Fastest way to process image patches?

Hi all,

My matlab script is almost entirely a big loop that searches through small patches of an image and computes sum-of-square-differences with a "target patch", like this:

for i = 1:num_pixels_in_image
  patch = image(i-5:i+5,j-5:j+5);
  ssd(i) = sum(patch(:) - target(:)).^2;
end

Naturally, this process is very slow when the number of pixels grows large. I'm wondering what the absolutely most efficient way to implement this is. Problems such as this, it seems, don't lend themselves easily to vectorization. Cheers!

Answer 1

[Image Analyst]

Not sure why you're doing that but I'm not sure you should do it that way. I think you maybe should use normalized cross correlation instead. There is a function that does that called normxcorr2(). I attach a demo. Basically it will give a high signal when the image patch is like the target patch and a low signal when the target patch is not like the image patch. Why do you want to do it the way you said? What is the overall goal of that algorithm? To determine where in the image is like the patch? That's what normxcorr2() is for.

[JohnDapper]

I'm implementing a very specific patch-based image processing algorithm which needs the L2 distance between certain patches.

https://en.wikipedia.org/wiki/PatchMatch

[Image Analyst]

You're showing (sum of differences) squared, not sum(differences squared). There is a BIG difference. To get what you showed, simply use conv2 and square it. For the other, try nlfilter().

[JohnDapper]

Good point. The parantheses is misplaced. For the sum of squared differences, you suggest nlfilter? I'll take a look at that, thanks.

[JohnDapper]

For anyone who may be interested, C++ mex is the way to go here.