GLLR-HAD: Global-local low-rank integration for hyperspectral image anomaly detection

• 出版年: Journal of King Saud University: Computer and Information Sciences
• 主要な著者: Yu Bai, Yanling Zhang, Lili Zhang, Tan Zhao
• フォーマット: 論文
• 言語: 英語
• 出版事項: Springer 2025-07-01
• 主題: Hyperspectral anomaly detection; Low-rank prior; Global-local feature interaction; Multi-scale autoencoder
• オンライン･アクセス: https://doi.org/10.1007/s44443-025-00118-y
• ISSN: 1319-1578; 2213-1248

Abstract Hyperspectral imaging (HSI), with its rich spectral and spatial information, offers unique advantages for anomaly detection. However, existing methods often struggle to simultaneously model global structures and local discriminative features, which hampers their ability to detect structural anomalies and adapt to spectral variability in complex scenarios. To overcome this limitation, we propose a novel framework, Global-Local Low-Rank Integration for Hyperspectral Anomaly Detection (GLLR-HAD), which synergistically integrates global and local low-rank modeling to effectively extract both global semantic representations and local fine-grained features from HSI data. Specifically, global low-rank decomposition captures structural anomalies by decomposing the image into a basis matrix and a coefficient matrix, effectively enhancing the recognition of large-scale background variations. Complementarily, local low-rank modeling introduces an adaptive step-size partitioning strategy and a Local Feature Enhancement and Multi-Scale Low-Rank Fusion Module (LFEMLS) to flexibly extract regional details, to perfect the modeling of local variations, and to adapt to localized spectral variability. Furthermore, a multi-scale autoencoder incorporating a Spectral-Spatial Dual Attention (SSDA) mechanism is designed to reconstruct the background, effectively suppressing noise and enhancing anomaly responses. Through the collaborative optimization of global and local components, the proposed method achieves robust detection performance across diverse scenarios and demonstrates strong adaptability to complex and variable spectral conditions. Experimental results show that the proposed method achieves outstanding performance on five publicly available hyperspectral datasets, attaining an average AUC of 0.9944, an index validating its excellent stability and cross-scenario generalization.