Efficient subpixel image registration by cross-correlation

Registers two images (2-D rigid translation) within a fraction of a pixel specified by the user.

Description

Registers two images (2-D rigid translation) within a fraction of a pixel specified by the user. Instead of computing a zero-padded FFT (fast Fourier transform), this code uses selective upsampling by a matrix-multiply DFT (discrete FT) to dramatically reduce computation time and memory without sacrificing accuracy. With this procedure all the image points are used to compute the upsampled cross-correlation in a very small neighborhood around its peak. This algorithm is referred to as the single-step DFT algorithm in [1].

[1] Manuel Guizar-Sicairos, Samuel T. Thurman, and James R. Fienup, "Efficient subpixel image registration algorithms," Opt. Lett. 33, 156-158 (2008).

Please refer to the attached HTML for more details and a sample implementation.

• Acknowledgements: This submission has inspired the following:; Automatic 2d Rigid Body Image Registration, STXM data analysis script collection - STACKLab
• MATLAB release: MATLAB 7.4 (R2007a)

[Chris Cookson]

easy to use, extremely fast, and easy to use. amazing work.

[Chris Cookson]

did i mention it was easy to use?

[Matt Sochor]

Very good and excellent documentation to get you started with the syntax. Great job and thanks for including the paper reference.

[Jiang Jun]

Thank you for you job, I will enjoy it.

[Zhi Zhang]

Excellent Work!

[Robert Reed]

Elegant solution in simple understandable code + sound documentation.

[Petter]

Always nice to see a published .m, but the images tell us nothing. A zoomed in display, abs error or something would be more relevant than two identially looking images.

[Robert Reed]

Is this method less robust to noise than standard upsampling?

[Chaos]

works but i had to re=write code for UINT16 images, very fast and accurate.

[Chaos]

works but i had to re=write code for UINT16 images, very fast and accurate. why are you using double? do you really need that accuracy in the dft?

[Chaos]

 forgot to rate. wow this code is brutally fast and accurate, if you define a ROI even faster. this code needs some kind of award. how much faster would it be if it was compiled in to a mex?

[Manuel Guizar]

Thanks to everyone for the positive comments! The published M-file is intended to get the user started with the syntax and explain what the code does. For technical documentation please refer to the reference above [1] and references within.

[Manuel Guizar]

This method is equally robust to noise as the standard FFT upsampling, but much faster. It computes the same upsampled cross correlation but only in a small neighborhood around the peak.

[Tobias Henn]

Great Work - fast and accurate!!!

[Lucas Koorneef]

Works right out of the box. Thanks for sharing !

[Eric]

Tremendously useful! Thanks very much.

[su su]

I want to register the pair of visible and theraml images with this one, but I don't get the required resutl.

[Raymond Cheng]

Thanks for your sharing.

[Sean]

Is there a way to extend this to 3D CT images?

Thank you!

[Rob Campbell]

Wonderful. Fast and accurate.

[Alex]

Code registers images very well. Question about the supplementary output:
output = [error,diffphase,net_row_shift,net_col_shift]

On my data it seems that some times the row_shift variable seems to 'wrap around' in away that is not intuitive. For example though the data is only 512 pixels wide, I get row shifts of -512.3, 1024.3 etc which I assume are actually very small pixel shifts that have a somewhat random multiple of the image size thown on for fun.

[Phillip]

Great code. Any plans to extend to cover angular rotations and scale shifts as well? (e.g. Fourier Mellin transform)

[Alex]

Re: my comment on 14 Jul 2010:
I owe an apology. It seems at the point when I wrote my comment I was feeding an accidentally reshaped array into the code. Because of the artifacts in my data, it caused a wrap around artifact in the output (and the registration worked in one axis only, but I didn't notice because that happened to be the axis containing almost all of the shifts).

Great code - no resrvations! Thanks much.

--Alex.

[Yu Yinghuai]

Great work! Question about the codes " kernr=exp((-i*2*pi/(nr*usfac))*( [0:nor-1].' - roff )*( ifftshift([0:nr-1]) - floor(nr/2) )) " :
Why does the vector [0:nr-1] need be ifftshift ?

[Nafise b]

Thanks for your code.
I want to work with your code, but I have a problem with it.
In fact, my images are 2 blocks which don't have the same size, and I want to find the best place for small block in the big block.
Something that I got from your code is for your code 2 images have same size.
What should I do now, if I want to use your code for subpixel image registration?

Best regards,
Nafise

[Sean de]

Nafise b, pad your smaller image with zeros so that it is the same size.

[Nafise b]

ThX sean,
I did it, but it is not true, because my image has black regions, and it determines that part with zero pad part.
WHAT SHOULD i DO NOW?!

[Sean de]

Nafise, post your situation/goals, your images, and what you've tried to the newsgroup (http://www.mathworks.com/matlabcentral/newsreader/). From there people can help you.

[Nafise b]

ThX, I did it, hopefully some one answer me.

[Yan G]

 

[Aron Sceidt]

can this be applied to colour images in some way?

[Michael]

do images need to be square - as in 2^n for the fft shift to work properly?

[Rui]

 

[Rui]

works very well. i'm amazed with the speed and accuracy of the algorithm.

thanks manuel!

[Greg]

 

[Greg]

Great algorithm, it has helped so much with my image stabilization project

[Daniel]

wow, that one is really fast and seems to be robust. thanks a lot!