Wednesday, April 17, 2013

Balancing false rejects and false accepts

It would be very easy to eliminate false rejects: just open up the inspection criteria so that everything passes.

Of course, you’ll probably ship a bunch of bad stuff, but Production will be happy: they’ll hit their numbers and the yields will be very good.

The reality though is that you, the poor vision engineer, gets beat up every day about the vision system kicking out good items. Yet you can’t ease off on the inspections because then bad stuff goes out the door. What to do?

I’ve been wrestling with this for a while, trying to get statistics to help me find a balance point. It’s complicated though, and I’m no numbers guy, so a while back I turned to master statistician Nate Silver for help.

Okay, I didn’t actually reach out to Nate, but I did read his book, “The Signal and the Noise”. This deals with how Bayesian probability can offer a better way of dealing with real-life problems, and I think it has much to offer the machine vision world.

How so? Well there’s a good example in a blog post by Eliezer S. Yudkowsky titled “An Intuitive Explanation of Bayes' Theorem.” This concerns instances of false accepts in mammography screening. I won’t repeat it all here, but let’s just say I think this offers some guidance for how to ‘dial-in’ or vision systems.

I don’t have it all figured out – maybe I never will – but as I work through I’ll keep you updated. And as always, if you have anything to add, please use the Comment function.

No comments: