The problem

A Flemish manufacturer of micro-ingredients and functional feed additives supplies products worldwide. Seam defects in filled bags can be difficult to spot before a package leaks in transit. For an export business, that creates a risk of container rejections, direct losses, and reputational damage. Manually checking every package on a continuous line was difficult, while sampling inevitably let some defects through.

Integrator Xylos describes a computer-vision system that inspects each bag after filling. Success criteria were set before development: detect at least 90% of defects, keep false alarms below 5%, and make a decision in under one second in real time.

How the system works

According to Xylos, YOLOX tracks each bag's position and movement on the conveyor so the inspection focuses on the right region. EfficientAD compares the seam with examples of correct products and flags deviations. This anomaly-detection approach can start with many normal seams without labeling every possible defect type in advance.

The models run locally on a network of Raspberry Pi devices with a central controller. This reduces the delay between the camera and a decision on the line and avoids sending every inspection image to the cloud. Before production use, the system ran alongside human inspectors for two weeks so detections could be compared and thresholds tuned. Operators remained as a safety net after the system became the primary check.

What the supplier reported

Xylos says the test period covered thousands of bags and met the targets for detection and false alarms. Six months after deployment, the integrator reports reduced product loss and fewer container rejections in export markets, as well as less inspector time spent on seam checks. These are supplier-reported results from one case, not an independent study or a guaranteed outcome for another production line.

How to adapt the approach to your factory

A pilot does not need to begin with “AI on every machine.” Start with one type of package and one measurable defect. Collect images of good products and confirmed defects under real lighting and conveyor speed. In observation mode, compare the model with inspectors without changing the production process.

Define four metrics in advance:
- share of defects detected;
- false stops or diversions of good products;
- decision latency and throughput;
- cost of a missed defect compared with a false diversion.

At first, the model can alert an inspector or mark a bag for re-checking. Automatic diversion is appropriate only after thresholds have been tested across shifts, SKUs, and lighting changes.

Limits and next step

Results depend on the camera, lighting, line speed, packaging, seasonal variation, and defect diversity. Local inference alone does not guarantee quality: version control, an event log, and a process for checking new product batches are still needed. Raspberry Pi worked in the described setup, but it is not a universal choice for every production workload.

Start with one line and a representative sample. First confirm what fraction of defects the system catches at an acceptable false-alarm rate; then evaluate payback and expansion to other operations.