Hyperspectral anomaly detection based on nonlocal group sparsity models with low-rank regularization

This is the authors' abstract. We don't add key points for this paper.

Hyperspectral anomaly detection (HAD) is an unsupervised detection task that distinguishes anomaly pixels from background pixels and is widely used in military and civilian applications. Nevertheless, the study of HAD still encounters various challenges, such as the unavoidable issue of anomaly contamination and the limited utilization of spatial information. Regarding this, we propose a HAD method based on nonlocal group sparse representation model with low-rank regularization (NGSM-LR). Specifically, to obtain purer group dictionaries, a singular value decomposition technique is applied to groups composed of nonlocal self-similar cubes, thereby eliminating the influence of anomaly pixels. Moreover, to learn the intrinsic geometric structure of hyperspectral images, an optimization model is proposed to simultaneously exploit the sparsity and low rank of the dictionary-domain reprensentation coefficients for each group, thereby exploring both local and global spatial information. Experimental results on four hyperspectral datasets demonstrate that this method outperforms several state-of-the-art anomaly detectors in terms of accuracy.