40 ms on the edge gateway. If the system faults, it fails to a state you choose — usually stopping the reject gate and raising a stack light rather than stopping the line.
What happens when the product or supplier changes?
Model drift is the most common reason inspection systems fade. Perflens Care monitors confidence and reject patterns, flags drift early and retrains the model on new material before detection drops.
Do our images leave the plant?
Not unless you want them to. Inference, storage and the Console run on-premise. Remote support uses an outbound, customer-controlled connection that you can switch off at any time.
Can we buy the system instead of subscribing?
Yes. Typical single-station systems are $48,000–$140,000 installed as a capital purchase, or from $2,950 per station per month as Inspection-as-a-Service with hardware replacement included. See pricing.
Send us ten good parts and five bad ones
A three-week Feasibility Sprint tells you, in writing, whether AI inspection will work on your defect — before you commit to a project.