Technical areas to assess – lighting

Continuing our discussion of how to interrogate someone who wants to build you a vision system:
Every machine vision application begins by designing an effective lighting system. Have prospective integrators show you how they would light your parts, and make sure you see actual images for all types of defect as well as good parts. (Good parts can vary yet still be acceptable.)

When you get the images, you want to see that good stuff is light and bad stuff is dark (or vice versa.) Ask for the actual grayscale values for the good surface and the bad. In a typical vision system there are 256 levels of gray, from 0 (totally black) to 255 (totally white, also called 'saturated'.)

You want to see good differentiation between the good and bad. 10 gray levels is not enough but 110 is good. Much over 200 is probably not good, the reason being that image analysis tools tend to use rates of change to determine the location of edges. (If the pixel values in a row go 25, 27, 26, 36, 58, 90, 110, 146, 147, 149 the software can work out where the edge is. If the values are 1, 2, 3, 250, 251, 249 it's less reliable.)

Don’t worry if the pictures you’re shown look ‘odd’ – these aren’t for publication in a glossy magazine; they’re for a computer to analyze, and as computers are dumb, good lighting design is essential to make the task as easy as possible.