> ML_RECIPE // AUTOMATED-OPTICAL-INSPECTION-SURFACE-DEFECTS_v1.0
Automated Optical Inspection (AOI) for Surface Defects
Segment and classify microscopic scratches, dents, and voids on manufactured PCB and metallic components at 60 parts per minute.
Business Outcome
Segment and classify microscopic scratches, dents, and voids on manufactured PCB and metallic components at 60 parts per minute.
Zero critical defect escapes (thermal cracked joints) allowed on quality assurance acceptance run.
Heuristic Baseline
Computer vision morphological filter: Canny edge detector + contour thresholding flagging surface blobs.
Canny edge filtering yielded 38% false rejection rate due to harmless surface oil reflections.
Phase 1: Prototype Path
Collect 2,000 labelled camera images. Fine-tune a lightweight feature extractor on developer GPU workstation. Evaluate mAP@50 and inference latency.
Phase 2: Production Path
Export trained weights to TensorRT engine. Deploy to industrial NVIDIA Jetson Orin edge appliance connected to GigE machine vision camera.
Compute & Placement Topologies
Quarterly training on developer/cloud GPU workstation (e.g. 1x RTX 4090 or A10G)
Industrial edge appliance (NVIDIA Jetson Orin or industrial IPC with TensorRT)
3-Plan Placement Alternatives
OpenCV thresholding and contour analysis on edge IPC CPU without deep learning weights.
Industrial edge PC with NVIDIA RTX 4060 or Jetson Orin running ONNX Runtime / TensorRT.
Pinned edge appliance with hardware watchdog + dual GigE cameras + TensorRT engine (< 15ms frame latency) + air-gapped model artifact updates via signed registry.
Recommended Libraries & Tools
Governance, Safeguards & Risks
- Licensing compliance notice: When employing YOLOv8 models, commercial use may occur under AGPL-3.0 subject to its obligations (including open-sourcing the derivative software under AGPL-3.0 upon network deployment); a separate Enterprise licence is required for proprietary/non-AGPL deployment.
- Artifact security: Enforce SafeTensors or verified PyTorch state_dict checkpoints with cryptographic SHA-256 validation; strictly prohibit unpickling untrusted model files on the factory IPC.
- Operator review: Route all rejected parts into a secondary physical bin with synchronized high-resolution image logs for manual operator confirmation.
- Optical calibration: Validate illumination stability with daily calibration target cards before operating shifts.