Finding correlation between two images

Suppose I want to perform correlation between two images. I know there is a function normxcorr2 which can be used to find the correlation between two images (img1, img2) like this

When I run the imfilter command and use the imshow() command to output the resulting image, I can see that it is different from what I get when I output the resulting image using normxcorr2 command.The image generated through normxcorr2 looks more like what i would expect from the correlation of the two images. What's the difference between these two methods?

C = normxcorr2(img1, img2)

But can I find the correlation using imfilter? Does this command perform correlation?

corr = imfilter(img1, img2, 'corr','replicate');

When I run the imfilter command and use the imshow() command to output the resulting image, I can see that it is different from what I get when I output the resulting image using normxcorr2 command.

imshow(C);
figure
imshow(corr);

The image generated through normxcorr2 looks more like what i would expect from the correlation of the two images. What's the difference between these two methods? Is using imfilter even logical?

[Bjorn Gustavsson]

Have you looked in the documentation how normxcorr2 normalizes the x-correlation values for the different shifts, it seems a "natural" guess that it properly scales with the number of samples, while imfilter might "not bother with that" since its purpose is mainly to filter images.

[Mohammad Saquib Khan]

I haven't yet looked into the documentation but I will surely do that now. What if after computing my variable corr, I normalise it? Will then the the two resulting images be same?

Answer 1

[Image Analyst]

normxcorr2() scans the image with a window and does a normalization at each location. imfilter() does not. It simply moves the window, multiplies the pixels together, and sums -- no normalization. You cannot normalize a xcorr2() or imfilter() output image to make it look like what it would look like if it had normalized the subimages at each location.

[Mohammad Saquib Khan]

That clears most part of my doubt. Thank you so much. But I'm still confused whether I should use normxcorr2 or imfilter when I'm asked to find correlation between two images. Using normxcorr2 makes a lot more sense now. Is it correct? Or I should rephrase my question and ask is using imfilter to find correlation wrong?

[Image Analyst]

If you want to find out how much they are different, use xcorr2. If the images could possibly have different overall brightness and contrast, I'd use normxcorr2() to try to ignore those differences.

[Mohammad Saquib Khan]

If I use xcorr2, I just get a simple white image. But when I used normxcorr2, I get an image where I can see a pattern. So I guess normxcorr2 is the right way. Thank you so much for your help!

[Image Analyst]

If your output is a floating point image you're going to have to make sure it's in the range 0-1 or else any values more than 1 will show up as white. Try casting to uint8 before display. If it's a gray scale floating point image, try

imshow(yourImage, []);