Help me please! : Machine Vision system resolution

[Purmpol]

Now I'm building the machine vision system for inspecting watch face by gauging features on it. The acceptable tolerance is 30um.
First, I contacted to vendor and they recommended the system which cliamed to achieve 25um-measurement-resolution. This system is composed of :
- Camera: IEEE 1394, PixeLINK PL-A776, 3.0 Megapixel CMOS Sensor (2048 x 1536)
- Lens:MVO 0.125 PMAG, Telecentric
- System FOV = 50x39mm

In practical, I tried to measure the features on the watch face. The results are,
Measuring by metallurgical microscope, the value is 0~5um different from the dimensions in the drawing.
Measuring by these machine vision system, the value is 0~80um !!!

So, I wonder wether or not my machine vision system can truly achieve 25um.
Anyone please help me about this, I have to complete my job soon.

If you need further information, please tell me.
Thank you in advance.

Purmpol
Project Manager
Digital Associates Co.,Ltd.

[BruceAmmons]

Each pixel in your system is about 25 um, which you can get by dividing the FOV by the camera resolution.  This matches the resolution advertised, but the advertising can be misleading.  Resolution does not guarantee the same level of accuracy.  Accuracy will depend on the lens quality, focus, and other factors.

The general rule is that your measurement capability is equal to two pixels, so I would predict an accuracy of about 50 um.  This can be improved by processing and multiple measurements, but you won't be able to get much better than 25 um accuracy.  It really depends on the software and algorithms used at this point.

Bruce

[Purmpol]

Hi Bruce
Ok, I see the problem and I feel familiar with your calculation. It's real.
Thanks!

Purmpol

[Purmpol]

Another quick question.

After finding out the configurations, we are currently working on how to get the accuracy down to 25-30 microns, but am having some trouble getting it there.

What are the multiple measurement and other methods that can accomplish the 25 micron?  Thanks!

[AnalogKid2DigitalMan]

Purmpol:

Ideally, your resolution should be 1/10th of the tolerance of your specification of 30um, which is 3 microns. In general practice, sometimes a factor of 1/5th can be acceptable which would require a resolution of 6 microns. The only way I can suggest to attain the required resolution is to increase your magnification via a more powerful lens by a factor of 5 to 10. The effect is a smaller field of view. If the features of interest still fit in the reduced field of view (10x8mm at 5x more or 1x0.8mm at 10x more) then the lens solution should be viable. If the features of interest are not within the field of view, then the approach typically taken is to use motorized stages with encoders to move the object around the field of view. Measurements are derived by measuring stage positions from the encoders in conjunction with coordinates of features located by the camera. The encoders will add an additional source of error to the measurements.

Also note that higher power lenses have a reduced depth of focus, so if the watch face is curved, some features may be out of focus and require refocusing between image acquisitions.

Some improvement in resolution, repeatability, and accuracy can also be attained by acquiring a lot of pixels on the features of interest and using averaging to attain sub-pixel resolution.

Can you post some images that you have acquired with your current setup?

-AK2DM

[BruceAmmons]

The software you are using may not be capable of the accuracy you need.

One way to get better accuracy is using sub pixel interpolation.  For example, when searching for an edge, you can estimate the position of the edge within a pixel based on the intensity of the pixel and its neighbors.  Your software has to have this capability built in.

Another method is averaging.  Measure the same thing several times and average the values.  You would probably need to shift the part enough to get different measurements each time.

Another method is multiple measurements at different locations.  For example, if you are trying to measure the width of a bar, measure the width in several positions along the bar and average them.

Another method is finding lines.  If you can find the edges along a straight edge, you can fit a line to these edges.  Once you have a straight line fit to each side of the bar, you can get a measurement anywhere along the bar.

You can also fit other geometries, such as a circle.  If you get enough points on the circle you can fit a circle pretty accurately.

Bruce