I remember the first time we put a camera above a bottling line, convinced it would catch every cracked cap before shipping. It caught almost all of them. The problem was it also flagged perfectly good bottles whenever the conveyor belt wobbled or the lighting shifted at 2 p.m. when the sun hit the skylight. We spent six weeks chasing false positives before I admitted the model wasn't the issue, the environment was.
That's the part nobody tells you about computer vision for quality control. The algorithms, even the off-the-shelf ones, are good enough these days. You can train a model to spot scratches, dents, missing labels, misaligned components. But a camera doesn't see like a person does. It sees exactly what you give it, and if you give it inconsistent lighting, vibration, or dust on the lens, it will faithfully report all of that as defects.
Our first deployment was a classic overreach. We tried to inspect too many things at once, surface finish, cap torque, label placement, and fill level. Every one of those needed a different angle, a different exposure, a different tolerance for what counted as bad. We ended up with a system that was technically accurate and operationally useless because the line operators stopped trusting it. They'd override the alarms or just turn the monitor away.
What actually worked was narrowing the scope. We picked the one defect that was causing the most returns, a hairline crack near the bottle neck, and focused everything on seeing that reliably. We added a polarized filter to kill the glare. We bolted the camera to a separate frame instead of the conveyor so it wouldn't shake. We took thousands of images across three shifts, including the graveyard shift when the plant lights dimmed. Only then did the false alarm rate drop below something people could live with.
There was a moment I still think about. We had a worker named Maria who had been inspecting bottles by hand for eleven years. She could spot a flaw from across the room. When we first showed her the camera's output, she laughed and said it would never replace her eyes. Six months later, she was the one who taught us to aim the lens at the spot where her gaze naturally landed. She wasn't wrong, and she wasn't replaced. Her job changed from staring at bottles all day to managing the exceptions the camera flagged.
That's the real shift. Computer vision in manufacturing isn't about removing people. It's about catching the repetitive, soul-crushing inspection work so people can handle the weird stuff that doesn't fit a pattern. And the weird stuff always shows up sooner or later.
If I were starting over, I'd spend less time tuning hyperparameters and more time standing next to the person who knows the line better than anyone. The camera is just a tool. The quality standard is a human decision.