1

I am performing image registration by extracting SURF features from 2 multi temporal satellite images, in Matlab, as explained here.

The steps provided seem to work great for me, and I can visually see that the target image is aligned correctly to the fixed image.

How do I access the accuracy of the method? In other words, is there any way to make a statement such as this "these 2 images are "XX" % aligned to each other" ?

Nancy
  • 315
  • 1
  • 5
  • 16
  • 1
    To be able to do that you must know the result of 100%. If you knew the result, then you wouldn't need registration! If the images are strictly equal, then correlation will give you this information. – Ander Biguri Feb 25 '16 at 08:27
  • When I see the earlier _slightly_ mis-registered(or you can say only slightly different) images after applying registration, I see the difference that yes, now these are perfectly aligned. But how do I say that in a mathematical way, kind of some metric that substantiates my claim that my program has indeed registered them? – Nancy Feb 25 '16 at 11:13
  • 1
    MY answer to your comment is to read my other comment again. In order to know how close you are to the truth, **you need to know the truth**. But if you knew the truth, why would you even be triying to solve it? – Ander Biguri Feb 25 '16 at 11:21

1 Answers1

1

There is not a way to get a true "XX%" answer, as discussed in the comments, unless the images are identical or you already know the answer. Other ideas:

1) You could use correlation to get a value and set a threshold for your results. (e.g., assume images with corr. coeff. > 0.8 are aligned well). Threshold depends on your data.

2) You could restrict the correlation to the immediate regions of the registration marks or features selected by the algorithm. Then assume that the registration features are identical between the two images, so that a corr. coeff. == 1.0 (mean value across all features) is perfect (100%) registration. This is likely to always be a high number, if the matching algorithm is working.

3) Pick one matched feature to exclude from the registration algorithm, and use the correlation of that region in the two images after matching as your % value. This is maybe the most rigorous method, but it requires a known feature to work with.

All the results based on correlation assume some normalization of the image data, adjusting for contrast, scale, etc.

mhopeng
  • 1,063
  • 1
  • 7
  • 17