Computer Vision on the Factory Floor: Where It Earns Its Keep

A camera and a model can now catch a defect that a human inspector misses at the end of an eight-hour shift. That part is genuinely new. What has not changed is that the value comes from the process around the camera, not from the model inside it.

Pick a defect you already measure

If nobody can tell you the current reject rate, there is nothing to improve against. Start where there is already a tally sheet: a surface flaw, a missing component, a label in the wrong place, a fill level. That tally sheet is your training data and your proof at the same time.

• The defect is visible in a photograph a person can judge
• It happens often enough to collect a few hundred examples
• Catching it earlier saves rework, scrap or a customer complaint
• The line can act on the result — stop, divert or flag

Lighting beats model choice

Most failed vision projects fail on physics: a reflective surface, a part that moves faster than the shutter, a shadow that changes with the time of day. Fixing the lighting and the camera mount improves accuracy more than any change of architecture, and costs a fraction as much.

If a person cannot judge the defect from the photograph, no model will.

Run it beside the inspector first

For the first weeks the system watches and the inspector decides. You compare the two, find the cases it gets wrong and retrain. Only once it agrees with your best inspector does it get to stop the line. This also answers the question the floor team is actually asking, which is whether the camera is there to replace them.

Safety and counting are the two other places vision pays quickly: protective equipment at a gate, and counting what leaves the yard. Both are unglamorous, both are measurable, and both are usually being done on paper today.