I remember the first computer vision quality control project I led at a mid-sized automotive parts supplier. The pitch was simple: replace three human inspectors with a camera system that never blinks, never gets tired, and catches microscopic cracks in cast aluminum brackets. We installed high-resolution cameras over the conveyor, trained a model on about ten thousand labeled images, and flipped the switch on a Friday afternoon. By Monday morning, the line was stopped more than it was running.
The problem wasn't the model. It was the lighting. We had calibrated everything during a quiet Sunday when the factory was empty and the overhead fluorescents were steady. But on a Monday, with the loading dock doors open and sunlight angling across the inspection station, the camera saw shadows as surface defects. It rejected nearly forty percent of good parts. The plant manager called me at 7 a.m. and used language I won't repeat here.
That failure taught me that computer vision in manufacturing is less about the algorithm and more about controlling the environment. We ended up building a light-tight enclosure around the camera with its own LED panels set to a fixed color temperature. We also added a simple lux sensor that would pause the line if ambient light leaked in. The false rejection rate dropped to under one percent within a week.
The next surprise came from data drift. The model we trained on that first batch of brackets had learned to recognize a specific texture from one casting machine. When the plant added a second machine from a different vendor, the surface finish was slightly smoother. The vision system started flagging those parts as suspicious, not because they were defective, but because they didn't look like the training data. We had to retrain with a balanced dataset from both machines and, more importantly, set up a weekly review of false positives to catch drift before it became a line stoppage.
What really changed my thinking was watching an experienced inspector work alongside the system. She could spot a hairline crack by the way it caught the light at a certain angle, something our camera initially missed because we had positioned it straight down. We added a second camera at a low angle and suddenly the model's recall on cracks jumped. I learned that the best computer vision deployments don't replace human knowledge; they encode it and then ask the humans to keep challenging the system's calls.
Today the system catches defects at a rate the plant's old manual inspection never achieved, but it still isn't perfect. Every quarter we sit with the quality team and review the images the system got wrong. Some of those misses hurt. A batch of brackets with porosity just below the surface went out because the camera couldn't see what an ultrasonic test could. That prompted us to integrate the vision system with other inspection data instead of treating it as a standalone oracle.
The lesson I carry into every new vision project is deceptively simple: the camera is the easy part. The hard part is the lighting, the changing materials, the integration with MES, and the trust you have to earn from the people who will still be accountable when the system says a bad part is good. If you go into a computer vision rollout thinking it's a camera and a model, you'll spend your first Monday morning on the phone with an angry plant manager. I know I did.