Hyperspectral Image Inpainting via Subspace-Based Nonlocal Total Variation

Hyperspectral images inevitably suffer from significant degradation during the data acquisition process, such as damaged or missing values. As a result, hyperspectral image restoration has gained widespread attention in recent years. Solving these ill-posed problems often requires exploring the prior structures of hyperspectral image (HSI) data and explicitly modeling them in a regularized form. Among the various priors, low-rankness and non-local self-similarity (NLS) are two widely studied features. However, the computational complexity of NLS methods is high, leading to a trade-off between performance and computational efficiency. To mitigate this issue and maintain good performance, the representative coefficient map, obtained through low-rank decomposition of the original data, inherits the prior knowledge of the original image while significantly reducing the dimensionality. In this paper, we propose a hyperspectral image inpainting algorithm based on the representative coefficient map and non-local total variation (NLTV). The proposed method applies NLTV on the representative coefficient map to simultaneously capture the low-rankness and non-local selfsimilarity of the image. Furthermore, we introduce an alternating direction method of multipliers (ADMM) algorithm to solve the proposed model. Through extensive experiments, we compare our method with existing state-of-the-art (SOTA) inpainting algorithms. The results demonstrate that our approach achieves a good balance between restoration performance and computational efficiency.