AutoVisionInspector
Edge-deployed automated visual quality inspection for manufacturing defect detection
AutoVisionInspector is an automated visual quality inspection pipeline designed for deployment directly on edge hardware — achieving near-cloud accuracy with sub-10 ms inference latency, making real-time defect detection feasible on factory floors without cloud connectivity.
Background
Visual inspection of manufactured goods (packaging integrity, surface defects, fill-level checking) is traditionally done by human inspectors or expensive vision systems costing $50,000+. Modern deep learning makes it possible to build accurate vision models, but cloud inference introduces unacceptable latency (>100 ms) and connectivity dependency for high-speed production lines.
What We Built
- A YOLOv8-based defect detection model fine-tuned on a custom dataset of packaging defects (incorrect labels, dents, seal failures) from a Nigerian food production facility
- TensorRT optimisation pipeline: FP16 precision conversion reduces model size by 2× and improves throughput 3.8× on Jetson Orin compared to native PyTorch
- Streaming inference server accepting RTSP camera feeds, processing at 120 FPS on Jetson AGX Orin
- Integration with OPC-UA industrial protocol for direct communication with PLC reject actuators
Performance
| Metric | Cloud (GPU server) | Edge (Jetson Orin) |
|---|---|---|
| Inference latency | 8 ms | 9.2 ms |
| Throughput | 125 FPS | 108 FPS |
| Accuracy (mAP@0.5) | 91.3% | 90.8% |
Accuracy degradation of <0.5% with full independence from cloud infrastructure.
Links
See (Igwegbe et al., 2025) for the full technical report.