Industrial Vision Inspection with AI

What problem does this solve?

Visual quality checks are the part of production that scales worst. An operator looking at parts for a full shift will find obvious defects reliably in the first hour and progressively less after that, and two operators rarely agree on borderline cases. When the line speeds up, sampling replaces full inspection, and the defects that escape are discovered by the customer instead of by the plant. In regulated sectors — food, pharma, packaging for export — the missing part is not only the detection but the record proving that each unit was inspected.

Many plants already own a classic machine vision system and are disappointed with it. Rule-based vision is excellent when the defect is a measurable geometric fact: a hole out of tolerance, a missing cap, a label offset by more than two millimetres. It becomes brittle when the defect is a class of appearance rather than a measurement — scratches on brushed metal, weld porosity, contamination on a textured surface, print smearing, colour drift. Every new product variant then needs another threshold, and the tuning work never ends.

The third failure mode is organisational. A vision cell is bought as a black box, the integrator leaves, and nobody in the plant can add a new defect class, retrain the model or explain to an auditor why a unit was rejected. Six months later the system is running with alarms disabled because it produces too many false rejects, and the quality team is back to manual checks with an expensive camera watching.

How the solution works

An AI vision inspection system treats the problem in the correct order: first make the defect reliably visible to a camera, then train a model on real production images, then wire the decision into the line and the plant systems. The imaging work — lens, working distance, resolution per defect feature, lighting geometry (bright field, dark field, diffuse dome, backlight), trigger and part presentation — decides most of the achievable accuracy, and no model recovers information that the optics never captured. Once images are consistent, a classifier, object detector or segmentation model is trained on labelled examples, validated on a held-out set, and deployed to an industrial edge computer next to the line so that decisions are made in cycle time without depending on a network link.

The output is not just a pass/fail. Each inspection produces a decision, a confidence value, a defect class and a stored image, which goes to the PLC as a reject signal and to MES or ERP as a traceable record against the batch, lot or serial number. Operators can review borderline cases in a queue; their corrections become new training data. Swedish Technology develops these systems as custom engineering projects and hands over the trained model, the labelling dataset, the retraining procedure and the integration code, so the plant can extend the system without depending on us.

1. 1
   Input / trigger A photo-eye, encoder pulse or PLC signal tells the camera exactly when the part is in position, so every image is taken from the same pose at the same point on the line.
2. 2
   Capture An industrial camera (GigE Vision / USB3 Vision, area or line scan) captures under controlled lighting. Exposure, gain and light geometry are fixed and documented; ambient light is shielded because it is the most common cause of a system that works in the morning and fails at night.
3. 3
   Pre-processing The edge node crops the region of interest, corrects distortion, normalises brightness and, where the part can move, locates the part first so the defect model always sees the same framing.
4. 4
   Inference The trained model runs on the edge device (GPU or NPU accelerated, typically ONNX Runtime or TensorRT) and returns a class, defect boxes or a defect mask with a confidence score, in a few milliseconds to tens of milliseconds per image.
5. 5
   Decision & integration A decision layer applies your thresholds and business rules — reject, divert, mark for manual review, or stop the line — and signals the PLC over digital I/O, OPC UA or fieldbus. The same event is posted to MES/ERP with batch, lot and defect code.
6. 6
   Action The rejector, diverter or robot acts within the line's time window; the operator HMI shows the defect image and reason so the shift can act on the cause, not just remove the part.
7. 7
   Reporting & learning Defect rates by class, shift, machine and product variant are reported; images from uncertain or disputed decisions collect in a review queue and feed the next retraining cycle, with model versions recorded for audit.

Reference architecture

Keep inference at the line, keep training separate, and keep the plant systems as the record of truth. A workable layout has five layers.

Deployment options: Everything needed for production runs inside the plant network: edge node at the line, training server in the data centre, no dependency on an external service. Where a customer prefers a private cloud in the UAE for training and dashboards, only images that the customer approves for export leave the site; fully air-gapped installations are supported with offline model updates on removable media.

