id = id(:);

I would like to compute all 8 neighbors for each pixel in a 20×40 image matrix except for the borders (1-pixel width). I do the following:

  [cols, rows] = meshgrid(2:19, 2:39);
  id = sub2ind(size(img), rows, cols);
  M = size(img,1);
  msk = [-M, -M+1, 1, M+1, M, M-1, -1, -M-1];
  nbrs = bsxfun(@plus, id, msk);

However, I get an error in the bsxfun():
Non-singleton dimensions of the two input arrays must match each other.

What should I be doing in order to get an 8-dimensional vector corresponding to each pixels in the ROI (2:19, 2:39) of my image? Is there a straight forward approach for this?

seed_pt=1;
neighbor_offsets =[M, M+1, 1, -M+1, -M, -M-1, -1, M-1];
nghood = bsxfun(@plus, seed_pt, neighbor_offsets);

Your initial seed_pt is at the upper-left of the image, on the boundary. Therefore your computation of its neighborhood using neighbor indexing on the next two lines doesn’t work correctly.

 if (((xn>=1)&&(yn>=1))&&((xn<=M)&&(yn<=N)))

This restricts the coordinates of the pixels so that only those in the image will be assigned to the local_data matrix. This works well with 4-neighbor but there is one wrong index when I use the 8-neighbor offsets. This is what I really want to know.

seed_pt=1;
neighbor_offsets =[M, M+1, 1, -M+1, -M, -M-1, -1, M-1];
nghood = bsxfun(@plus, seed_pt, neighbor_offsets);
nghood=nghood';
nd=size(nghood,1);
cnt=1;
clear local_data
for j=1:nd
       x0=nghood(j);
       yn=ceil(x0/M); %convert to column coordinate
       xn=rem(x0,M);  %convert to row coordinate
       if xn==0, xn=M;end
       if (((xn>=1)&&(yn>=1))&&((xn<=M)&&(yn<=N)))
         local_data(cnt,:)=reshape(I(xn,yn,:),1,3);
         cnt=cnt+1;

       end
end