The Thursday our little computer vision model spotted a cracked gear before I did
So I help out at this small machine shop run by my uncle in Akron, and for a while we were just eyeballing parts as they came off the line, which meant stuff slipped through. Two years back I started messing with a basic image model on a spare laptop, just training it on phone photos of good gears and bad ones. Nothing fancy, maybe 400 pictures to start. On a Thursday toward the end of a busy run, it flagged a hairline crack on a gear that I had already looked at and passed. I pulled it under the light and sure enough, there it was, right on the edge of the tooth. We had been shipping those out for who knows how long, and my uncle just stood there quiet for a second and then asked how fast we could point a camera at the whole line. Now we catch maybe 5 or 6 bad ones a shift that we used to miss. Curious if anyone here has set up a simple vision check on a small budget, and what got you the most gains first, more training photos or better lighting?