Reinforcement Learning Based Plug-and-Play Method for Hyperspectral Image Reconstruction

Hyperspectral images have multi-dimensional information and play an important role in many fields. Recently, based on the compressed sensing (CS), spectral snapshot compressive imaging (SCI) can balance spatial and spectral resolution compared with traditional methods, so it has attached more and more attention. The Plug-and-Play (PnP) framework based on spectral SCI can effectively reconstruct high-quality hyperspectral images, but there exists a serious problem of parameter dependence. In this paper, we propose a PnP hyperspectral reconstruction method based on reinforcement learning (RL), where a suitable policy network through deep reinforcement learning can adaptively tune the parameters in the PnP method to adjust the denoising strength, penalty factor of the deep denoising network, and the terminal time of iterative optimization. Compared with other model-based and learning-based methods and methods with different parameters tuning policies, the reconstruction results obtained by the proposed method have advantages in quantitative indicators and visual effects.