Knowing when it’s a bin job

In Britain, when they say something’s a “bin job” they mean it’s fit only for the bin, or trash, to use the American word. Thus “binning” means you’re throwing something in the trash.

In machine vision “binning” has an entirely different meaning. “Binning” in this context means adding together the charges from adjacent pixels. Thus a 2x2 bin means adding pixels pairs both horizontally and vertically. In the case of a 640 x 480 image, it becomes 320 x 240.

Now since we usually want more pixels rather than fewer, you might wonder why anyone would want to bin. Well, writing on the Adimec blog, Gretchen Alper has an explanation for you.

It all gets a bit technical, so I’m going to jump over the details and skip to the part where I say, “check out the Adimec blog. It’s really good.”

How to acquire spectral data

When you think about it, monochrome cameras are pretty crude devices. All they do is collect photons, with no regard for the wavelength of each one. Wouldn’t it be cool if a camera could sort the incoming photons into wavelength “buckets”?

Of course, that’s what a color camera does, using Red, Green, and Blue filters to ensure only a narrow range of wavelengths reach each pixel. The software has to figure out exactly what color was seen, interpolating between what was captured.

Hyperspectral imaging takes this further, enabling wavelength “slices” to be captured and analyzed. The concept has been around for some years, but to date the hyperspectral imagers I’ve seen have all been rather clunky, not to mention, expensive.

Imec and Adimec are attempting to change that, as described in these two posts on the Adimec blog, “Imec demonstrates a compact, low cost, fast hyperspectral imaging solution, part 1” and the follow up “Part 2”. There’s also a link to video demo, but you’ll have to read the Adimec posts to get to that.

I said before that color imaging is really hard. (“Why color machine vision is really difficult”) and it might seem that more spectral “channels” will make it harder, but I think the opposite is true. With less need to interpolate, we’ll be better able to zero in on the specific color we want, always assuming we can get the lighting right.

How to choose a camera

There’s lots of advice online about sensors and pixel sizes and so on, but one of the best tips I’ve seen in a while is on the Adimec blog.

The accurately-but-not-very-concisely-titled, “Evaluating Machine Vision Cameras versus Comparing Camera Specification Documents” talks about measuring the photon transfer curve, which sounds like hard work, but what it boils down to is this: choose the camera that works best in your application.

Easy to say, I know, but what I believe Adimec are getting at is that the camera is just part of a system. That system includes the lighting, the enclosure, the lens, and of course, the object you’re looking at. So you put the camera in that environment and find out how well it works.

Now as a good engineer, you want to evaluate performance on the basis of measurable criteria. But I think many machine vision people, (myself included,) have a tendency to look at a pair of images from two different cameras and pick the one that “looks” best. And this is where the photon transfer curve comes in to play.

I’m not going to regurgitate all the info about this that’s already on the Adimec website – use the link above and you’ll find it all – but I do urge you to learn more and figure out how to use it. Or if you want to go to the source, look through my list of “Friends” and go to Harvest Imaging

Maybe it is more work, but it gives you data as a basis for selecting your camera. Or just go with the vendor that offers the best freebies.