Robust Tensor Completion via Spatial-Spectral Constrained Deep Low-Rank Tensor Factorization for Hyperspectral Image Recovery

Robust tensor completion of hyperspectral image (HSI) is a challenging task in the field of remote sensing. Recently, nuclear norm minimization-based methods have made certain progress in robust tensor completion. However, the tensor nuclear norm applies the same constraint to all singular values, resulting in insufficient capturing power for the global structure of the HSI. In addition, as a convex surrogate of global low-rankness, tensor nuclear norm minimization leads to an overall low-rank approximation that cannot capture the details of the HSI. In this letter, we propose the spatial-spectral constrained deep low-rank tensor factorization (SDLTF). More precisely, the low-rank tensor factorization is used to dynamically assign penalty weights, aiming to preserve the main information and maintain the global structure of the HSI. The spatial-spectral constrained unsupervised deep prior is applied within a deep convolutional neural network to capture spatial-spectral correlations and local details of the HSI. We develop an efficient algorithm to tackle the corresponding model based on the ADMM. Extensive experiments demonstrate that our model has superior performance compared with several state-of-the-art methods.