Figure 1(a) shows the performance for the first 10 iterations of our SFP algorithm for denoising a smoothly varying image (with no visible jumps in it); for this problem, option 2 is almost the same as option 3 in performance. However, Figure 1(b) shows that option 3 is clearly better for denoising an image with a lot of jumps (edges). We also tried other
∆t i,j i,j i,j
M for M = 2,3. Then we found that the refinements of option 3 with T VM (u) = ∇ · ∇u
|∇u|
resulting SFP method is more sensitive to the selection of γ.