This is a stylised problem which could be solved easily using a processor that has more prior information: for example, if the shape of the target image were known. That is not the point. What is required is a primary vision system that obeys Marr's ``principle of least commitment'' [ 34 ] but preserves as much structure as possible. It appears that the sieve preserves structure and does so more robustly than the diffusion system.
As contenders for primary vision systems, sieves have some advantages. They use some of the recent results from mathematical morphology to form a processor that satisfies many of the desirable scale-space axioms and robustly separates noise and occlusion according to scale. Furthermore, they do so with a realistic amount of computation.
A. Bosson ESE PG