Building AI-Powered Visual Quality Control for Plastics Manufacturing
Deploying a Custom Vision defect-detection system to catch flash, shrink, parting-line, and surface defects on the bottling line - running fully offline, in real time, at the point of inspection.
Industry
Time Required
Engagement Model
Solution
Current Phase
Technologies
All Time Plastic manufactures PET bottles at high volume, where even a small number of defective bottles reaching the packing line can mean rejected shipments, customer complaints, and rework costs. Manual visual inspection was slow, inconsistent between shifts, and couldn’t scale with production speed.
This case study explores how Zelite Solutions deployed an Azure Custom Vision defect-detection system that runs fully offline, in real time, at the point of inspection.
All Time Plastic needed a solution to address:
- Manual inspection bottlenecks: Human inspectors could not consistently catch subtle defects – flash lines, shrink marks, parting-line irregularities, scratches, and silver streaks – at production-line speed.
- Inconsistent defect classification: Different inspectors judged borderline cases differently, with no standardized, repeatable criteria for pass/reject decisions.
- No offline, on-premises option: Any inspection tooling needed to run locally on the factory floor without dependency on constant cloud connectivity.
- Zero tolerance for false passes: A defective bottle marked as “passed” was a far more serious risk than a healthy bottle flagged for a second look.
Zelite Solutions partnered with All Time Plastic to design and build a two-module AI quality inspection system: one module for training and refining a custom object-detection model, and a second for running that model live against bottle images entirely offline.
- Custom-Trained Object Detection Model: Using Azure Custom Vision, an object-detection model was trained on real bottle imagery from multiple camera angles (Side and Top views) to recognize five defect types – flash, parting line, shrink, scratch, and silver streak – as well as defect-free bottles.
- Fully Offline Local Inference Engine: The trained model is exported to ONNX and run locally via ONNX Runtime, so every evaluation happens on-site with no dependency on live cloud calls.
- Three-Tier Verdict System: Rather than a simple pass/fail, a “Needs Review” tier catches borderline-confidence detections, so an uncertain result is never silently reported as a confident pass.
- Model Evaluation and Reporting: An evaluation pipeline scores batches of reference images against verified ground truth and generates visual HTML reports with bounding boxes and confidence scores.
What was once a manual, inspector-dependent process is now a consistent, auditable, AI-assisted quality gate running directly on the factory floor – with no dependency on a live internet connection.
The AI quality inspection system delivered clear improvements for All Time Plastic:
- Consistent Detection: Repeatable defect detection across shifts, removing variability between individual inspectors.
- Safety Net Against False Passes: Borderline-confidence detections are automatically routed to human review instead of being silently approved.
- Fully Offline Operation: No cloud dependency or added latency once the model is deployed to the local machine.
- Auditable Reports: Evaluation reports for every test batch give the quality team full visibility into what the model detected and why.
- Scalable Foundation: The same pipeline extends to new defect types, bottle SKUs, and camera angles as the line evolves.
All Time Plastic now has a repeatable system for catching defective bottles before they reach the packing line, reducing the risk of rejected shipments and customer complaints while cutting the manual burden on their quality team.
