Spectral Quadratic Variation Regularized Autoweighted Tensor Ring Decomposition for Hyperspectral Image Reconstruction

The structure information of hyperspectral image (HSI) is well-characterized by tensors, surpassing the capabilities of traditional compressive sensing reconstruction models based on vectors and matrices. Tensor decomposition has been integrated with other regularizations in the model-based reconstruction algorithms to capture more priors. However, the existing tensor decomposition fails to achieve the best low-rank approximation. The effectiveness of model-based reconstruction methods can be promoted. In this article, a subspace-based model utilizing spectral quadratic variation regularized autoweighted tensor ring (TR) decomposition is proposed to explore the multiple-layer spatial-spectral priors of HSI. The original HSI is decomposed into the feature image and spectral basis to explore the first-layer spectral low-rankness. To capture the second-layer low-rank prior, TR decomposition is applied to obtain the effective low-rankness approximation and low computational complexity. The tensor nuclear norm is employed to describe underlying structure priors of the TR factors, which address the deficiency in tensor rank robustness and enhance the reconstruction quality. An autoweighted mechanism is utilized to account for the varying contributions of different TR factors to the low-rank approximation. Moreover, embedding spectral quadratic variation into subspace decomposition enhances spectral smoothness and continuity. Alternating minimization is used to optimize the spectral basis and feature HSI. Through comparative experiments on three datasets, the superiority of the proposed model is demonstrated.