wouldnt the solution be more accurate if you partitioned the picture down into smaller segments and then just use the algorithm on each part?