Two-Stage Framework For Enhanced Hyperspectral Anomaly Detection

Hyperspectral anomaly detection has garnered significant attention due to its wide applicability across industrial fields. Traditional unsupervised anomaly detection methods typically adopt a one-stage framework, where object and background classes are combined into a unified "normal class". However, this approach can easily suffer from performance degradation in dynamic and noisy industrial environments, primarily caused by background complexity or contamination. To address these limitations, this study purposes a novel two-stage framework. The first stage isolates the background classes, and the second stage focuses on detecting anomalies within the remaining regions. Additionally, a spectral distance-based centralization strategy is introduced to emphasize the relative positional relationships among object, background, and anomaly classes, further improving detection performance. Extensive experiments on benchmark hyperspectral datasets, including Pavia University and Salinas Valley, demonstrate that the proposed framework consistently outperforms traditional one-stage methods, achieving significant performance improvements in scenarios with complex background and object relationships. These findings highlight the potential of the proposed approach in precision agriculture, quality inspection, and other industrial applications.