Moving average with overlapping windows

Hi,

I am looking for an efficient way to compute a moving average with overlapping windows for a matrix X along the first dimension.

A non-efficient solution is the following.

 N = 3500000;
 X = randn(N,20);

 windowSize  = 600;
 stepSize    = 100;

 b = (1/windowSize)*ones(1,windowSize);
 a = 1;

 windowEnd = windowSize:stepSize:N-windowSize+1;

 tic
 X = filter(b,a,X) ;
 X = X(windowEnd)  ;
 toc

In this solution, I am computing a lot of averages that I am throwing away at the end... This is what I meant by 'non-efficient'.

Thank you in advance for your help.

Answer 1

[Image Analyst]

You can use blockproc() to move a 600x600 window along a matrix in "jumps" of 100. Look at all the arguments and let us know if you can't figure it out.

[Pierre S.]

Thank you for the suggestion. I tried your suggestion. But unfortunately, blockproc() is much slower than filter()... Any other idea? Thanks again.

[Image Analyst]

You can use conv2() or imfilter(). They are very efficient and fast and highly optimized. However, they don't move in "jumps" of 100, they move by 1 pixel so they do a lot more computations. Whether it will be faster or not is for you to test.

tic;
kernel = ones(101) / 101^2;
blurredImage = conv2(X, kernel, 'same');
elapsedSeconds = toc