Creation of multiple geometric template

Question

i have various templates and want to create a multiple geometric template file
i am unable to do it.
i have uploaded the templates
please help me with it
-$hiv@
Solved!
Go to Solution.
Attachments:
five(OK).png ‏342 KB
five3(OK).png ‏269 KB
five4(OK).png ‏1694 KB

Answer 1

Both geometric matching techniques rely on curves extracted from image to perform the matching. The two geometric matching techniques differ in how the curve information is used to perform the matching. The edge-based geometric matching method computes the gradient value of the edge at each point along the curves found in the image and uses the gradient value and the position of the point from the center of the template to perform the matching. The feature-based geometric matching method extracts geometric features from the curves and uses these geometric features to perform the matching.
You can find more detailed information in the NI Vision Concepts Manual here:
http://zone.ni.com/reference/en-XX/help/372916L-01/nivisionconcepts/geometric_matching_technique/
The article also explain when it is better to use one versus the other.
I couldn't test the templates yesterday, because you did not attach a match image.
I might be wrong, but another thing I noticed is that it looks like you thresholded the images prior to extracting the templates. In general, algorithms that need to extract edges from images, such as geometric matching, will work better when you provide them with the grayscale image, as opposed with a thresholded image. The contour extraction will be more accurate. Also, the algorithm does not only use contour information, but also grayscale information of the image. By thresholding the image, one removes a lot of information that the algorithm can use.
On a last note, combining templates into one is only going to help with speed. If speed is not an issue in your application, you can still learn with Edge Based matching and match all template. If there is a match, the character/number you looked for is the right one.
Another solution might be to use a reverse technique, still using Edge Based matching that seems to work better in your case. We sometimes use it to solve classification applications similar to what you are trying to do:
Instead of learning multiple templates and trying to locate them in your image, take the opposite approach: create a synthetic image that contain your templates (all rotated 1s on one line... all rotated 5s on another line etc). Then if you can locate and extract the number (using blob analysis) on the original image, you can extract that region, learn it as a template, and look for it in the synthetic image. The line you find it indicates which digit it is.
Again, this technique might work if you can extract the digit you're trying to identify from the original image. Performance wise, it is only interesting if learning the extracted digit on the fly is faster than trying to match all the separate templates (learning takes more time than matching).
It looks like you're trying to perform OCR on the image. Again, if the text is always located at the same place and you can draw a region of interest around it, you can try the OCR algorithm and define different classes that have the same label for the different rotated digits.
I tried to give you different possible approaches. Hope that was not too confusing.
Best regards,
-Christophe