Hyperspectral Image Denoising via Dynamically Adjusted Regional Rank Guided by Adaptive Superpixel Segmentation.

Hyperspectral image (HSI) is inevitably contaminated by mixed noise during acquisition and imaging, which severely affects subsequent applications. Therefore, HSI denoising is an indispensable part of HSI preprocessing. To solve this problem, this paper proposes an HSI denoising method based on adaptive superpixel segmentation and mathematically grounded regional rank regulation. To avoid the influence of empirical parameters on clustering accuracy, we design an adaptive segmentation framework. This framework autonomously determines the optimal number of homogeneous regions by evaluating internal pixel regional uniformity and boundary discrimination. Instead of relying on a fixed empirical rank, the rank of the CP tensor decomposition for each region is dynamically adjusted. Specifically, the detail complexity of each region is explicitly quantified by combining its spectral variance and spatial gradient energy. The adaptive rank is then mathematically determined by a formulated regression function linking this computed complexity to the region's estimated noise level. Furthermore, a masked completion strategy is introduced to apply CP decomposition to these irregularly shaped superpixels, accurately extracting clean spatial-spectral factors without artifacts. Finally, the Alternating Direction Method of Multipliers (ADMM) optimization framework is used to make the objective function quickly converge. Extensive simulated and real experimental results demonstrate that the proposed method achieves satisfactory denoising performance and robustness. Code will be released at https://qzhang95.github.io.