The question.
Inspection models must work under changing lighting, device limits, camera movement, and rare defects rather than only curated benchmark images.
The design covers the complete perception loop: capture, inference, review, labeling feedback, and controlled model updates.
The proposed approach.
- Define defect classes and ambiguous cases with operators
- Compare supervised and anomaly-detection approaches
- Profile inference on the intended hardware and image resolution
- Feed reviewed cases into a versioned annotation and evaluation loop
Target architecture.
Camera capturePreprocessingEdge inferenceUncertainty / alertsOperator reviewDataset feedback
The difficult parts.
- Rare defects and imbalanced labels
- Latency and memory limits on devices
- Changing capture conditions and false alarms
How to evaluate it.
- Recall and false alarms by defect and capture condition
- Latency, memory use, and thermal behavior on target hardware
- Detection stability as cameras and environments change
Planned deliverables.
- An inspection and annotation specification
- An edge-deployment reference architecture
- A drift and review feedback plan
Design principles.
Define ambiguous visual cases before choosing a model.
Profile the complete capture-to-alert loop.