Brief
A real-time visual quality inspection system running on edge devices on the production line. Custom-trained YOLOv8 models detect defects at 99.2% accuracy, replacing 12 manual QA inspectors and letting the line run around the clock.
99.2%
Accuracy
38 FPS
Edge Inference
0.94
mAP (from 0.71)
Problem
Visual QA was done by people, which capped the line at the hours those people worked and made defect rates a function of fatigue. Sending frames to the cloud was not an option at line speed.
Solution
- A custom dataset and augmentation pipeline that lifted mean average precision from 0.71 to 0.94 before any deployment work started.
- A quantised ONNX deployment achieving 38 FPS on an NVIDIA Jetson Orin Nano, so inference happens on the line rather than in a data centre.
- A cloud-synced retraining loop fed by flagged false positives, so the model improves from the cases it got wrong.
- A live operator dashboard with annotated defect visualisations, so the people on the floor can see what the model saw.
Result
Detection runs at 99.2% accuracy and 38 FPS directly on the line, replacing 12 manual inspection roles and letting the line run around the clock.
Under the hood
| Models | YOLOv8, custom dataset and augmentation pipeline |
|---|---|
| Training | PyTorch, OpenCV |
| Deployment | Quantised ONNX on NVIDIA Jetson Orin Nano |
| Feedback | Cloud-synced retraining loop from flagged false positives |
