A novel total variation optimization method and its application on blind super-resolution

The total-variation (TV) regularization is very popular because of its ability to deal with noise while preserving important image features. This paper proposes a novel TV-based algorithm which can be applied to many inverse problems such as image de-convolution and super-resolution. The idea is to break the cost function into two parts: a linear part and a nonlinear part containing the TV term, then handling them one after the other in each iteration. This method has overall advantages considering both quality and speed. Then it is applied to blind super-resolution (SR), serving as the solver for both image estimation and the point spread function (PSF) estimation. The PSF estimation is also improved in this work by eliminating boundary pixel values that have not been computed from complete data. Synthetic and real data experiments show the nice performance of the proposed method.