A multiplicative noise removal model based on TGV with spatially adaptive regularization parameters

In this article, we propose a total generalized variation (TGV) [1] based model for removing multiplicative Gamma noise. To preserve edge more, we adopt a nonconvex regularizer to TGV regularization term. The model integrates the data-fitting energy proposed in [2] with a spatially adaptive regularization parameter (SARP) approach. The data-fidelity term enables to deal with heavy multiplicative noise. And the SARP allows to preserve textures and edges while effectively removing the noise in homogeneous regions. Since nonconvex regularizer is adopted, the objective function is nonconvex function. To deal with nonconvex function, we use iteratively reweighted `1 algorithm introduced in [3]. We present the first-order optimality characterization of our model and derive a SARP algorithm from it. Numerical results are shown to demonstrate the efficiency of our model, with comparisons with some existing TGV based models.