More specifically, how do you feel about sub-pixel interpolation?
My question is prompted by another article from Control Engineering (September 2009,) which is fast becoming my favorite place for machine vision application stories. This concerns an intriguing story of a high precision measurement system.
The magazine describes a system from integrator DWFritz that uses two “high-resolution Cognex smart cameras,” to deliver, “repeatability better than two microns.”
That’s a pretty bold claim, and I think it has to be based on sub-pixel interpolation. I say this because the high-res Cognex cameras provide ‘only’ 1600 by 1200 resolution, (Natik – please correct me if I’m wrong.) If the system is set up to give one micron per pixel, the field of view will only be 1.6 x 1.2 mm, and based on the picture in the print magazine (not available on-line,) the part under inspection seems to be an order of magnitude greater.
Now I know that Cognex have some powerful sub-pixel tools in their armory, but when all is said and done, that’s just using math to calculate the probable location of an edge. So can they give two micron repeatability? I’d sure like to see some data.
Subscribe to:
Post Comments (Atom)
Posts
4 comments:
I find sub-pixel repeatability difficult as it can depend on many parameters. When evaluating a potential application I am generally not comfortable claiming sub-pixel repeatability. After sufficient testing with the final configuration and algorithms, sub-pixel repeatability can be verified but unfortunately this doesn't usually occur until late in the development.
Assuming DW Fritz used In-Sight cameras, you are correct about the resolution.
If I can get data on the two micron repeatability, I'll post it here.
The two cameras used on this DWFritz Automation project were both Cognex CDC-200 CMOS digital cameras with 1280 X 1024, 6.7 micron pixels.
One camera has a 1.0X lens and the other camera has a 0.5X lens. For the 1.0X lens, each pixel in the camera is imaging a 6.7 micron square portion of the object, with a field of view of approximately 8.58 X 6.86 mm.
These Cognex cameras provided excellent results. Each pixel's analog signal is converted to digital within the camera and then sent to the frame grabber over a digital interface. This avoids the problem of "pixel jitter" sometimes encountered with analog cameras and conventional frame grabbers.
The Cognex software is doing sub-pixel interpolation, but they don't call it that. In our experience, the software easily achieves 1/5 pixel repeatability with most measurements.
The digital cameras produce better images than analog cameras. Because 1/5 pixel corresponds to 1.34 micron (1.0X lens), it is no surprise that the machine is capable of better than 2 micron repeatability on its measurements.
We did extensive Gage R&R testing with this machine and documented the "better than 2 micron" repeatability results in order to achieve buyoff on the machine from our client.
I'd like to thank John at DWFritz for taking the time to provide more detail on the sub-pixel interpolation issue I raised. From the facts provided, it would seem that Cognex can indeed do an excellent job when such precision is needed, although this may well also be a tribute to the expertise of the people at DWFritz.
Thanks also to 'passion' for helping get more details. Openness and transparency are always good.
Post a Comment