Converting 3D patches to 3D image

Hey guys,

I have a bunch of patches generated from a 3D image of size 64x64x25. I now want to retore the image with these pixels. I was able to do this with the following code:

function [img] = patch2im_2d_time(patch, size_img, size_patch, size_skip, border)
Nx = size_img(1);
Ny = size_img(2);
Nt = size_img(5);

psz1 = size_patch(1);
psz2 = size_patch(2);
psz3 = size_patch(3);

%Extract blocks. This is not necessary
patches  = reshape(patch, [psz1 psz2 psz3 size(patch,2)]);
c = 1;
img2 = zeros(squeeze(size_img));
%Count for each pixel how many times we added smth to it.
add_count = zeros(size_img);

%The first three loops, loop through all the pixels in the image
for d=1:Nt-psz3+1
    for j=1:Nx-psz2+1
        for i=1:Ny-psz1+1
            %Here we get the next patch. The next patch is always
            %the patch that has the pixel at i,j,d at its top front corner.
            current_patch = patches(:,:,:,c);
            %counter for the next patch
            c = c + 1;
            %In this loop we add the patch values of each pixel in the
            %patch to the image. i,j,d is the base. We add the offset
            %ii jj and dd to it. This iteration takes psz^3 many
            %iterations.
            for dd=1:psz3
                for ii=1:psz2
                    for jj=1:psz1
                        img2(i+ii-1,j+jj-1,d+dd-1) = img2(i+ii-1,j+jj-1,d+dd-1) + current_patch(ii,jj,dd);
                        add_count(i+ii-1,j+jj-1,d+dd-1) = add_count(i+ii-1,j+jj-1,d+dd-1) + 1;
                    end
                end
            end

          end
      end
  end
      img = flipud(rot90(img2 ./ add_count,1));
  end

The problem is that it runs for 15 (!) seconds on my machine. I have an implementation that does it for the 2D case in under 1 second. I know, the for loops are just horrible, but I can't figure out a better way. If you quickly want to try some things, I suggest that you use

reshape(linspace(1,1000,1000),[10 10 10]);

To generate a 10x10x10 matrix that has the indices as entries. That is how I came up with the bad solution. It would be great if somebody could provide an alternative, as I am using this in an iterative algorithm and the overhead is horrible (even for testing).

[Duncan Lilley]

If it is possible in your case, using vectorization instead of loops may improve the runtime of your function.

Here is a documentation page for vectorization in MATLAB.

Answer 1

[mouna barhoumi]

hello, I need this fuction but in python, someone help me and render this code in python :)

[madhan ravi]

this forum is for Matlab not python