Superpixel-Based Hyperspectral Image Denoising via Local-Global Low-Rank Approximation

Recently, superpixel segmentation-based hyperspectral image (HSI) denoising methods have attracted increasing attention, since they could obtain the size-adaptive superpixel fiber rather than a cube with fixed spatial size. The superpixel fiber flexibly exploits the local similarity at different scales and leads to significant low-rankness. In this paper, we propose the parallel HSI denoising models which simultaneously consider the local and global low-rankness of the HSI based on superpixel segmentation. In the proposed models, the non-convex but smooth log-determination function is adopted to better characterize the low-rankness of the HSI. We also propose an adaptive weighted strategy to optimize the restored HSI. An efficient iterative algorithm is developed to solve the parallel models. Several experiments verify the superior performance of the proposed approach over other competing methods.