home / case studies / European FMCG contract manufacturer (composite)
AI / ML

Computer Vision Quality Inspection on a Packaging Line

We deployed edge-based vision models that inspect labels, seals and fill levels at line speed. Defects are caught on the line instead of at customer returns, with images retained for traceability.

European FMCG contract manufacturer (composite) · AI / ML

Computer Vision Quality Inspection on a Packaging Line

Challenge

Frequent packaging artwork changes and reflective films. We used synthetic augmentation, per-SKU few-shot fine-tuning and controlled lighting fixtures to keep accuracy stable across changeovers.

Approach

We captured and labelled a dataset with the client's QA team (CVAT), trained YOLO-family detection and segmentation models, and optimised them with TensorRT for NVIDIA Jetson edge devices beside each line. A PLC integration triggers rejection actuators. Uncertain images are routed to a human review queue, and reviewed samples flow back into a retraining pipeline. A central dashboard shows defect trends by line, shift and SKU.

Outcome

Consistent, full-line inspection instead of periodic sampling Illustratively ~50–70% fewer label and seal defects reaching customers Faster onboarding of new SKUs through a repeatable labelling workflow Image evidence for every rejected unit to support audits QA staff redeployed from repetitive checking to root-cause work

✓Consistent, full-line inspection instead of periodic sampling
~50–70%Illustratively fewer label and seal defects reaching customers
✓Faster onboarding of new SKUs through a repeatable labelling workflow

Problem solved

Manual visual checks were sampling-based and inconsistent between shifts, so mislabelled or poorly sealed batches still reached retailers.

What we built

We captured and labelled a dataset with the client's QA team (CVAT), trained YOLO-family detection and segmentation models, and optimised them with TensorRT for NVIDIA Jetson edge devices beside each line. A PLC integration triggers rejection actuators. Uncertain images are routed to a human review queue, and reviewed samples flow back into a retraining pipeline. A central dashboard shows defect trends by line, shift and SKU.

Benefits

Consistent, full-line inspection instead of periodic sampling Illustratively ~50–70% fewer label and seal defects reaching customers Faster onboarding of new SKUs through a repeatable labelling workflow Image evidence for every rejected unit to support audits QA staff redeployed from repetitive checking to root-cause work

YOLOTensorRTNVIDIA JetsonCVATPLC integrationPython

Talk to us