Gradient low-rank sparse representation: spatial-spectral group sparse total variation for hyperspectral image denoising

Hyperspectral images (HSIs) mixed noise removal is a challenge and essential preprocessing part in remote sensing fields. In this paper, To address the challenge of mixed noise suppression in HSI, we introduce a composite denoising framework that unifies a novel weighted group sparsity regularization based on l2,PS and low-rank Tucker decomposition (LRPWGS). Rather than applying l2,1-norm constraints directly to the difference image, a weighted l2,PS-norm is employed to capture the shared group sparse pattern in the gradient domain of HSI and a shrinkage operator is proposed to effectively solve l2,PS-norm minimization problem. Furthermore, the conventional low-rank Tucker decomposition is utilized to exploit the global correlations across all HSI tensor modes. The alternating direction multiplier method (ADMM) algorithm is adopted to solve the proposed LRPWGS model. Experimental results demonstrate the superior performance of proposed LRPWGS method in HSI denoising task.