How to align two images which are only several pixels away in vertical direction?

Hi, Guys,

There are two images and they are several pixels away in the vertical (Y) direction. I want to align them automatically. My plan is to draw a line in the two images and calculate the Normalized cross-correlation, and find coordinates of peak. The idea is from Registering an Image Using Normalized Cross-Correlation . But I don't know how to find the offset between the two lines.

normxcorr2_2diff = normxcorr2(template_line, other_line);
[max_c, imax] = max(abs(normxcorr2_2diff(:)));
[ypeak, xpeak] = ind2sub(size(normxcorr2_2diff),imax(1));

In the above code, size(template_line) = 89, size(other_line)=176, size(normxcorr2_2diff)=264, ypeak=135. You can load template_line and other_line from the attached .mat file.So how can we find the offset from these results? If the offset can be determined, I can use imtranslate to move the line (image). Let's use template_line as the reference (fixed) line and want to move other_line to align with it.

So this is my thought. Otherwise, is there other easier or more direct approaches to do this? Since two images are away only several pixels in Y direction, is there a globe way (based on two entire images and then find the offset in Y direction) to align them? I also consider imregister . But haven't figure out how to use it in my case.

Thank you very much.

Answer 1

[Image Analyst]

You want to use imregister() instead of normxcorr2(). There is an example in the help that you can easily adapt.

[Stelios Fanourakis]

I use imregister and I get the error "The value of 'MovingImage' is invalid. All dimensions of the moving image should be greater than 4.". Why is that for? Any idea?

[Image Analyst]

You can't do a regression if you have less than the number of points than the order. For example, you can't fit a parabola through 2 points can you? You need at least 3. Likewise you can't register a 3-by-3 image to another 3-by-3 image. Why are your images so small?

[Stelios Fanourakis]

Sorry, it was a mistake of mine but now has fixed. Thank for responding

Answer 2

[Alex Taylor]

Also look at imregcorr. It uses phase correlation and is easier to use than imregister. I'd start there.

[YZ]

Thank you, Alex.

The difference in two image is subtle, probably only 1 or 2 pixels away along Y direction. I used the following codes:

fixed = img_alldyn(:,:,1);
moving = img_alldyn(:,:,2);
tformEstimate = imregcorr(moving,fixed, 'translation');
movingReg = imwarp(moving,tformEstimate);
figure;
subplot(1,3,1); imshow(fixed,[]); imcontrast;
subplot(1,3,2); imshow(moving,[]); imcontrast;
subplot(1,3,3); imshow(movingReg,[]); imcontrast;

As a result, tformEstimate.T = [1,0,0;0,1,0;0,0,1]. What does this mean?

Thank you.

Answer 3

[martin hauhana]

How can I write a code to align NIR and RGB images