Machine vision has a way of making grown engineers want to run for the hills. I spent 20 years selling and supporting it, so believe you me, I get it. A lot of that pain comes from two mistakes made before anybody ever writes a line of inspection logic: using vision where a simpler sensor would've done the job, and treating lighting like an afterthought.
Start with the first one. Before you price out a camera, ask what you're really checking. Is a part there or not? That's a photo eye. Is the cap on? Often a photo eye or a proximity sensor. Is the label the right color, or is the registration mark there? A color or contrast sensor. Is it the right height? A laser distance sensor. Reading a barcode? A dedicated barcode reader beats a general-purpose camera nine times out of ten.
Those sensors cost a fraction of a vision system, set up in minutes, and there's nothing to tune when the lighting in the building changes. If one simple sensor answers your question, use it and go get a coffee.
Vision earns its keep when the job really needs it: several checks on the same part at once, parts that show up in a different spot or orientation every time, actual measurements, reading printed text, or defects that come in all shapes and sizes. That's when a camera pays for itself.
Now the second mistake. In vision, lighting is most of the battle. The software can only judge what's in the picture, and the light decides what's in the picture. Get the lighting right and the feature you care about jumps out, bright on dark or dark on bright, and the inspection logic gets easy. Get it wrong and you'll spend weeks fiddling with thresholds, chasing glare and shadows, and still reject good parts every time a cloud goes over the skylight. Different lights make different features show up, and the right light depends on what you're looking for. That's a whole topic on its own, so there's a separate guide on lighting for defect detection.
A couple more things I'd do before buying anything. Collect real parts: plenty of good ones, and as many bad ones as you can find, including the weird ugly rejects that only show up once a month. Then test with those parts, the lighting you plan to use and a camera, ideally in conditions close to the line. A feasibility test on real parts tells you more in an afternoon than a month of spec sheets. That's actually why I built Angora Vision: you can try your own parts there and get a feel for whether the inspection is doable before anybody spends a dime on hardware.
Also look hard at how the part shows up in front of the camera. If it's wobbling, rotated or a different distance away every time, the vision system has to work a lot harder. Fixing part presentation with a guide rail or a nest is often cheaper than making the software smarter.
And the AI question, since everybody asks. Deep-learning inspection tools are real, and they're genuinely good at the messy stuff: defects that vary a lot, natural materials, cosmetic flaws that are hard to describe with rules. But they still need good images and a healthy pile of example pictures to learn from, and you still have to prove they work on your parts. Bad lighting doesn't get better just because the software got smarter. Garbage in, garbage out...just with more GPUs.
Do the cheap checks first, sort out the light, test on real parts, and then pick the camera. In that order.