L. Xiao, F. Heide, W. Heidrich, B. Schölkopf, M. Hirsch
IEEE Transactions on Image Processing, (2018)
Recently, several discriminative learning approachesave been proposed for effective image restoration, achieving onvincing rade-off between image quality and computational fficiency. However, these methods require separate training for ach restoration task (e.g., denoising, deblurring, demosaicing) and problem condition (e.g., noise level of input images). This akes it ime-consuming and difficult to encompass all tasks nd conditions during training. In this paper, we propose a iscriminative transfer learning method that incorporates formal roximal optimization and discriminative learning for general Image restoration. The method requires a single-pass discrimi- ative training and allows for reuse across various problems and onditions while achieving an efficiency comparable to previous iscriminative approaches. Furthermore, after being trained, our odel can be easily transferred to new likelihood terms to solve ntrained tasks, or be combined with existing priors to further mprove image restoration quality.