Key capabilities

Integrations

The vision system must be a good citizen on the plant floor: fast and deterministic towards the PLC, and well-structured towards the plant's record systems. Integration work is custom development against the customer's own interfaces.

Industry use cases

Implementation approach

1. 1
   Feasibility (1–3 weeks) Define the defects in words and pictures, agree what a 'defect' actually is with the quality team, review sample parts and existing images, and state whether the defect is detectable at all with reasonable optics.
2. 2
   Imaging trial Test lens, working distance and lighting geometry on real good and bad parts, at the line's speed if possible. This step decides the project; a poor imaging setup cannot be fixed later with a bigger model.
3. 3
   Data collection and labelling Collect images across shifts, product variants, materials and lighting conditions; label with the quality team using an agreed defect taxonomy. Rare defect classes may need parts collected over weeks or produced deliberately.
4. 4
   Model training and evaluation Train and validate on held-out data; report false-reject and false-accept rates per defect class at chosen thresholds, and agree the operating point with the quality manager, not with the data scientist alone.
5. 5
   Integration build Edge node installation, PLC signalling, MES/ERP interface, operator HMI, image archiving and retention rules, alarm and health monitoring.
6. 6
   Shadow run (2–6 weeks) The system inspects and records but does not reject. Its decisions are compared with the current inspection process, disagreements are reviewed, and thresholds are adjusted with real evidence.
7. 7
   Go-live and acceptance Switch to active rejection with agreed acceptance criteria, operator training in Arabic and English, and documented fail-safe behaviour.
8. 8
   Hand-over and retraining cycle Source code, model artefacts, labelled dataset, retraining runbook and a schedule for periodic re-evaluation as products and materials change.

Security & deployment

The inspection cell sits on the plant network, not on the office network, and communicates with MES/ERP through a defined interface rather than direct database access. Edge nodes run a hardened OS with signed model artefacts and a versioned deployment, so an auditor can establish exactly which model version made a given decision. Image archives are subject to a retention policy agreed with the customer, since they can contain product and process information; access to review queues and to the retraining environment is role-based. Where the plant policy requires it, the whole solution operates without any outbound internet connection, with model updates delivered as signed packages on approved media.

Limitations & prerequisites

• If the defect is not visible in the image, no model will find it. Sub-surface flaws, internal cracks or contamination under an opaque layer need X-ray, ultrasound or thermography, not a normal camera.
• Accuracy depends on labelled examples. Rare defect classes are the hardest: a defect that occurs a few times per year cannot be learned from a two-week data collection, and anomaly-detection approaches trade that for a higher false-reject rate.
• There is always a trade-off between false rejects and missed defects. Pushing missed defects towards zero raises the number of good parts thrown away; the operating point is a business decision, not a technical constant.
• Changes to the product, material, supplier, lighting or camera position can silently degrade accuracy. Periodic re-evaluation and a monitoring plan are part of running the system, not an optional extra.
• Line-speed and cycle-time limits are real. Very fast lines may need line-scan cameras, multiple cameras, or a smaller model, and some inspections cannot be done in the available window at all.
• Environmental conditions — dust, steam, vibration, wash-down, heat — affect optics more than they affect software. Enclosures, air purge and cleaning routines have to be budgeted.
• The system reports what it sees; it does not explain the root cause. Linking defect classes to process parameters is a separate analytics exercise.

Manual inspection vs rule-based machine vision vs AI vision

These three are not competitors; most plants end up with a combination. The question is which defect belongs to which method.

A pragmatic plant keeps calibrated measurement in rule-based vision, gives appearance defects to the AI model, and keeps a human review queue for the borderline cases that neither method should decide alone.

FAQ

Related topics & solutions

• Predictive Maintenance AI for Industrial Assets
• Site Safety Computer Vision (PPE, Zones, Vehicles)
• Pipeline & Corrosion Vision Inspection AI
• GeoAI Object Detection with ArcGIS
• Smart Inspection
• AI Products Portfolio
• RAQEEB ÔÇô AI Surveillance
• Business Intelligence