How do I compute the maxpool of a image? Let us say stride of 2,2 on a mxn matrix?

If I were to implement just the max pooling operation on an image as mentioned in the following page https://www.quora.com/What-is-max-pooling-in-convolutional-neural-networks What is the most efficient way of computing it without going into for loops

Answer 1

[Kannan U V]

The following does the trick

fun = @(block_struct) max(block_struct.data(:));
b = blockproc (a, [X Y], fun);

[Matt J]

But it is not very efficient. Compare:

a=rand(5000);
X=4; Y=4; %window sizes

tic
fun = @(block_struct) max(block_struct.data(:));
b = blockproc (a, [X Y], fun);
toc
%Elapsed time is 19.764354 seconds.

tic
b=sepblockfun(a,[X,Y],'max');
toc
%Elapsed time is 0.092457 seconds.

It is probably in fact the least efficient approach you could use. Even a double for-loop is faster:

    tic;

         [m,n]=size(a);
         ex=ones(1,m/X)*X;
         ey=ones(1,n/Y)*Y;

         ac=mat2cell(a,ex,ey);

         for i=1:m/X
             for j=1:n/Y

                 ac{i,j}=max(ac{i,j}(:));

             end 
         end

         b=cell2mat(ac);

    toc
    %Elapsed time is 6.203763 seconds.

Answer 2

[Matt J]

What is the most efficient way of computing it without going into for loops

The most efficient way in the entire universe is to use SEPBLOCKFUN (Download) as follows,

X=2; Y=2; %window sizes

maxpool=sepblockfun(yourImage,[X,Y],'max');

This assumes the image dimensions m,n are evenly divisible by X,Y respectively. Otherwise, you must pad the image to make it so.

[Kannan U V]

Thanks for your time and answer. It looks like the following does the trick

fun = @(block_struct) max(block_struct.data(:));
b = blockproc (a, [X Y], fun);

[suWits Mohrarsii]

thank you for this function

Answer 3

[Anton Semechko]

Here is an example:

% Sample image
im=imread('cameraman.tif'); % sample image

% 4 pixels comprising non-overlapping 2-by-2 neighbourhoods  
im_nw=im(1:2:end,1:2:end);
im_sw=im(2:2:end,1:2:end);
im_se=im(2:2:end,2:2:end);
im_ne=im(1:2:end,2:2:end);

% Select pixel with maximum intensity
im_max=max(cat(3,im_nw,im_sw,im_se,im_ne),[],3);

% Visualize
figure('color','w')

ha=subplot(1,2,1);
imshow(im,imref2d(size(im)))
title(ha,'original','FontSize',20)

ha=subplot(1,2,2);
imshow(im_max,imref2d(size(im_max)))
title(ha,'2x2 max-pool','FontSize',20)

Note that even though two images appear to have the same size when visualized using 'imshow', the dimensions of im_max are half that of im. Recursive application of 2-by-2 max-pool will result in downsampled images with sizes 1/2, 1/4, 1/8, etc. of the original image.

[Kannan U V]

Thanks for your time and answer. I am looking for variable X, Y window sizes too.