Computer vision AOI pipeline for automated production-line quality control
Defect and deviation detection from industrial camera streams: real-time inference, calibration, batch traceability and MES/SCADA integration.
Delivered functionality
Zone-level defect classification/localization; rules: scrap / rework / rerun.
Operator UI: borderline adjudication, active-learning labels, shift reports.
Model monitoring: scores, FP/FN alerts, A/B weights.
Business outcomes
Missed defects −40–60% vs sampling (pilot); false rejects −15–25%.
Evidence frame lookup per batch: hours → minutes.
Highlights
- Defect detection/segmentation (YOLO / mask models) + ROI-aware post-processing for the line.
- Inference on GPU servers and/or edge (ONNX Runtime, TensorRT) with per-frame latency SLAs.
- Evidence image storage, labeling and active retraining; ML/data drift monitoring.
- MES integration: reject signals, downtime reasons, FP/FN dashboards and measurement repeatability.
Stack
Python 3.12
PyTorch
OpenCV
ONNX / TensorRT
FastAPI (inference gateway)
Redis + Celery
PostgreSQL
MinIO (кадры, артефакты)
MLflow
Prometheus + Grafana
React + TypeScript (операторская панель)
Want a similar architecture for you ?