AI ·
Shopfloor Defect Vision
Deployed an on-edge computer-vision inspector that flags surface defects on a moving production line at 60 FPS, replacing slow manual spot-checks.
Python OpenCV ONNX Runtime C++ Docker FastAPI
Problem
Manual QA sampled one in fifty units and still missed hairline defects. Cloud inference was a non-starter — the line could not wait on a round trip, and the factory floor had no reliable uplink.
Architecture
A classical OpenCV pre-processing stage normalizes lighting and isolates the part, then a quantized ONNX model classifies defects entirely on an edge box.
# Lighting varies across the shift; normalize before the model ever sees a frame.
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
norm = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8)).apply(gray)
defect = session.run(None, {"input": preprocess(norm)})[0]- Throughput: pipelined capture → preprocess → infer holds 60 FPS on a single GPU.
- Edge-first: runs in a sealed Docker container, no cloud dependency.
- Explainability: every reject is saved with a heatmap for line-lead review.
Impact
- Defect escape rate cut from ~3% to under 0.4%.
- 100% inspection coverage replaced 2% manual sampling.
- Line operators retrain the model on new defect types without my involvement.