Computer Vision in Manufacturing: A Real Case Study

Manufacturing QA used to mean a person standing on a line for eight hours looking for surface defects. It's hard work, it's inconsistent across operators, and at speed it misses things. Here's how we replaced that process for a UK industrial customer using YOLOv8 and a small inference box that lives next to the conveyor.

The Setup Cameras above the line captured each unit. Lighting was industrial and imperfect. We trained a YOLOv8 detector on labeled defect classes (scratch, dent, missing fastener) with heavy augmentation for motion blur and specular highlights.

Edge Inference Cloud round-trips were too slow and risky for plant networks. We deployed ONNX Runtime on an industrial PC next to the line. Decisions came in under 40ms so reject actuators could fire before packing.

Results Defect escape rate dropped and QA labor cost fell about 67% on the covered lines. Operators moved from continuous visual inspection to exception review — a better job and a clearer audit trail.

Lessons Computer vision for manufacturing fails when teams ignore lighting, camera vibration, and labeled edge cases. Budget for data collection days on-site, not just model training in the cloud.

If your plant has a high-volume inspection bottleneck, MindVersa can assess feasibility from a short video sample and one week of labels.