Content-Aware Subspace Low-Rank Tensor Recovery for Hyperspectral Image Restoration

The low-rank tensor model has made great progress for hyperspectral image (HSI) restoration. Recently, the low-rank tensor methods have further been boosted with subspace learning by transforming original HSI into a low-dimensional subspace with reduced computational burden and discriminative feature representation. However, existing subspace-based methods consistently employ a fixed subspace dimension for all patches, which may violate the intrinsic dimension discrepancy of different image content, leading to information loss or redundancy. In this work, our key observation is that the intrinsic subspace of different image patches along different dimensions is different, which should be adaptively modeled for compact feature extraction. Therefore, we propose a content-aware subspace low-rank tensor recovery (CSLRTR) method by leveraging both deep network and low-rank tensor model. Specifically, we first analyze the intrinsic discrepancy of different HSI patches among both spatial and spectral dimensions and design a simple network to adaptively learn the optimal subspace dimension. The adaptive subspace learning and low-rank tensor recovery are iteratively performed and mutually promote each other. On one hand, the learned subspace would contribute to more compact low-rank representation for better restoration; on the other hand, the low-rank tensor recovery with less degradations would definitely ease the difficulty of the subspace estimation. Note that the adaptive content-aware subspace strategy has been simultaneously employed on both spectral and nonlocal dimensions, where the spectral-spatial relationship has been further strengthened with better restoration. We have performed extensive experiments on different datasets and restoration tasks and extended the content-aware subspace strategy to previous methods.