What I Learned When We Replaced Our Inspection Team With Three Cameras

Three years ago, I walked a fabrication shop in northwest Indiana with a plant manager who had four full-time inspectors working two shifts. Their job was simple: look at stamped brackets coming off a progressive die at roughly 45 parts per minute, flag geometry and surface defects, and sort scrap from rework from good parts. The inspectors caught about 87 percent of actual defects. The rest shipped.

The cost was not just labor. It was throughput. Parts stacked up on the inspection bench. Cycle time bloated. The line ran at 35 pieces per minute because the downstream constraint was not the die but the eyeballs. I suggested they look at inline vision systems. They were skeptical. Every vision vendor they had talked to promised 99.9 percent accuracy and delivered a system that either flagged every part with a minor cosmetic blemish or missed the stuff that actually mattered.

What changed was not the cameras. Cameras have been good enough for a decade. What changed was how we trained the models.

Here is what we actually did: We integrated three GigE cameras mounted above the press discharge, angled to capture top and side geometry. We collected 8,000 labeled images of known defects: burrs, dimensional drift, surface gouges, incomplete forms. No synthetic data. Real parts. Real defects. We trained a YOLOv8 detector on an NVIDIA Jetson AGX Orin running inference at 22 milliseconds per frame. That was fast enough to keep pace with the line and output a pass/fail decision before the part reached the sort chute.

Accuracy looked like this: precision at 94 percent, recall at 91 percent, F1 score of 0.92. Not theoretical. Not on a validation set in a lab. On production parts, shift after shift, under fluorescent lighting and conveyor vibration. We missed some defects. We false-flagged some good parts. But the miss rate dropped from 13 percent to 9 percent in the first month. Within six months, to 6 percent.

The real surprise came in week two. The system caught a class of defects nobody had been looking for systematically: micro-cracks in the formed web, barely visible to the human eye, that would fail in fatigue testing downstream at the customer. Our inspectors had missed them because they were trained to look for obvious geometry problems and surface finish. The machine learned the data, not the habit.

Here is the part that matters to your bottom line: we removed two inspectors. Moved one to a quality engineering role. Kept one as a secondary validation because automated systems fail in ways humans do not and vice versa. Throughput on the line jumped to 44 pieces per minute. Cycle time dropped eight percent. Scrap reduction from the improved defect catch rate was worth roughly $120,000 in the first year. The system paid for itself in less than ten months.

What did not work: trying to detect cosmetic defects below 0.3 millimeter. The cameras had enough resolution but the lighting setup was brutal. We solved it with custom LED backlighting and a second camera on a different spectrum. What did not work: generic pre-trained models. Every off-the-shelf detector we tried required so much tuning on our specific part geometry that we trained from scratch. Time cost but accuracy was night and day different.

If you are running manual inspection, ask yourself: at what defect rate is my customer going to notice? What is the cost per escaped defect? How much of my line capacity is sitting in the inspection queue? Then get samples of your actual parts, your actual defects, in your actual lighting. Talk to a systems integrator who has done this in your process family, not someone selling cameras. And be honest about what your inspectors can actually catch in eight hours of repetitive work.

The technology is ready. The question now is whether your operation is structured to use it.