indexing takes a very long time

hi

i have a large multi dimentional matrix that i need to iterate pixel by pixel, lets say size(A) = [2000,2000,100];

each pixel that i iterate through i take it's neighbourhood .

btw, to my appication I have to iterate by for loop.

the problem: indexing , for example

b = A(i:i+10,i+i+10,:)

takes a very long time for some reason.

i performed timing with profiler and that line is the most expensive one (timely wise).

why is indexing such an expensive operation? is there another way to do this?

thanks

[KSSV]

Obviously iterations will be slow........welcome to MATLAB, in most of the cases you can avoid iterations. We cannot help you unless we know what exactly you are doing?

[Adam]

Is the indexing really a very expensive operation or is it just that you are doing 400 million times? Or potentially some multiple of that depending how many such operations you are doing per pixel.

[Vinny]

it is really an expensive operation. (for some reason)

I checked with profiler. this line is my bottelneck

[Bruno Luong]

If you have image processing toolbox you should take a look at function im2col, blockproc or such.

[David Goodmanson]

Hi Vinny,

By

b = A(i:i+10,i+i+10,:)

I assume you mean

b = A(i:i+10,i:i+10,:) 

I realize that you are looking for more than just modest speedup. However, it's interesting that if you implement

B = permute(A,[3 1 2])

to put the third dimension first,

b = B(:,i:i+10,i:i+10)

then b has a different shape but the same elements as before, and the code.is about two and a half times faster. I believe that's because the way that the elements are read out, now the 'colon' elements have consecutive memory locations.

The question doesn't say either way, but If you are doing this process in order to average A over a moving window, i.e. convolution, then you can use fft and gain perhaps an additional factor of four.

[Walter Roberson]

Could you explain more about the reason you use a for loop? Is that an external requirement (for example it might be imposed in homework)? Does each iteration depend upon the results of the previous iteration?

If you need to process a neighborhood then conv2 or filter2 or blockproc are often useful, and for more complex cases, nlfilt can be useful.

[Vinny]

All these functions require an image (meaning: mXnX1 or mXnX3 array).

I deal with a general mXnXC array. that is the reason i need a for loop.

[Vinny]

@David Googmanson

this solution works! my code is a lot faster

thank you

Answer 1

[Matt J]

It takes a long time because every time you execute the line

b = A(i:i+10,i+10,:) 

memory is allocated for b. Instead of iterating over image neighborhoods in Mcode, you should try to write your operations in terms of use built-in functions like convn, which can access the neighborhoods in a more economical way.