Different high order approaches for R(u) have been proposed. For example, [55, 35, 36, 13] all use second order information (i.e., second order derivatives), so it is expected for them to be able to model noise better than those using only first order information such as the well-known total variation (TV) model [41, 18].