Anomaly Detection via Tensor Multisubspace Learning and Nonconvex Low-Rank Regularization

Abstract

Hyperspectral anomaly detection represents a crucial application of intelligent sensing, focusing on the identification and localization of anomalous targets. However, the complicated background distribution of hyperspectral imagery (HSI) and the lack of exploration of the intrinsic structure raise enormous challenges for efficient anomaly detection. To address these issues, we introduce the tensor multi-subspace learning strategy with nonconvex low-rank regularization (TMNLR) for anomaly detection in HSI. The HSI is considered as a third-order tensor and is decomposed to background and anomaly, where the tensor subspace and the coefficient tensor are obtained from the background via the tensor multisubspace learning strategy. To improve detection accuracy, the nonconvex low-rank regularization is introduced for suppressing the background, where the optimization process is designed to extract the background coefficient tensor. And the nonisotropic total variation (TV) regularization is jointly implemented to maintain the local spatial similarity of HSI and promote spatial smoothness. Results demonstrate that the proposed framework could achieve an average detection accuracy rate of 97.98% on four real-scene datasets. Extensive experiments validate the effectiveness and robustness of the TMNLR over the comparative methods.

Keywords

• Anomaly detection
• nonconvex tensor low-rank
• tensor multisubspace learning
• total variation (TV)