FIRNet: Forward-Inverse Reinforcement Network For Image Restoration Through Scattering Media

Image restoration through scattering media is a desired yet challenging task in numerous scenarios. Deep learning (DL)-based approaches have seen significant advancements in recent years, achieving impressive performance. However, most existing solutions have been limited to a single network framework to model the inverse scattering process, leading to relatively poor recovery performance. In this paper, we introduce a forward-inverse reinforcement network (FIRNet) to enhance image recovery performance through scattering media, in which two distinct neural networks are designed to model the inverse and forward scattering processes, respectively. A reinforcement training strategy combined with a variation of GAN loss function is implemented to fully exploit the feature extraction ability of these networks, resulting in higher-fidelity image recovery compared to a single network. Experimental results on different datasets showcase the effectiveness and superiority of our proposed two-stage framework. The promising recovery results indicate that FIRNet could pave the way for new opportunities to enhance image restoration performance in related fields.