imregcorr() misaligns images badly

[tform, peakregcorr] = imregcorr(SourceImageAdj,SourceRef, ...
    TargetImageAdj,TargetRef, ...
    'transformType', 'similarity', ...
    'Window', true);
RegisteredImage = imwarp(SourceImageAdj,SourceRef, tform, 'linear', ...
    'OutputView', TargetRef);

disp(peakregcorr)
figure, imshowpair(TargetImageAdj,TargetRef,RegisteredImage, TargetRef)

above is my code to correlate 2 images of different sizes. the goal is to find the correlation between the images to determine what pixel location a point of interest might lie in both images. However, imregcorr consistently misses the mark badly.

Whereas the code above might output a tform object - Dimensionality: 2, Scale: 0.4254, RotationAngle: 2.4851, Translation: [-0.3269 -0.0904]

A more correct tform object (based upon the output image would look like - Dimensionality: 2, Scale: 1, RotationAngle: -1, Translation: [0.0640 0]

The SourceRef and TargetRef seem correct, so Im not sure what is causing the issue

[Garmit Pant]

Hello

Can you please provide the input images that you are using?

Regards

[Nathan Nguyen]

these are the input images that I am using

srcimg - source image

srcref - source reference imref2d

targetim - target image

tgtref - target reference imref2d

tform - transformation object that is formed from the images

The problem is that my original source image is 512x512 whereas the target image is 2048x2044

[Steve Eddins]

See my answer below, or my 16-Oct-2025 blog post, for updated information related to releases R2024b and later.

Answer 1

[Matt J]

Phase correlation isn't going to be a good algorithm for images of a starry field. I think your best bet is to do landmark extraction -- either manually with cpselect or automatically if you have some way to do that -- and then use fitgeotform2d.

[sp,tp] = cpselect(SourceImageAdj, TargetImageAdj ,'Wait',true);
 [sp(:,1), sp(:,2)]=intrinsicToWorld(SourceRef, sp(:,1), sp(:,2))
 [tp(:,1), tp(:,2)]=intrinsicToWorld(TargetRef, tp(:,1), tp(:,2))

 tform = fitgeotform2d(sp,tp,'similarity')

RegisteredImage = imwarp(SourceImageAdj,SourceRef, tform, 'linear', ...
    'OutputView', TargetRef);

figure, imshowpair(TargetImageAdj,TargetRef,RegisteredImage, TargetRef)

[Nathan Nguyen]

Thanks yeah manual selection seems to work well!

This isn't relevant to the question specifically but if I found all of the feature detection algorithms from matlab on this webpage. These seems to be pretty useful https://www.mathworks.com/help/images/techniques-supported-by-registration-estimator-app.html

[Steve Eddins]

See my answer below, or my 16-Oct-2025 blog post, for updated information related to releases R2024b and later.

Answer 2

[Steve Eddins]

As of R2024b, the function imregcorr uses a new algorithm called normalized gradient correlation. See my 16-Oct-2025 blog post for more information. The function can now register these star-field images successfully.

load srcimg SourceImageAdj
load srcref.mat SourceRef
load targetim.mat TargetImageAdj6
load tgtref.mat TargetRef

tform = imregcorr(SourceImageAdj, SourceRef, ...
    TargetImageAdj6, TargetRef, ...
    TransformType = "similarity")

tform = 
  simtform2d with properties:

    Dimensionality: 2
             Scale: 0.9981
     RotationAngle: -1.0487
       Translation: [0.0617 -0.0056]
                 R: [2×2 double]

                 A: [ 0.9979    0.0183    0.0617
                     -0.0183    0.9979   -0.0056
                           0         0    1.0000]

registered_image = imwarp(SourceImageAdj, SourceRef, tform, "linear", ...
    OutputView = TargetRef);
imshowpair(TargetImageAdj6, TargetRef, registered_image, TargetRef)