Block average images with MATLAB's blkproc or blockproc function

I recently had to perform a block average on an image to reduce its size for processing in MATLAB and learned about a useful function called blkproc or blockproc (in the newer versions) in the Image Processing Toolbox. Block averaging is a process by which you average non-overlapping blocks of an image, which becomes a single pixel in the block averaged image. The standard image resize functions use a filter to resize the image and does not allow the blocks to be non-overlapping.

MATLAB's blkproc or blockproc functions facilitate this by making it easy for you to specify a function to be applied to blocks in the image. If you have an image img and want to do a M x N block average (i.e. the pixels of the resultant image are the average of M x N blocks in the original), you can use:

img = blkproc(img, [M N], 'mean2');

In newer versions of MATLAB, blockproc is preferred and is used this way:

fun = @(block_struct) mean2(block_struct.data);
img = blockproc(img, [M N], 'fun');

mean2 is an Image Processing Toolbox function that takes a 2D array and returns the mean of all values. The distinction between blkproc usage and blockproc usage is that with blockproc, the function that it expects in place of mean2 takes a block struct structure instead of an array. Hence, we need to define a new inline function fun which provides the right interface by wrapping the mean2 function.

What blkproc / blockproc does is divides the image up into M x N blocks and feeds each one to mean2 and then takes the result and puts it into a single pixel in the new image.

You can replace mean2 with other functions that take an M x N array and returns a M' x N' array (where M' and N' are arbitrary numbers; for mean2 this is always 1 x 1) and the result will be constructed by tiling the M' x N' arrays in the order that the M x N blocks occur in the original image. The figure below shows what happens:

The original image is 20 x 20 pixels. Setting M = 4 and N = 4 and using mean2 (which outputs M' = 1 and N' = 1), the final image is 5 x 5 pixels.

When the image cannot be divided up into an integer number of M x N blocks, you will get border effects as the blkproc function pads images out by zeros. The newer version blockproc allows you to specify how you want to treat the partial blocks, either process them as is or pad them with zeros, with the 'PadPartialBlocks' parameter. See the MATLAB documentation for details.

The MATLAB documentation does not indicate the actual order of processing of the blocks (i.e. the actual sequence by which the blocks are processed). If the blkproc or blockproc implementation is parallel or multithreaded, there will not be any guaranteed order at all.

I believe that the blkproc function is single threaded on a single core computer, so the order of processing should be simple. To find out, I used the following test function, which does nothing except print the value of each image block pased into it. When I use this function with blkproc, I will use a block size of 1 x 1 so it will only print one value each time the function is executed.

function a = testblkproc(imblock)
disp(sprintf('%g', imblock))
a = imblock;

I then constructed a matrix of values that I can easily interpret:

>> a = [1 4 7; 2 5 8; 3 6 9]
 
a =
 
     1     4     7
     2     5     8
     3     6     9

Finally, I ran blkproc with my test function testblkproc using a block size of 1 x 1. This makes it output the order in which the 1 x 1 blocks are procesed:

>> blkproc(a,[1 1],'testblkproc');
1
4
7
2
5
8
3
6
9

According to this test, blkproc processes entire rows first before moving onto the next row. I should re-iterate that this is only applicable to the blkproc function running on a single core computer.