Computer VisionShipped★ Featured
Automated QC on Shoe Production Line
QC Tự động trên Dây chuyền Giày
Camera + YOLOv8 detect stitch & sole defects on shoe line — 12 fps, accuracy 0.96.
YOLOv8TensorRTJetson OrinOpenCVPythonBaslerPyQt
Problem
A shoe assembly line outputs 1,200 pairs/shift and needs 12 QC inspectors staring at every pair. Fatigue → missed defects → EU customer returns. Defect leak rate ~1.4% — max acceptable 0.3%.
Architecture
Basler 5MP cameras × 2 (top + side) + white dome light → Jetson Orin Nano 8 GB → fine-tuned YOLOv8m (18k images, 6 defect classes: open stitch, crooked stitch, glue leak, dented sole, fabric stain, wrong size) → TensorRT INT8 → MQTT alert to HMI panel + relay tower light.
Stack & rationale
- Basler 5MP: global shutter, no blur when the line runs at 1.2 m/s.
- Jetson Orin Nano: 40 TOPS, runs 2 streams of YOLOv8m + tracker.
- TensorRT INT8: 22 fps FP16 → 38 fps INT8, <1% accuracy drop.
Results
| Metric | Before | After |
|---|---|---|
| Defect leak rate | 1.4% | 0.18% (−87%) |
| QC headcount/shift | 12 | 5 (only confirm flagged items) |
| Line throughput | 1,200 pairs/shift | 1,450 pairs/shift |
| EU customer returns | 1.8% | 0.4% |
Lessons
Dataset diversity beats a bigger model: 18k varied images (angle, lighting, fabric lot) crush 50k uniform ones. White dome light > LED bars (less reflection on glossy shoes).