0

Imagine a simple drawing of let's say a duck. And you have an photo of it in your database.

Now a user takes a second photo of the same drawing of the duck. What approaches exist to check if both photos show the same drawing?

Is it possible to extract the vectors and compare them afterwards? I'm completely new to this topic so I already have trouble searching for it on the web.

Improve this question
5

1 Answer 1

Reset to default
3

This is commonly called image registration (https://en.wikipedia.org/wiki/Image_registration). You might also look through OpenCV (https://docs.opencv.org/3.4.1/db/d61/group__reg.html) for an actual implementation.

This used in processing MRIs to get rid of motion by the patient and I've heard it is used in mask alignment in chip manufacturing (finding and aligning fiducial marks) - exactly the same problem that you are describing in your question.

It is also possible to convert from pixel based images to vector based images. A classmate of mine found work in a company that I cannot remember its name (it was close to 30 years ago) doing just that. However, I am uncertain if that would get you any closer to the problem of image registration.

Improve this answer
3
  • 1
    for the last part this is finding the affine tranformation between the two no? math.stackexchange.com/questions/584272/… Commented Mar 14, 2018 at 13:09
  • 2
    @jk. Depends on what kind of perspective differences are between the views. Under general assumptions (that everything is rigid, i.e. the object, etc.) and the only thing that changed is the camera position and pointing direction, 2D or 3D perspective transform (homography) should be sufficient. In practice, there are lots of complicated details, not for the faint of heart. The task is easier if a human is allowed to manually tweak the parameters slightly for each individual task (i.e. pair of images). Commented Mar 15, 2018 at 7:10
  • 1
    If, the problem is not formulated as "given two images, determine if they're the same up to transformation", but rather, "given one image, determine if the image is closest to one of the images from a database of N images", then the whole issue becomes "image-based content retrieval", which requires completely different approaches, such as extraction of fiducial marks, invariant signatures, feature points (KLT, SIFT, SURF, LBP, ...) other features, etc. Commented Mar 15, 2018 at 7:12

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.