Visible and infrared image fusion using ℓ0-generalized total variation model

The tuning of the values of the proposed method’s parameters is an interesting and complex topic. The parameters of the proposed method are presented in Table A1 , i.e., κ, β and λ. It should be noted that β is a positive penalty parameter. β is important to the success of the iteration. Some methods are developed to update this parameter [1], such as increment by a constant multiple (5 or 10). In our experiments, the constant multiple is √ 10. A detailed discussion about the strategy for choosing and updating the penalty parameter can be found in [1] . Base on the definition of proximal ADMM in [2] and the sub-problem (5) in our paper, we assume that κ lies in (0, 1 β‖∇‖2+β ). It can be seen that κ varies with β. Moreover, ‖∇‖2 is pre-computed. For these parameters, we verified experimentally. However, after setting β as described earlier, we observe that constant values for these parameters tended to lead to near-optimal results. We choose these values after evaluating the corresponding quality indexes for each experiment. It is well-known that the choice of regularization parameter λ is a sophisticated topic. A number of techniques have been developed to determine this parameter, such as unbiased predictive risk estimator method, generalized cross validation, discrepancy principle, L-curve method [3]. A comparison of these methods can be found in [3]. We set it experimentally by evaluating the corresponding quality indexes. Then, we chose a set of parameter values that were the same for all experiments.