1

What are the steps in feature-based registration?

If I have a transformed image matching to a stack of images (of which only one is the best match), what are the steps in registration?

Once I find matches and use them to estimate a transform, what's usually the next step?

haxtar
  • 1,962
  • 3
  • 20
  • 44

1 Answers1

3
  1. Find points/areas which are invariant to different kind of transformations like: translation, rotation, scale.. because this points are invariant to those transformation we can find them in the different images that we want to register and than can have a corresponds between the images. Example for those points/areas are corners,blobs.. .
  2. After we have those points we need a way to describe them that will be insusceptible to noise, change in contrast lighting etc... We do this by using the neighborhood around the point using some descriptor like: SIFT,ORB,SURF .... Those descriptors use different kind of ways to tackle the challenge using the gradients, normalization of the color, histograms etc...
  3. After we have a description for points in all the images that we want to register them one to the other we try to find the corresponds between the points by calculating some distance measure (for example euclidean) between the descriptors.
  4. When we find some corresponds between the points in the different images we can calculate the transformation between them (The number of corresponds points depends on the type of transformation we want to find Rigid,Affine,Prepspective). A further way to deal with outliers is using something like Random sample consensus(RANSAC) or a robust distance measure to find the outliers.
  5. After we calculated the transform in step 4 we apply it to the image we want to register.
Amitay Nachmani
  • 3,259
  • 1
  • 18
  • 21
  • Thanks, Amitay. My question is mostly about step 5. I have a transformed img (affine transform) and a reference img. To which image would I apply the transform? To me, it seems that the answer is - apply it to the transformed img. In fact, this is the correct thing to do for alignment. It's just not intuitive to me how applying a transform to an ALREADY transformed image almost behaves as an inverse transform. – haxtar Aug 18 '16 at 14:20
  • 1
    The titles "reference image" and "transformed image" are arbitrarily. You decide who is which when you use the corresponding points to calculate the transform. If you want to find how to convert coordinates from the transformed image to the reference image you will get the "inverse transform". – Amitay Nachmani Aug 18 '16 at 14:21
  • I am writing this all myself - raw code. I use OpenCV to detect features, but that's not important for estimating the transform. I have used RANSAC to estimate the transform matrix, and after that, I apply it to the transformed img. After that, the transformed img looks almost identical to the reference. I was just wondering... how does that make sense mathematically. – haxtar Aug 18 '16 at 14:24