A packaging line running faster than human inspectors could follow needed consistent detection of seal, print and label defects without slowing production. The proposed system inspects every unit at the edge and feeds results into the plant's existing quality workflow.
The manufacturer ran several packaging lines where sealed pouches passed inspection stations at a rate that made full manual inspection impossible. Operators sampled units periodically, which meant intermittent defects such as incomplete seals, smeared print and skewed labels were often caught only at the end of a batch or by downstream customers.
Each late discovery triggered rework, quarantine of entire batches and occasional customer complaints. Quality leadership wanted inspection of every unit, with results recorded against batch and shift so that recurring issues could be traced back to specific machines and settings.
Any solution had to fit into the existing line without adding stops, work under the plant's mixed LED and daylight conditions, and be maintainable by the on-site engineering team rather than a data science department.
Defects were rare relative to good units, so a naive training set would have been overwhelmingly imbalanced. The team needed an approach that could learn the appearance of acceptable product and flag deviations, while still classifying the known defect types that quality engineers wanted tracked.
Line speed meant each unit was visible for only a short window. Image capture had to be synchronised with the conveyor, and inference had to complete within that window so results could drive a reject mechanism rather than only a report.
Lighting varied across the shift as daylight through roof panels changed. Reflections on glossy film produced highlights that resembled print defects to an untrained model.
Several product variants shared the same line. The system had to switch inspection profiles when a changeover happened without requiring manual reconfiguration at the camera.
Industrial cameras triggered by the line encoder, with controlled strip lighting to suppress film reflections and stabilise exposure across the shift.
An NVIDIA Jetson module at each station running optimised ONNX models, so every unit is inspected locally and reject decisions are made in the line's timing window.
A segmentation model locates seal, print and label regions; a classifier and an anomaly detector then score each region against known defect types and learned normal appearance.
A FastAPI service records each inspection result with batch, variant and station metadata in PostgreSQL and emits reject signals to the line PLC.
A React interface for quality engineers to review flagged units, confirm or override decisions and export trends by shift, machine and variant.
Collected and labelled footage across all product variants and shift conditions, with quality engineers defining defect categories and acceptance thresholds.
Trained segmentation and classification models in PyTorch, then exported through ONNX and benchmarked on the target Jetson hardware to confirm they met the per-unit timing budget.
Ran a shadow deployment in which the system recorded decisions without controlling rejects, allowing side-by-side comparison with manual sampling and tuning of thresholds.
Integrated the reject signal with the line PLC and the inspection records with the existing quality management workflow via API.
Established a feedback loop in which engineer overrides in the dashboard become candidate training examples for the next model revision.
Every unit on the line is inspected rather than a periodic sample.
Defects are flagged and rejected at the station instead of being discovered downstream.
Recurring defect patterns can be traced to specific machines, variants and shifts.
Quality engineers maintain the system through the review dashboard without model expertise.
Defect detection, counting and packaging verification that run at line speed and connect directly to your control systems.
PPE compliance, restricted access and unsafe-behavior detection that help safety teams intervene before harm occurs.
Automate visual inspection, anomaly detection, product counting and manufacturing intelligence.
We can walk you through how this approach would translate to your environment and what it would take to pilot it.