AI Vision for Quality Control
Development of an AI vision system that detects product and packaging defects in real time on the packaging line of an East Frisian food manufacturer – right at the belt via edge cameras, with no cloud dependency and without slowing down the production cycle.
AI Vision for Quality Control
Artificial Intelligence
The Challenge
A family-run food manufacturer in East Frisia produces packaged fresh products at high throughput. Until now, final inspection was done by spot checks: staff pulled individual packs off the line and checked for misaligned labels, leaking or damaged film, missing best-before-date prints, and foreign objects. With several thousand units per shift, only a fraction could actually be inspected. Defects that slipped through led to retail complaints and, in the worst case, recalls – and the monotonous continuous inspection fatigued staff and tied up people needed elsewhere. A pure cloud solution was out of the question for the production environment: image analysis had to happen in milliseconds right at the line, independent of the internet connection, and could under no circumstances slow the production cycle.
Our Solution
Solution insights
We developed an AI-powered vision system that works directly on the packaging line. Industrial cameras capture every single pack; analysis runs on an edge device (NVIDIA Jetson) right on site – without a detour through the cloud. An object-detection model trained on the manufacturer's products (YOLOv8, optimised via ONNX Runtime) checks in real time for typical defect patterns: skewed or missing labels, damaged film, illegible or missing best-before dates, and visible foreign objects. When the system detects a defect, it triggers automatic ejection of the affected pack via a PLC connection and reports the event by MQTT to a dashboard. The model was trained together with the staff: over several weeks we annotated good and defective parts and progressively refined the system until false ejections dropped to a minimum. A Grafana dashboard shows defect types, rates, and trends per line and shift – turning quality assurance from a spot check into a complete metric.
Results
In the pilot, the system inspects 100% of a line's packs instead of the previous spot check – at full production speed. The rate of defective units shipped fell noticeably during the pilot, because label and film defects are now reliably sorted out at the belt before they leave the hall. Staff previously tied up in monotonous continuous inspection now focus on assessing the cases the system flags and on value-adding tasks. After the learning phase, detection accuracy for the trained defect patterns is around 98%, with a very low false-ejection rate. Because analysis runs entirely on the edge device, the system stays available even during internet outages – and sensitive production images never leave the plant.
Features
Feature overview
- Real-time inspection of every pack right at the belt (100% instead of spot checks)
- Detection of label defects, damaged film, and missing prints
- Best-before-date check for legibility and presence
- Foreign-object and visual inspection via object detection
- Automatic ejection of defective units via PLC connection
- Edge processing without cloud dependency (NVIDIA Jetson)
- Available even during internet outages
- Live dashboard with defect types, rates, and trends per line/shift
- Joint model training with annotation by the QA team
- Continuous refinement for new products and defect patterns
- Sensitive production images never leave the plant (data protection)
- Connection to existing line equipment via MQTT
Project Details
Client (scenario)
Food manufacturer (East Frisia, Germany)
Completed
2025
Technologies
Related decision paths
How a gut feeling in final inspection became a solid metric
Hilke has worked in final inspection for twelve years. Her job used to be: pull a pack, turn it, check it, put it back – shift after shift, thousands of times. 'You know you can't see everything,' she says. 'At some point it all blurs, especially on the late shift.' Today there's a screen in front of her showing only the packs where the system found something. She assesses the borderline cases, gives feedback on whether the AI was right, and thereby makes the system better day by day. 'At first I was sceptical whether a camera could do my job. But it doesn't do my job – it takes the strain off me. And for the first time I can show management in black and white how good our quality really is.'
Team Voices
"It was important to us that nothing has to go to the cloud. Our product images stay in the plant, analysis runs right at the line – even if the internet drops out."
Jens O.
Production Management
"They didn't just drop a model on top. We looked at good and defective parts together for weeks until the hit rate was right. That was real collaboration."
Sandra B.
Quality Assurance
"It helped that this comes from a company in the region: short distances, someone drops by and looks at the line in person instead of just sending tickets."
Heiko M.
Engineering & Maintenance
Partnership Instead of Project
- 1We trained the model not in a lab but on the real line with real products – together with quality assurance.
- 2Edge instead of cloud: the solution was built so that sensitive production images never leave the plant and the cycle never waits.
- 3Made in Germany from East Frisia – short distances, personal contacts, and on-site visits at the line.
- 4After go-live we continuously refine the model for new products and defect patterns.
More References
Planning a similar project?
Use our interactive cost calculators for an initial estimate – free and non-binding. Or schedule a consultation directly with our experts.