Troubles with Geometric Pattern Machine

[altelec]

I need to analyze flat paper form, back lighted images as attached, to:

1] Correct position camera, if image are clipping or partially occluded
2] Compare with a master template for geometric matching, and reject it, if there is any shape

mismatching focusing en each of four corners.
3] Get accurate X and Y coordinates displacement and T angle rotation, most important objective in

process.

I’m using VA 7.1.1 I Geometric Matching because I don't care gray tones and create five templates one

for the whole image and one for each corner.
I think to develop this project with LabVIEW 7.1

Geometric matching work fine for whole image but I find some problems finding corners working with same

“ideal” version of rotate and /or shifted image.
1] Doesn’t find one or more corners at all. Nor lowering score, to very low setting.( 400)
2] In some cases are necessary swapp modes on extraction mode and/or edge filter size settings, to

get recognition or when recognize it, to get correct rotation angle o shift coordinates. I can’t find a

general rule to know how setting correct parameters. A general rule is better to set Row and Column

serch step size to 1-5 but processing come very slow
3] Acute angle as bottom left corner seem to be more sensitive to rotations. Changes in curve setting,

seemingly, can’t improve results.
4] Worst, with more “permissive” setting, some flaw was detected as good with score setting over 800 !.

In real world, I will work with a variable background / target contrast and illumination gradients

relationships, making thinks worst.

I read Geometric Matching in Vision Concepts, all help in relation with; analyze the only one example

of geometric matching I found

What I’m doing wrong?

Mensaje editado por altelec

Mensaje editado por altelec

[mfisher]

Here is what I would do:

1) Use Geometric Matching to get the position and angle of the target.

2) Now that you know the position and angle of the target, you can infer where you expect to find the "corners".  One easy way to do this is to set up a coordinate system based on the geometric matching step.  You can either use Vision Builder or LabVIEW to do this (but it's not a feature of Vision Assistant).

3) To inspect the corners, do a post processing step about a small region of interest where you expect to find them.  I used a combination of Find Straight Edge and geometric measurements.  I'm attaching some screenshots from Vision Builder.

Matthew Fisher

IMAQ Vision R&D

[altelec]

Hello Matthew
Is interesting your solutions, as flaw detector, with precise measurement of parameter errors.  But, as I need to work with a large number of similar patterns with different distances and relationships between flanges, is important to me to avoid multiples ROI definitions.  In few words, minimize manual interaction when a user beginning to work with a new pattern.

I’m trying to get a solution, using Geometric PM, because I need to detect ALL kind of flaws, i.e. missing “V” cut between flanges. I see GPM as a very powerful tool in Machine Vision, in applications like these, but I think to full exploit it, I need even, more examples and detailed info.

I’m attaching a zip with
a) A script in VA 7.1.1
b) Templates
c) More images that work well with my script with the exception of Desp Izq 100 Up 100.jpg and CENT1quad4ocluded 50%.jpg

My questions;
1) What do you think, why doesn’t work c)
2)  How can I refine parameters setting for this application, to minimize false rejects or pass fail ones (Please, re-read my originals questions).
3) You can load my “real “images from Processing variable gray image, post. Do you think that still can I use GPM? I get poor results using those real images, with a script like the attached one, of course, changing templates and trying with different parameters settings.

By the way, I'm intrigued with some aspects of your Screenshots:  can you send me your Create VI Code output for LabVIEW 7.1?

Thank You for your help.

Mensaje editado por altelec

Mensaje editado por altelec

[mfisher]

We should talk in more detail.  Post your email address here and I will get back with you.

Matthew Fisher

IMAQ Vision R&D

[altelec]

Hello Matthew

My email address

altelec@bigfoot.com

Thank you