Total Variation Regularized Low-Rank Model With Directional Information for Hyperspectral Image Restoration

Hyperspectral images (HSIs) are unavoidably polluted by various kinds of noise, which decrease the potential of subsequent processes in HSIs applications. Due to the diversity and complexity of HSIs mixed noise, including impulse noise, Gaussian noise, stripe noise and deadlines, traditional restoration technology cannot be used directly. In this paper, a novel HSIs restoration approach is proposed that integrates low-rank (LR) prior and spatial-spectral total variation with directional information. Specifically, by analyzing the characteristic of spatial-dependent edge and texture directional structure, a spatial-spectral directional total variation (SSDTV) regularization is defined. Then, considering the HSIs as a cube data, the proposed method utilizes SSDTV regularization to characterize spatial-spectral smoothness, as well as LR regularization to constrain spectral consistency. Finally, an extended alternating direction method of multipliers algorithm is designed to achieve simple and fast implementation, in which the complex optimization problem is separated into several easier subproblems. Both simulated and real-world HSIs experiments indicated that the proposed method is effective and numerically feasible for HSIs Restoration.