Learned anisotropic gradient fidelity term-driven unfolding network for single image deraining

Rain streaks often reduce outdoor imaging system performance. Ample research has been done on single image deraining (SID) tasks. Specifically, learned model-driven networks have shown powerful SID ability owing to their interpretability and fast inference abilities, so they have attracted more attention. However, existing model-driven networks typically leverage image intensity to constrain the error between derained results and the ground truth, which limits further improvement of image deraining performance. In addition, they have numerous model parameters. To address the above issues, we propose a learned anisotropic gradient knowledge-driven unfolding network for SID. Specifically, within the maximum a posteriori framework, we propose using anisotropic gradient-based fidelity terms, instead of the common intensity-based fidelity term, to control errors. Moreover, we employ trainable anisotropic gradient operators instead of the fixed gradient ones to enhance the adaptability of gradient representation. Our method can be easily solved by the proximal gradient descent method, and it can be unfolded into a trainable network. The experimental results demonstrate that, compared with the state-of-the-art methods, our approach achieves comparable or even better deraining performance but reduces model parameters by over 30%.