How to reduce the number of "for" loops (while implementing an overlapping block-wise processing on an image)?

I have the following code section to implement overlapping blocks in any square image (my image size 128*128). For each block, I would like to find HOG (Histogram of Oriented Gradients) features and then concatenate them at last to use it for classification.

The block size = 64

The step size = 16 (Both Horizontal & Vertical)

How can I implement a faster version of the following code section without using "for" loops? :

      window_size=64;
      step_size=16;

      %the input image is stored in the variable "image"

      for k=0:4
            for n=0:4
                   for i=((step_size*n)+1):window_size+(step_size*n)
                       for j=((step_size*k)+1):window_size+(step_size*k)   

                              img(p,q)=image(i,j);     %get the block and store it in "img"
                              q=q+1;
                       end
                       p=p+1;q=1;
                   end
                   hog_block{k+1,n+1}=extractHOGFeatures(img);       %storing each block's HOG features in a cell
                   p=1;
             end
       end
             hog_blocks_mat=cell2mat(lbp);      %convert to matrix
           hog_vector=reshape(hog_blocks_mat',1,numel(hog_blocks_mat));    %convertto a row vector

Please help.

Answer 1

[Jan]

These loops:

for i=((step_size*n)+1):window_size+(step_size*n)
  for j=((step_size*k)+1):window_size+(step_size*k)   
    img(p,q)=image(i,j);     %get the block and store it in "img"
    q=q+1;
  end
  p=p+1;q=1;
end

can be rewritten to:

iIni = (step_size*n) + 1);
iFin = window_size+(step_size*n);
jIni = (step_size * k) + 1;
jFin = window_size + (step_size * k);
img  = image(iIni:iFin, jIni:jFin);

[Tintumon]

Dear Jan Simon,

Thank you. This is exactly what I needed. It's working fine. :)

Answer 2

[Tohru Kikawada]

You can use blockproc for block processing. 'UseParallel' option enables to execute in parallel. See this link for details.

[Tintumon]

Dear Tohru Kikawada,

Thank you for the reply.

I have also used "blockproc" function. It is as follows:

fun = @(block_struct) extractHOGFeatures(block_struct.data);    
hog_blocks_mat=blockproc(image,[window_size window_size],fun);
hog_vector=reshape(hog_blocks_mat,1,numel(hog_blocks_mat));

My question is: 1) By how much step (in horizontal and vertical direction) does the block move using "blockproc" function?

2) And, is there any way to change the step size in the function?

Note that I got an improvement in testing accuracy of 2%, while I used the code with the "for" loops, as compared to the "blockproc".

[Image Analyst]

Yes to both questions. It will move in steps/jumps of window_size. You can change the value of window_size. You can also move sliding (not full jumps). See the help.