Abstract: Aiming at the problem of mixed noise in hyperspectral images, we proposd a method based on subspace representation and weighted low-rank tensor regularization to remove mixed noise from hyperspectral images. The subspace representation used the correlation between spectral bands to select an appropriate orthogonal matrix and project the hyperspectral image into a low-dimensional subspace, so that  the proposed algorithm had lower complexity and simplified the denoising process while  removing part of the noise in the image. The denoising process was based on the low-rank tensor extracted from the simplified image. The regularization term of the weighted low-rank tensor was introduced to represent the prior information of the simplified image subspace. A reasonable weighting mechanism was constructed based on the physical meaning of the nuclear norm in the Tucker decomposition, and the intrinsic structural correlation of the hyperspectral image was preserved. We designed  a method based on  iterative minimization to solve the proposed non-convex denoising model. The experimental results on simulated and real datasets show that the proposed subspace representation and weighted low-rank tensor regularization method achieve high denoising performance in both quantitative and qualitative analysis.

Key words: hyperspectral image denoising, subspace representation, weighted low-rank tensor regularization