Computer vision QC for a manufacturing line

Context

A manufacturer running a single-shift QC station with a senior inspector, which meant when she was on leave, defective product shipped. Goal: a camera + edge box that flagged the same defect classes she did, live on the line.

Problem

Off-the-shelf industrial CV is expensive and the SKU-set was a moving target. The client also didn’t want a cloud round-trip: the line was in a region with intermittent connectivity, and the rejection signal had to fire within ~150ms of a part hitting the inspection zone.

Approach

Build small. A YOLO-family detector fine-tuned per defect class beat any “general industrial” model the client had trialled, because defect patterns are highly product-specific. Run it on a Jetson box at the inspection station; only sync the labelled imagery and a daily metrics summary to the cloud.

Solution

• Labelled the first 4,000 frames by hand alongside the inspector, paying attention to the edge cases she flagged that an off-the-shelf model would miss.
• Fine-tuned a YOLOv8 variant; converted to ONNX for Jetson inference.
• Built a thin FastAPI service that took camera frames, ran inference, and published rejection signals over MQTT to the line’s existing rejector arm.
• Daily nightly retrain trigger: when the inspector overrides the model 3+ times in a shift, the new images get added to the training set and the model rebuilds at midnight.

Outcome

Above. The “accept the inspector’s overrides as ground truth” loop is what kept accuracy climbing instead of plateauing.