INDUSTRIAL AI
Computer Vision Quality Inspection: What Works on a Real Production Line
Amey Kadle
3 February 2026
7 min read
Most computer vision pilots succeed and most rollouts fail. The reason is rarely the model — it is the imaging chain.
The Imaging Chain
- Camera (resolution, frame rate, sensor type) — the most under-specified link.
- Lens (focal length, aperture) — must be matched to the working distance.
- Lighting (intensity, colour, polarisation) — controls > 60% of the model’s accuracy.
- Mechanical fixturing — the part must present consistently.
- Trigger — PLC or photo-eye driven, not software-polled.
Spend twice as long on the imaging chain as you do on the model. The model can be retrained. The imaging chain decides whether the model has any signal to learn from.
Model Architecture
- Object detection (defect location): YOLO v8 / v10 variants.
- Classification (defect type): ResNet / EfficientNet.
- Anomaly detection (unknown defects): autoencoders, PatchCore.
- Multi-task heads where the same image needs both detection and classification.
Integration with MES
- Vision system receives the trigger and the part ID from MES.
- Inference runs at the edge, result returned in < 200 ms.
- Reject decision is published to the line controller (PLC).
- Reject reason and image are logged to MES quality module for genealogy.
- Continuous evaluation runs against a held-out labelled set every shift.
Practitioner note
A vision QA system that does not feed its outputs back into MES is just an expensive camera. The genealogy link is what makes the data compound.
Frequently asked
Do I need a GPU on the line?
For real-time inference at line speeds above 30 fps, yes. Jetson Orin or RTX A2000 typically suffice for most vision tasks.
Amey Kadle
Founder & CEO, Ajinkya Technologies. 20+ years of building MES, ERP and AI systems for India’s most demanding manufacturing plants.