Just the point(s) r away from each point in the input image array in each plane. If the image were 256x256, and r were 5, say, then xlow, xupp for each point would start out at
xlow = max (1,(i-r));
for i=1, max(1,1-5) --> 1 similarly until i>r at which time then max(1,I-r) --> max(1,5-5) --> max(1,0) --> 1
xupp is symmetric excepting it's preventing the subsequent loop over ii from exceeding M just as xlow prevents addressing elements outside the array on the lower range.
Take and element out in the middle somewhere-- i=130, j=100
xlow=max( 1,130-5) --> 125
xupp=min(256,130+5) --> 135
ylow=max( 1,100-5) --> 95
yupp=min(256,100+5) --> 105
so, the variance is computed for each point over that square of points around each i,j of the array.
The function blockproc in the Image Processing Toolbox does this for any aribtrary function handle rather than having to "roll your own".
